On diagnostic codes

On diagnos ic codes
Amalie Dahl Haue
This hesis has been submi ed o he
G adua e School o Heal h and Medical Sciences
Uni e si y o Copenhagen
Augus 19, 2021
P eamble
The i le o his hesis joins an in eg al pa o clinical p ac ice and a co ne s one o any compu e
language. While diagnosis codes ep esen a sys ema ic answe o causes o disease and dea h;
a diagnos ic code may bo h e e o he con en o a sc ip and he dual na u e o clinical p ac ice.
Since he ea lies e idence o clinical p ac ice, diagnos ics has e ol ed om a p ac ice ha concluded
manual inspec ion, palpa ion and unc ional assessmen o s anda ds e ol ing in a sys em ha is
hea ily dependen on compu e s. This hesis explo es he bounda ies o diagnos ics in a digi ized
eali y.
Diagnosis Pa ien
Physician
TRUE
s ands o
CORRECT
symbolizes
e e s o
ADEQUATE
Diagnos ics and eali y.∗
∗Adap ed om The Meaning o Meaning: A S udy o he In luence o Language upon Though and o he Science o Symbolism.
Cha les Kay Ogden &I o A ms ong Richa ds. Ha cou , B ace & Wo ld, Inc. New Yo k (1925); p. 11.
Table o con en s
FRONT MATTER
P e ace i
Lis o manusc ip s ii
Summa y iii
Summa y in Danish
PART I SYNOPSIS
0 Objec i es and o e iew 1
1 In oduc ion 3
1.1 Diagnos ics in a web o mul iple causes and consequences . . . . . . . . . . . . . . . 3
1.1.1 Ischemic hea disease is a common, ch onic, mul i- ac o ial disease . . . . . 3
1.1.2 Ischemic hea disease o en co-exis s wi h o he common ch onic diseases . 5
1.2 Classi ica ion sys ems ha monize heal hca e da a . . . . . . . . . . . . . . . . . . . . 6
1.2.1 Classi ica ion sys ems o medical causes and consequences . . . . . . . . . . 6
1.2.2 Classi ica ion sys ems ha e lec clinical p ac ice and examina ion . . . . . . 8
1.3 S eng hs and limi a ions o medical classi ica ion sys ems in esea ch . . . . . . . . 8
2 Ma e ials 10
2.1 The Danish heal hca e sys em as a esea ch esou ce . . . . . . . . . . . . . . . . . . . 10
2.2 Da a om millions o people o e mul iple decades . . . . . . . . . . . . . . . . . . . 11
2.2.1 The Danish Na ional Pa ien Regis y . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 The Danish Regis y o Causes o Dea h . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 The Danish Na ional P esc ip ion Regis y . . . . . . . . . . . . . . . . . . . . 12
2.3 Deepe da a cha ac e ize pa ien s wi h g ea e p ecision . . . . . . . . . . . . . . . . 12
2.3.1 The Eas e n Danish Hea Regis y . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.2 Elec onic Heal h Reco ds om Eas e n Denma k . . . . . . . . . . . . . . . . 13
2.3.3 Copenhagen Hospi al Biobank . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Me hods 14
3.1 Da a, science, and scien i ic models e ol e o e ime . . . . . . . . . . . . . . . . . . . 14
3.2 Comp ehensi e analysis o na ionwide heal h egis ies . . . . . . . . . . . . . . . . . 16
3.2.1 Desc ibing mul i-mo bidi y using empo al disease ajec o ies . . . . . . . . 16
3.2.2 Ta ge ing disease ajec o ies mul i-mo bidi y in ischemic hea disease . . . 18
3.2.3 P esc ip ion ajec o ies model mul i-mo bidi y in he gene al popula ion . . 19
3.3 Clus e analysis o de ine ischemic hea disease subg oups . . . . . . . . . . . . . . . 20
3.3.1 A model whe e mul i-mo bidi y and ex documen s sha e p ope ies . . . . 20
3.3.2 A ma hema ical g aph ep esen a ion o ischemic hea disease pa ien s . . . 21
3.3.3 A clus e algo i hm minimizes a cos unc ion . . . . . . . . . . . . . . . . . . 22
3.3.4 Cha ac e iza ion o clus e s using su i al analysis . . . . . . . . . . . . . . . 23
3.4 Da a-d i en p edic ions o mo ali y in ischemic hea disease . . . . . . . . . . . . . 25
3.4.1 A concep ual in oduc ion o a i icial neu al ne wo ks . . . . . . . . . . . . . 25
3.4.2 Ranking inpu ea u es based on le el o in o ma ion . . . . . . . . . . . . . . 27
3.5 A p obabilis ic map o in-hospi al d ug dosage al e a ions . . . . . . . . . . . . . . . 28
3.5.1 Bayesian models can be decomposed in o h ee dis ibu ions . . . . . . . . . . 28
3.5.2 Applica ion o Bayes’ heo em o elec onic heal hca e da a . . . . . . . . . . 30
4 Da a p o ec ion and p i acy 31
4.1 In o med consen and ins i u ional pe missions . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Legal egula ions and da a managemen . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.3 Summa y o au ho i y app o als . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5 Resul s 34
5.1 A empo al analysis o mul i-mo bidi y in ischemic hea disease . . . . . . . . . . . 34
5.2 A s a egy o de ine mo e homogeneous ischemic hea disease pa ien subg oups . 36
5.3 Mo ali y in ischemic hea disease om a ime- o-e en machine lea ning algo i hm 38
5.4 E idence o d ug-d ug in e ac ions in elec onic heal h eco ds . . . . . . . . . . . . 40
5.5 Es ablishing a map o longi udinal, na ion-wide p esc ip ion pa e ns . . . . . . . . . 42
5.6 In e ne applica ions display pe spec i es wi h da a-d i en esea ch . . . . . . . . . 43
6 Discussion 46
6.1 Conclusion........................................... 46
6.2 S eng hsandlimi a ions................................... 47
6.3 Pe spec i es .......................................... 48
Re e ences 50
Lis o abb e ia ions 60
PART II FULL-LENGTH MANUSCRIPTS 63
7 Tempo ali y in ischemic hea disease mul i-mo bidi y 64
8 Risk s a i ica ion o 72,249 pa ien s wi h ischemic hea disease: a e ospec i e s udy
linking p io mul i-mo bidi y, biochemical da a, and gene ics 89
9 PMHne -alpha: a neu al ne wo k-based disc e e- ime su i al model o mo ali y p e-
dic ion in ischemic hea disease 112
10 Polypha macy and d ug dosage modi ica ions: a longi udinal analysis o 3.5 million
elec onic heal h eco ds 135
11 Op imizing d ug selec ion om a p esc ip ion ajec o y o one pa ien 168
12 Disease ajec o y b owse o explo ing empo al, popula ion-wide disease p og es-
sion pa e ns in 7.2 million Danish pa ien s 199
PART III APPENDICES 210
A Appendix 211
B Appendix 214
C Appendix 237
D Appendix 253
E Appendix 269

P e ace
This hesis was w i en o comple e he G adua e p og amme in Bios a is ics and Bioin o ma ics
a The Facul y o Heal h and Medical Sciences a Uni e si y o Copenhagen and he eby ul ill he
equi emen s o ob aining a PhD deg ee in ag eemen wi h he minis e ial O de on he PhD P o-
g amme a he Uni e si ies and Ce ain Highe A is ic Educa ional Ins i u ions (PhD O de ).
Financially, he wo k was ca ied ou wi h gene ous suppo om The Depa men o Heal h and
Medical Sciences, Uni e si y o Copenhagen (CAG P ecision Diagnos ics in Ca diology), The No o
No disk Founda ion (g an ag eemen s NNF14CC0001 and NNF17OC0027594) and Inno a ions-
onden (p ojec numbe s 60833 and 3-3031-1731/1).
Fo coun less aluable discussions, imely insigh , and i m us , I hank amily (immedia e and
ex ended), iends, colleagues, and he academic ad iso s. Indi idually and in conce you ha e ex-
panded ho izons, induced pe spec i e; and encou aged me o ind my own ajec o y in his model
o eali y called science and he co ne o eali y known as clinic p ac ice.
Amalie Dahl Haue
Copenhagen 2021
i
Lis o manusc ip s
1. ∗Haue AD, A men a os JJA, Holm PC, Moseley PL, Købe LV, Bundgaa d H, B unak S. Tem-
po ali y in ischemic hea disease mul i-mo bidi y. Manusc ip in p epa a ion.
2. ∗Haue AD, Holm PC, Banasik K, Lundgaa d AT, Muse VP, Rode T, Wes e gaa d D, Chmu a
P, Siggaa d T, Ch is ensen AH, Weeke PE, Sø ensen E, Os owski SR, DF Gulbja sson, Holm
H, I e sen KK, Købe LV, Ullum H, Bundgaa d H, B unak S. Risk s a i ica ion o 72,249 pa-
ien s wi h ischemic hea disease: a e ospec i e s udy linking p io mul i-mo bidi y, bio-
chemical da a, and gene ics. Manusc ip submi ed.
3. ∗Holm PC, Haue AD, Banasik K, B unak S, Bundgaa d H. PMHne -alpha: a neu al ne wo k-
based disc e e- ime su i al model o mo ali y p edic ion in ischemic hea disease. Manu-
sc ip in p epa a ion.
4. †Rod íguez CL, Mazonni G, Haue AD, E iksson R, Biel JH, Can well L, Wes e gaa d D,
Belling KG, B unak S. Polypha macy and d ug dosage modi ica ions: a longi udinal analy-
sis o 3.5 million elec onic heal h eco ds. Manusc ip in e ision.
5. †Aguayo-O ozco A, Haue AD, Jø gensen IF, Wes e gaa d D, Moseley PL, Mo ensen LH,
B unak S. Op imizing d ug selec ion om a p esc ip ion ajec o y o one pa ien . Manusc ip
in e ision.
6. †Siggaa d T, Reguan R, Jø gensen IF, Haue AD, Lademann M, Aguayo-O ozco A, Hjal-
elin JX, Jensen AB, Banasik K, B unak S. Disease ajec o y b owse o explo ing empo al,
popula ion-wide disease p og ession pa e ns in 7.2 million Danish pa ien s. Na Commun.
2020 Oc 2;11(1):4952. DOI: h ps://doi.o g/10.1038/s41467-020-18682-4.
∗: Main manusc ip s
†: Con ibu o y manusc ip s
ii
Summa y
In oduc ion As he digi ized eali y is ans o ming mode n medicine, da a olumes and a ie y
a e g owing conside ably while he isions o heal hca e a e also being e ised. Ye i emains an
open ques ion how o con e his weal h o da a in o clinically ac ionable e idence ha is becom-
ing e en mo e needed in an ageing popula ion. In his hesis, I p esen six o iginal s udies ha
in es iga e he po en ial o Danish heal hca e da a as a pla o m o pe o m da a-d i en esea ch.
The hesis has a s ong ocus on mul i-mo bidi y in ischemic hea disease (IHD) and he applied
me hods ange om classical eg ession models o machine lea ning me hods. Mul i-mo bidi y
in IHD is bo h cha ac e ized wi h espec o he empo ali y be ween diagnoses and by analyzing
mo e ine-g ained heal hca e da a, such as elec onic heal h eco ds (EHRs) and gene ic da a. The
p ac ical consequences o mul i-mo bidi y a e also s udied based on analyses o polypha macy in
an in-hospi al se ing and in he gene al popula ion. Finally, he hesis p esen s an example sha ing
esea ch esul s as a web ool made a ailable o o he s wi hou sha ing pe son-sensi i e da a.
Manusc ip 1: Tempo ali y in ischemic hea disease mul i-mo bidi y. This s udy cha ac e izes
mul i-mo bidi y in IHD using The Danish Na ional Pa ien Regis y (NPR) da a om IHD pa ien s
in he pe iod yea s 1994–2018. Mul i-mo bidi y is cha ac e ized by applica ion o a new e sion o
he so-called disease ajec o y p og am ha es ablishes empo al associa ions be ween diagnoses.
The s udy showcases di e ences be ween disease ajec o ies con aining condi ions ha a e solely
isk ac o s, and condi ions ha can be bo h isk ac o s and complica ions. The s udy a gues ha
disease ajec o ies can be used o sys ema ically assess he empo ali y in mul i-mo bidi y, which
is impo an , ye o en an unde app ecia ed aspec , o mul i-mo bidi y.
Manusc ip 2: Risk s a i ica ion o 72,249 pa ien s wi h ischemic hea disease: a e ospec-
i e s udy linking p io mul i-mo bidi y, biochemical da a, and gene ics. In his s udy, mul i-
mo bidi y in IHD is add essed by cha ac e izing dis inc pa ien subg oups ob ained by applica ion
o a clus e algo i hm o diagnosis codes. In a coho o oughly 75 000 pa ien s, pa ien subg oups
a e iden i ied based on simila i ies be ween pa ien -speci ic ec o s ha ep esen all diagnoses eg-
is e ed in NPR up o 24 yea s be o e IHD onse . Da a om EHRs and Copenhagen Hospi al biobank
(CHB) a e included o assess pheno ypic di e ences in hese pa ien subg oups ob ained by appli-
ca ion o he clus e algo i hm. The pa ien subg oups associa e wi h di e en p og ession a es
as well as biochemical and gene ic di e ences, indica ing ha he pa e ns cap u ed by he clus e
algo i hm a e biologically ele an .
iii
Manusc ip 3: PMHne -alpha: a neu al ne wo k-based disc e e- ime su i al model o mo a-
li y p edic ion in ischemic hea disease. This s udy p esen s PMHne -alpha, which is a neu-
al ne wo k-based su i al model o p edic ion o all-cause mo ali y in IHD pa ien s. PMHne -
alpha is based on a de i a ion coho o app oxima ely 40 000 IHD pa ien s wi h e i ied co ona y
pa hology acco ding o The Eas e n Danish Hea Regis y (EDHR) da a. Roughly 600 inpu ea-
u es ex ac ed om EDHR, NPR, EHRs, and CHB, including diagnoses, blood es s, and polygenic
sco es a e he basis o de elopmen o PMHne -alpha ha ou pe o ms he GRACE2.0 isk sco e.
PMHne -alpha is also subjec ed o explainabili y analysis whe e p edic i e alueso inpu ea u es
a e anked and cha ac e ized as posi i e o nega i e depending on how hey a ec he p edic ion.
Manusc ip 4: Polypha macy and d ug dosage modi ica ions: a longi udinal analysis o 3.5 mil-
lion elec onic heal h eco ds In his s udy, polypha macy is s udied wi h e e ence o co-medica-
ion pai s ha a e de ined by a combina ion o wo dis inc d ugs whe e he likelihood o a dosage
adjus men is mo e likely du ing concomi an ea men episodes han when adminis e ed as a
mono he apy. By ex ac ing in o ma ion on d ug egimens om EHRs om abou one million in-
hospi al pa ien s in he pe iod yea s 2008–2016, he s udy p esen s 3993 co-medica ion pai s ha
we e c oss- e e ences wi h 15 publicly a ailable d ug-d ug in e ac ion da abases. We a gue ha
he analysis p o ides e idence o up o 600 p e iously undesc ibed d ug-d ug in e ac ions which
we e iden i ied as co-medica ion pai s and no p esen in a leas one o he d ug-d ug in e ac ion
da abases.
Manusc ip 5: Op imizing d ug selec ion om a p esc ip ion ajec o y o one pa ien . This
s udy add esses polypha macy by analyzing da a om The Danish Na ional P esc ip ion Regis y
(DNPR) in he pe iod yea s 1995–2019 comp ehensi ely. The s udy p esen s he so-called p esc ip-
ion ajec o ies ha map he p esc ip ion pa e ns om mo e han se en million people p esen in
DNPR. P esc ip ion ajec o ies con ain up o se en dis inc edeemed d ugs ha a e mo e likely
o be edeemed in one speci ic o de . In he s udy, we a gue ha he analysis can suppo op i-
mized ea men decisions by analyzing di e ences be ween people subjec ed o many changes in
ea men egimens which can hope ully lead o mo e indi idualized ca e.
Manusc ip 6: Disease ajec o y b owse o explo ing empo al, popula ion-wide disease p o-
g ession pa e ns in 7.2 million Danish pa ien s. This s udy p esen s The Danish Disease T ajec-
o y B owse , which is a web applica ion ha enables use s o explo e da a om NPR in he pe iod
yea s 1994–2018 based on esul s om he disease ajec o y p og am. Thus, he s udy exempli ies
newe possibili ies o sha ing esea ch esul s. In he b owse , use s can sea ch o one o mo e
diagnoses and sea ch esul s will be displayed in he o m o disease ajec o ies. The b owse has
an a ay o sea ch il e s ha use s can apply o e ine sea ches as well as an in e ace ha allows o
na iga e he sea ch esul s and modi y he isual ep esen a ion. Sea ch esul s can be expo ed in
he o m o g aphical ep esen a ions o disease ajec o ies and summa y s a is ics.
i
CHAPTER 1. INTRODUCTION ON DIAGNOSTIC CODES
Figu e 1: Decline in ca dio ascula diseases in ela ion o scien i ic ad ances. CABG: Co o-
na y a e y by-pass g a ing. PCI: Pe cu anous co ona y in e en ion. MI: Myoca dial in a c ion.
NHBPEP: Na ional High Blood P essu e Educa ional P og am. CASS: Co ona y a e y su ge y
s udy. TIMI: Th ombolysis in myoca dial in a c ion. NCEP: The Na ional Choles e ol Educa ion
P og am. GISSI: G uppo I alioano pe la Spe imen azione della s e ochinasi nell’IN a o Mio-
ca dico. ISIS-2: The Second S udy o In a c Su i al. SAVE: Su i al and Ven icula Enla gemen
T ial. ALLHAT: An ihype ensi e and Lipid-Lowe ing T ea men o P e en Hea A ack T ial.
F om [9].
To da e, he Global Regis y o Acu e Co ona y E en s (GRACE) has p o ided he da a ounda-
ion o one o he mos widely used isk sco ing schemes o IHD pa ien s[11]. GRACE e e s o a
la ge mul i-na ional egis y, whe e pa ien s wi h acu e co ona y synd ome we e ec ui ed om 14
di e en coun ies o e he pe iod om yea s 1999 o 2009. Da a om GRACE comp ise he oun-
da ion o a isk sco e o he same name ha es ima es sho - and long- e m mo ali y in pa ien s
wi h acu e MI. The GRACE isk sco e has been upda ed se e al imes and exp esses he isk o six
mon h mo ali y on a scale o 1 o 372, whe e a sco e o 1is conside ed minimal isk[12]–[14].
Recen ad ances wi hin gene ic esea ch ha e also pa ed he way o de elopmen o polygenic
sco es (PGSs) ha ha e ueled pe spec i es wi hin he p ecision medicine agenda, including mo e
p ecise isk es ima es[15]. The e is al eady da a suppo ing he hypo hesis ha polygene ic p edic-
o s can be used o iden i y indi iduals a inc eased isk o IHD and o he mul i- ac o ial diseases
equi alen o ha o gene ic p edic o s o monogenic diseases. Fu he , i has been s ipula ed ha
he inc easing in e es in ecycling popula ion-wide heal hca e da a combined wi h con inuous ex-
pansion o compu a ionalcapaci y, p esen endlessoppo uni ies o b oadly in eg a inggene icda a
in o clinical decision making[16], [17]. Ideally, p inciples ha in eg a e con en ional pa hology ob-
ained om CAGs and o he clinical examina ions combined wi h deep pheno ypes ha es ed om
4

CHAPTER 1. INTRODUCTION ON DIAGNOSTIC CODES
human da a collec ions will inc ease heal hca e quali y o he indi idual[18]. Howe e , he speci ic
s a egies ha will ansla e hese pe spec i es in o clinical p ac ice ha e no been ully es ablished
ye and acco dingly much explo a i e wo k is being conduc ed wi hin his ield[19].
1.1.2 Ischemic hea disease o en co-exis s wi h o he common ch onic dis-
eases
Up o 85% o IHD pa ien s a e co-mo bid, b oadly de ined as a pheno ype comp ised o mo e han
one ch onic disease [9], [20], [21]. Co-mo bidi y and mul i-mo bidi y a e o en used in e change-
ably. Technically, co-mo bidi y desc ibes he pheno ype wi h e e ence o an index disease, whe eas
mul i-mo bidi y deno es he pheno ype wi hou e e ence o a pa icula index disease[22], [23]. In
his hesis, I ha e chosen o use he e m mul i-mo bidi y in IHD o ensu e consis ency and s ess
ha a any poin in ime se e al condi ions may be equally impo an [24].
The impo ance o mul i-mo bidi y has been a icula ed o mo e han 50 yea s[25]. Back hen,
he e was al eady a ocus on he consequences o ailu e o classi y and analyze mul i-mo bidi y ow-
ing o he po en ial causal ole o mul i-mo bidi y in p ognosis and he apeu ic e ec s. In andom-
ized con olled ials (RCTs) and o he ypes o clinical s udies, unacknowledged mul i-mo bidi y
may con ound o modi y he esul s, making e icien me hods o measu ing mul i-mo bidi y an
unme need[26]. Ne e heless, he e is no consensus ega ding mul i-mo bidi y measu es and he
exis ing ools o measu ing mul i-mo bidi y a y g ea ly in alidi y, eliabili y and impo an ly,
a ge popula ion[26].
His o ically, he mos widely used co-mo bidi y measu e is he Cha lson co-mo bidi y index
(CCI), which was de i ed om he co ela ion o 19 selec ed diseases wi h mo ali y among can-
ce pa ien s. In a coho o 685 b eas cance pa ien s, he in es iga o s ound ha only age and
co-mo bidi y we e isk ac o s o co-mo bid dea h, i.e. dea h om o he causes han me as a ic dis-
ease[27]. Al hough he e does no exis any speci ic mul i-mo bidi y sco e o IHD pa ien s, he
e idence ha mul i-mo bidi y associa es wi h he p ognosis and bu den o he indi idual pa ien is
endless and in p ac ice, CCI is widely used i espec i e o main pheno ype[28]. In acknowledging
he p ognos ic in e dependency o seemingly dis inc diseases, i has been sugges ed ha deli e y
o IHD heal hca e should become mo e mul idisciplina y a he han p ima ily designed o a ge
one condi ion a a ime[20], [29]. Examples o mul idisciplina y app oaches include in eg a ion o
knowledge ac oss di e en medical special ies and enhanced ocus on pa ien p e e ences wi hin
a amewo k ha is a ge ed ea men o IHD and o he co-exis ing ch onic condi ions, i.e., mul i-
mo bidi y[9].
A common he apeu ic consequence o mul i-mo bidi y is a condi ion known as polypha macy,
e e ing o he simul aneous usage o mul iple d ugs by a single pa ien [30]. And gene ally, he e
is a knowledge gap ega ding bene i s, isks, and in ensi y o ea men in ca dio ascula pa ien s,
wi h an inc easing p e alence as well as age[31]. Majo conce ns wi h polypha macy a e ha i in-
c eases he isk o ad e se eac ions, educes he likelihood o compliance, and he e is e y limi ed
e idence ega ding he long- e m consequences o he d ugs ha a e o en p esc ibed li e-long[30],
[32]. In ac , he inc easing complexi y o medical he apy is gaining a en ion in so-called adap i e
pla o m ials, whe e mul iple he apeu ic in e en ions may be compa ed in he same ials[33].
5
CHAPTER 1. INTRODUCTION ON DIAGNOSTIC CODES
1.2 Classi ica ion sys ems ha monize heal hca e da a
1.2.1 Classi ica ion sys ems o medical causes and consequences
O iginally being he backbone o compa able mo ali y s a is ics da ing back as long as o 1893, he
In e na ional S a is ical Classi ica ion o Diseases and Rela ed Heal h P oblems (ICD) has u ned
in o an impo an esou ce in heal hca e sys ems which a e inc easingly in luenced by ad ances in
in o ma ion echnology[34]. One inhe en ad an age o he ICD e minology is ha i is gene aliz-
able ac oss many di e en heal hca e sys ems, as i is used in many coun ies[35]. Since 1994, he
10 h e sion o he ICD sys em (ICD-10) has been he o icial diagnos ic epo ing sys em be ween
Danish hospi als and he Danish heal h au ho i ies[36]. The hie a chical s uc u e o he classi ica-
ion sys em is comp ised o 21 chap e s ha la gely co espond o di e en unc ional-ana omical
body sys ems. Fo example, chap e IX co esponds o he Diseases o he ci cula o y sys em. Each
chap e can be di ided in o blocks ha a e u he segmen ed in o diagnosis codes a di e en es-
olu ions, e.g. le el 3 and 4 codes wi h le el 4codes being he mos speci ic. Le el 4codes amoun
o a abula lis ing o mo e han 155 000 dis inc diagnosis codes (Figu e 2).
Figu e 2: A sea ch in he ICD-10 e minology o IHD whe e he le el 3ICD-10 code ch onic IHD
was selec ed as indica ed by he a ows in he panel on he le side. ICD-10 chap e s, blocks, le el
3and le el 4codes a e lis ed in he panel on he le . Desc ip ions o he selec ed le el 3and 4
codes a e on a ailable on he igh side. ICD-10: In e na ional Classi ica ion o Diseases and Rela ed
Heal h P oblems, 10 h e sion. IHD: Ischemic hea disease. Re ie ed om h ps://icd.who.in /
b owse10/2019/en#/I25.
6
CHAPTER 1. INTRODUCTION ON DIAGNOSTIC CODES
The Ana omical The apeu ic Chemical (ATC) Classi ica ion Sys em was o iginally de eloped as
a ool o co ec ly in e p e da a on d ug u iliza ion[37]. Like ICD-10 codes, he ATC Classi ica ion
Sys em has a hie a chical s uc u e, whe e g oups a i e di e en le els a e used o classi y each
d ug a inc easing speci ici y. The g oups ange om he 1s le el codes ha consis o 14 ana omical
g oups speci ying he p ima y o gan sys em o ac ion, o he 5 h le el codes, which speci y he ac i e
chemical subs ance o a gi en d ug[38]. In con as o he ICD-10 e minology ha has one unique
code pe diagnosis; he ATC Classi ica ion Sys em may ha e mo e han one code o one d ug.
In cases whe e a single chemical subs ance has mul ipe ou es o adminis a ion, his is s anda d.
Simila ly, al hough non-s anda d, chemical subs ances wi h mo e han one app o ed indica ion
may also ha e mo e han one ATC code (Figu e 3).
Figu e 3: Sea ch esul s in he ATC Index o he 5 h le el ATC code B01AC06 which is ace ylsal-
icylic acid co esponding o i s indica ion as an an i h ombo ic agen . G een inse indica es 1s
h ough 4 h ATC Index le els. The desc ip ion below explains he eason o he exemp ion ha
ace ylsalicylic acid may also be classi ied in he chemical subg oups N02BA ha con ains salicyclic
acid. In his case, he indica ion de e mines he ATC code. ATC: Ana omical The apeu ic Chemical.
Re ie ed om h ps://www.whocc.no/a c_ddd_index/.
7
CHAPTER 1. INTRODUCTION ON DIAGNOSTIC CODES
1.2.2 Classi ica ion sys ems ha e lec clinical p ac ice and examina ion
The Nomencla u e, P ope ies and Uni s in Labo a o y Medicine (NPU) e minology, ep esen ed
by NPU codes, index analyses wi hin clinical labo a o y sciences. NPU codes ha e been used since
1987 and a e owned by he In e na ional Fede a ion o Clinical Chemis y and Labo a o y Medicine
and In e na ional Union o Pu e and Applied Chemis y. Since 1987 he numbe o a ailable exami-
na ions om clinical labo a o ies has inc eased om a ew hund eds o mo e han 30 000 di e en
ypes o es s[39]. NPU codes we e de eloped o con ey in o ma ion ega ding he p ope y o a
s udied objec , i.e. he pa ien . The h ee essen ial elemen s o an NPU code a e (i) he sys em, (ii)
he componen , and (iii) he kind-o -p ope y. Fo example, an NPU code o oponin is NPU18583
which is de ined by P—T oponin I, ca diac muscle; subs .c.= ? nmol/L. The componen is "T oponin
I, ca diac muscle". Any hing be o e he componen e e s o he sys em (P in his case o plasma)
and any hing a e e e s o he kind-o -p ope y, which in his case is he concen a ion measu ed
in nmol/L. Impo an ly, NPU codes do no p o ide any in o ma ion ega ding he co ec ness o
he measu emen also meaning ha e e ence in e als a e de ined locally a he clinical labo a o y
whe e he analyses a e being pe o med[39].
Clinical examina ions and p ocedu es a e ye ano he aspec o medical ca e ha is being o ga-
nized acco ding o a e minology. Fo example, he No dic Medico-S a is ical Commi ee (NOMES-
CO) was es ablished in he ea ly 1980s o compa e he equencies o su gical ac i i ies in he No dic
coun ies. The i s NOMESCO Classi ica ion o Su gical P ocedu es (NCSP) was published in 1996
and con ains 20 chap e s. E e y NCSP code consis s o h ee le e s and h ee digi s. The i s le e
o he NCSP code co esponds o he chap e . Simila o he case o ICD-10 and ATC codes, NCSP
chap e s a e a anged acco ding o unc ional-ana omic body sys ems. Fo example, he NCSP code
o PCI is FNG05, whe e F indica es ha i is su ge y pe o med on he hea o majo ho acic es-
sels and N speci ies ha i is pe o med on he co ona y a e ies[40]. In epo ing om he Danish
heal hca e sys em, p ocedu e codes a e gene ally mo e eliable han diagnosis codes[41].
1.3 S eng hs and limi a ions o medical classi ica ion sys ems in
esea ch
In many heal hca e sys ems a ound he wo ld, ICD-10 and ATC codes a e an e icien means o
compa ing dea h s a is ics, d ug usage, and handling billing pu poses. Simila ly, NPU and NCSP
codes a e ins umen al adminis a i e esou ces. Recen ly, he alue o heal hca e da a de i ed om
o he sou ces han RCTs has been e e ed o as eal-wo ld e idence and is gaining a en ion as a
esou ce o de ine o he endpoin s han all-cause mo ali y[42]. Po en ially, his is highly aluable
in an ageing popula ion whe e ac o s such as quali y o li e and he apeu ic e icacy a e becom-
ing inc easingly impo an aspec s o heal hca e[31]. Al hough classi ica ion sys ems and clinical
e minologies a e necessa y in his espec , hey a e no pa icula ly designed o e lec di e en-
ial diagnos ic didac ic no o encompass he ue complexi y o mul i-mo bidi y. In o he wo ds,
he e is no gua an ee ha a pa ien who has been assigned a diagnosis code also was subjec ed o
adequa e diagnos ic examina ion. This also applies o ATC codes in he sense ha i a pe son has
edeemed a p esc ip ion o a d ug, he e is no e idence o compliance, i.e. whe he o no he pe -
8
CHAPTER 1. INTRODUCTION ON DIAGNOSTIC CODES
son ook he d ug as p esc ibed. Simila ly, an NCSP code as such has no in o ma ion ega ding he
conclusion o he p ocedu e ha i desc ibes. Essen ially, in isola ion no e minology p o ides any
in o ma ion ega ding he ue pheno ype, whe e inc easing p e alences o mul i-mo bidi y and
polypha macy complica e ma e s e en u he . Mo eo e , e en in he simples cases, diagnos ics
emain an incomple e ep esen a ion o eali y and s ill au oma ized diagnos ic sys ems a e no used
ou inely[43], [44]. Ye , he classi ica ion sys ems p esen ed in sec ion 1.2 ha e become impo an
ools o classi y da a poin s and ha monize heal hca e da a ac oss coun ies. These e o s a e pa -
icula ly impo an in o de o ans o m as much da a as possible in o a ounda ion o imp o ed
medical ea men [42]. The adap ion o ICD-10 and ATC wo ld-wide has pa icula ly been d i en
by he Wo ld Heal h O ganiza ion (WHO) leade ship d i ing he ICD[45]. The sys ems a e also
being upda ed con inuously e lec ing a medical ield ha is cons an ly unde going de elopmen .
Fo example, ICD-11 in which he ATC codes a e embedded, has been designed o use in a digi al
wo ld and planned o come in o e ec in 2022[46].
The apho ism“ga bagein, ga bage ou ”has esona edbehind so wa e e e since he ea lies days
o compu e powe [47]. And i s implica ion emains he same. Thus, o make up o he ac ha
diagnos ics in he classical sense was no pe o med o e he cou se da a-d i en s udies, including
he onesp esen ed in his hesis; success uladap a ion o medicalclassi ica ionsys emsis absolu ely
necessa y o gene a e aluable da a-d i en models. In ac , i has been sugges ed qui e ecen ly
o c ea e mo e da a-d i en and mechanis ically o ien ed disease classi ica ion sys ems ha would
ma ch he aims o he p ecision medicine agenda be e [48]. By disc imina ing be ween appa en ly
simila pheno ypes, using a wide ange o da a (e.g. egis y, molecula , and gene ic da a) he
diagnos ic p ocess may become much mo e indi idualized and in e ec suppo be e ea men
decisions. A conc e e a emp is he Human Pheno ype On ology (HPO) ha now inc easingly
makes i s way in o heal hca e sys ems[49]. The wo k in his hesis has a simila aim and pe spec i e.
9

2|Ma e ials
This chap e gi es an o e iew o he da a sou ces ha we e he basis o explo ing he po en ial o
conduc ing da a-d i en esea ch in he se ing o Danish heal hca e da a. Sec ion 2.1 in oduces he
basic s uc u es wi hin he Danish heal hca e sys em o ele ance o s udies p esen ed in his hesis,
while sec ion 2.2 desc ibes he ele an na ionwide egis ies and sec ion 2.3 p esen s he sou ces o
deepe pheno ypic and gene ic da a.
2.1 The Danish heal hca e sys em as a esea ch esou ce
Two o he essen ial aspec s o he Danish heal hca e sys em is ha edis ibu ion o axes co e s
abou 85% o he heal hca e and he ole o pe sonal iden i ica ion numbe s[36]. Pe sonal iden i i-
ca ion numbe s a e en-digi numbe s ha since 1968 ha e iden i ied e e y Danish ci izen uniquely.
The Danish Ci il Regis a ion Sys em (CRS) adminis e s all pe sonal iden i ica ion numbe s[50].
Pe sonal iden i ica ion numbe s unc ion as unique iden i ie s o ci izens ac oss all public sec o s
including heal hca e se ices[51]. Thus, in medical esea ch based on Danish heal hca e da a, CRS
acili a es linkage o di e en da a sou ces gi en he s udy has been app o ed app op ia ely (c .
sec ion 4). In addi ion o he linkage unc ion, CRS con ains in o ma ion such as da e o bi h, sex
and s a us (e.g. ali e, dead, o emig an ) o he indi idual ci izen. Figu e 4 shows he da a sou ces
ha we e linked ia CRS in his hesis p io o analysis.
1970 1980 1990 2000 2010
CRS†
DAR∗NPR∗
DNPR∗EDHR•BTH•
CHB‡
Figu e 4: O e iew o da a ounda ion and esou ces o de ed by yea o i s obse a ion in e-
sou ce. †: Founda ion o Danish heal hca e da a in as uc u e. ∗: Na ionwide egis ies. •: Re-
gional heal hca e da a wi h deepe pheno ypic da a. ‡: Gene ic da a. CRS: The Danish Ci il Regis-
a ion Sys em. DAR: The Danish Regis y o Causes o Dea h. NPR: The Danish Na ional Pa ien
Regis y. EDHR: Eas e n Danish Hea Regis y. BTH: BigTempHeal h (EHRs om Eas e n Den-
ma k). CHB: Copenhagen Hospi al Biobank. EHRs: Elec onic heal h eco ds
10
CHAPTER 2. MATERIALS ON DIAGNOSTIC CODES
2.2 Da a om millions o people o e mul iple decades
2.2.1 The Danish Na ional Pa ien Regis y
The Danish Na ional Pa ien Regis y (NPR) was es ablished in 1977 and is one o he oldes na ion-
wide heal hca e egis ies in he wo ld[36]. I was es ablished wi h he p ima y aim o con inuous
moni o ing o hospi al and heal hca e se ice u iliza ion o he Danish Heal h Au ho i y. NPR con-
ains da a ela ed o all discha ges om Danish hospi als, which a e epo ed in he o m o ime-
s amped pa ien con ac s e lec ing ha ac i i ies wi hin Danish hospi als a e s uc u ed acco ding
o he so-called danske kon ak model[52]. A pa ien con ac con ains in o ma ion such as du a ion and
ype o con ac (e.g., in-hospi al, ou -hospi al o eme gency oom isi ), diagnosis codes assigned
du ing he con ac , ype o diagnosis code, codes ha documen p ocedu es pe o med du ing he
con ac and a my iad o o he ypes o s uc u ed adminis a i e in o ma ion. One con ac may span
a ew hou s o se e al yea s, depending on he con ac ype and eason[53].
In NPR, egis a ions a e epo ed in acco dance wi h Sundheds æsene s Klassi ika ionssys em
(SKS), which is o ganized in chap e s co esponding o di e en aspec s o heal hca e.4Fo exam-
ple, chap e D co esponds o he ICD-10 classi ica ion and chap e K con ains NCSP codes. Ye
o he chap e s con ain Danish classi ica ion sys ems. One such example is chap e U ha con ains
codes o non-su gical p ocedu es such as adiologic examina ions[40]. Diagnosis codes a e unique
o NPR in he sense ha i is manda o y o epo a leas one diagnosis code o e e y single con ac .
In con as , no all con ac s ha e a ele an p ocedu e code, which is ob ious o con ac s whe e no
p ocedu es we e pe o med. Howe e , i was no un il yea 2000 ha i was equi ed by he hos-
pi als o epo he pe o med p ocedu es o he Danish heal h au ho i ies[36]. His o ically, he
diagnosis codes ha e been used mos ex ensi ely in esea ch. They can ei he be assigned as a p i-
ma y o non-p ima y diagnosis. P ima y diagnosis codes e e o he diagnosis ha bes desc ibes
he eason o he con ac and non-p ima y codes may complimen ha desc ip ion making hem
highly ele an in he con ex o mul i-mo bidi y[36], [50].
2.2.2 The Danish Regis y o Causes o Dea h
In Denma k, comple ion o dea h ce i ica es has been manda o y by law since 1871. I was no un il
1970 ha dea hs among ci izens dying in Denma k we e egis e ed in indi idual elec onic eco ds
in DAR. Un il hen he da a om dea h ce i ica es we e a chi ed on punched ca ds. Da a in DAR
o igina es om dea h ce i ica es which can only by w i en by a physician. Today The Danish Reg-
is y o Causes o Dea h (DAR) is main ained by he Danish Heal h Au ho i y. Dea h ce i ica es
ha da es back be o e 1994 a e a chi ed in he Danish Na ional A chi es. Va iables in DAR in-
clude manne (i.e. na u al, acciden , iolence, suicide, and unce ain), ime, place, and causes o
4The ull SKS is a ailable a h ps://medin o.dk/sks/b ows.php [Danish].
11
CHAPTER 2. MATERIALS ON DIAGNOSTIC CODES
dea h[54]. Unde lying and con ibu o y causes o dea h ha e been classi ied acco ding o in e na-
ional classi ica ion sys ems since 1948 whe e hey ha e been main ained by WHO. This means ha
causes o dea h a e now being classi ied by ICD-10 codes. Be o e 1948 causes we e egis e ed using
Danish and Scandina ian disease classi ica ions. The egis y was cen ally alida ed un il 2002 and
since hen, he egis y has elied solely on in o ma ion om he medical doc o who has e i ied
he dea h[54]. The e is some o e lap be ween da a in CRS and DAR as he ime o dea h is a a iable
ha is a ailable in bo h CRS and DAR. The o he a iables ela ed o dea h a e unique o DAR[51],
[54].
2.2.3 The Danish Na ional P esc ip ion Regis y
The Danish Na ional P esc ip ion Regis y (DNPR) con ains da a on all indi idual p esc ip ions
dispensed a any Danish communi y pha macy om 1994 and onwa ds. DNPR is a sub egis y
o Regis e o Medicinal P oduc S a is ics, which is main ained by he Danish Medicines Agency.
DNPR is adminis e ed by S a is ics Denma k. The ou main a iable ca ego ies a e a iables ela ed
o he use (i.e. he ci izen o pa ien ), he p esc ibe (i.e. he ea ing physician), he d ug and he
pha macy. Only edeemed p esc ip ions a e en ies in DNPR. Dispensed d ugs a e egis e ed in
acco dance wi h he global ATC Classi ica ion Sys em ha is equi alen o chap e M o SKS. O e -
he-coun e (OTC) d ugs a e only en ies in DNPR i hey we e dispensed as p esc ip ions[55].
2.3 Deepe da a cha ac e ize pa ien s wi h g ea e p ecision
A key elemen o his hesis is o explo e he po en ial o he Danish heal hca e da a by analyzing
i comp ehensi ely. Deepe pheno ypic da a was he e o e analyzed in addi ion o he na ionwide
egis ies p esen ed in sec ion 2.2. In b oad e ms, deep pheno ypes a e based on indi idual-le el
in o ma ion ha is mo e ine-g ained han he in o ma ion in o example heal hca e egis ies[19],
[56].
2.3.1 The Eas e n Danish Hea Regis y
The Eas e n Danish Hea Regis y (EDHR) is as subse o he Danish Hea Regis y ha was es-
ablished in 2000. I con ains da a on all pa ien s who ha e been subjec ed o CAG, PCI, co ona y
a e y bypass g a ing o hea al e su ge y in Denma k. EDHR con ains da a om in asi e ca -
diac p ocedu es pe o med a a hospi al in Eas e n Denma k (equi alen o he Capi al Region and
Region Zealand) and pa ien s younge han 14 a ime o he p ocedu e a e excluded om he eg-
is y. Cu en ly, EDHR con ains da a on abou 100 000 unique pa ien s many o whom ha e mul iple
en ies in he egis y. Each en y is desc ibed wi h up o 70 a iables. The a iables co e in o -
ma ion on e.g., demog aphics, disposi ions, p ognos ic ac o s, indings (e.g. co ona y pa hology),
he apeu ic consequences, and p ocedu e- ela ed complica ions[57].
12
CHAPTER 2. MATERIALS ON DIAGNOSTIC CODES
2.3.2 Elec onic Heal h Reco ds om Eas e n Denma k
The EHR da a used in he s udies p esen ed in his hesis o igina ed om BigTempHeal h (BTH),
which deno es a esea ch p ojec based on EHRs om Eas e n Denma k in he pe iod yea s 2006
o 2016. The e m EHR da a is used o desc ibe any ou inely collec ed heal hca e da a, ha a e
no necessa ily a chi ed in a heal h- egis y, a biobank o a simila specialized ins i u ion[19]. Ide-
ally, EHR da a would con ain con inuous and o p ac ically pu poses comple e moni o ing o all
pa ien s du ing hospi aliza ion[56], [58].
Biochemical es s, in-hospi al p esc ip ions and ee ex o igina ing om he aw clinical no es
a e he h ee componen s ha make up BTH. The biochemical da a we e o iginally s o ed in he
wo da abases Labka and BCC, co e ing The Capi al Region and Region Zealand, espec i ely[59].
The biochemical es s we e indexed using NPU codes o local sys ems and e e ence anges we e
p o ided by he labo a o y ha pe o med he analyses. The p esc ip ion da a a ailable in BTH
we e o iginally a chi ed in he h ee sys ems Elec onic Pa ien Medica ion 1,3and OpusMedicine,
co e ing The Capi al Region and Region Zealand, espec i ely. Analogously o he p esc ip ion
da a in DNPR, he p esc ip ion da a in BTH we e indexed acco ding o he ATC Classi ica ion. In
con as o he da a in DNPR, he p esc ip ion da a ob ained om EHRs also con ain in o ma ion
on he d ug dosages o he indi idual dispensa ion. The clinical no es include uns uc u ed da a,
whe e clinical a iables such as blood p essu e, heigh and weigh can be ex ac ed by ex mining.
Simila ly, alle gies and symp oms can in p inciple be iden i ied in his pa o he da a.
2.3.3 Copenhagen Hospi al Biobank
CopenhagenHospi alBiobank(CHB)wases ablished in2009 omakeuseo esidual blood samples
ha we e p e iously disca ded. I was ini ia ed a Rigshospi ale and la e expanded o include sam-
ples om all public hospi als in he Capi al Region o Denma k. CHB con ains gene ic da a ob ained
om blood ha was o iginally d awn o blood- ype es ing o ed cell an ibody sc eening[60]. The
gene ic analyses pe o med on hese samples a e cu en ly being conduc ed in collabo a ion wi h
deCODE gene ics[61]. Since he biological ma e ial is le o e om ou ine es s, s udy pa icipan s
should in p inciple be asked o in o med consen p io o inclusion. Howe e , he biobank has a
dispensa ion om he gene al ule equi ing consen . In e ec , s udy pa icipan s a e no asked o
in o med consen p io o inclusion bu can op ou a any ime. Samples ompa ien s younge han
18 yea s a e excluded. Likewise, one pa ien can maximum be included once meaning ha samples
om pa ien s who a e al eady included in CHB a e disca ded. As o Feb ua y 2020, gene ic da a
we e a ailable on abou 155 000 pa ien s[60]. A he ime o w i ing, his has now exceeded 300 000
pa ien s. The analyses a e pe o med only i s udies a e app o ed by he ele an au ho i ies (c .
sec ion 4).
13
CHAPTER 3. METHODS ON DIAGNOSTIC CODES
edeemed o e he s udy pe iod, P12 indica e ha p esc ip ion P1 was edeemed be o e p esc ip ion
P2 and P02 indica e ha only p esc ip ion P2 was edeemed in he s udy pe iod (Figu e 8).
Signi ican p esc ip ion pai s we e hen de ined by es ablishing a Poisson eg ession model o
each p esc ip ion pai . The numbe o people who had edeemed a p esc ip ion o P2 was he de-
penden a iable. The independen a iables we e age, sex, calenda yea and whe he a bina y
a iable indica ing whe he he s udy pa icipan had edeemed a p esc ip ion o P1 p io o e-
deeming P2. Like o he eg ession models, a Poisson eg ession can be sol ed using maximum
likelihood es ima es and signi icance can be assessed using he Wald s a is ics[66], [76]. The in-
e p e a ion o he RR was ha i exp esses he isk o edeeming a p esc ip ion o P2 a e P1
ela i e o people who had no edeemed he p esc ip ion o P1. Longe p esc ip ion ajec o ies
we e pieced oge he om signi ican p esc ip ion pai s by applica ion o he same p inciples as
desc ibed in sec ion 3.2.1.
3.3 Clus e analysis o de ine ischemic hea disease subg oups
This sec ion desc ibes a s a egy o s udying mul i-mo bidi y in IHD pa ien s ha is based on clus-
e analysis. In con as o he disease ajec o y app oach desc ibed in sec ion 3.2, he s udy o
mul i-mo bidi y based on clus e analysis only included mul i-mo bidi ies ha we e diagnosed be-
o e IHD onse . Sec ion 3.3.1 desc ibes he unde lying concep o applica ion o clus e analysis in
he con ex o mul i-mo bidi y used in his hesis. Nex , sec ion 3.3.2 desc ibes he ma hema ical
concep s ha clus e analysis elies on. Sec ion 3.3.3 desc ibes how clus e analysis was applied o
he en i e spec um o mul i-mo bidi ies p esen in an IHD coho , ha was iden i ied in NPR and
inally sec ion 3.3 exempli ies how he physiological ele ance o he clus e s was assessed.
3.3.1 A model whe e mul i-mo bidi y and ex documen s sha e p ope ies
The p emise o his model o mul i-mo bidi y was ha he se o diagnosis codes assigned o a pa -
icula pa ien in a coho can be conside ed a ex documen , whe e diagnosis codes co espond
o he wo ds in he documen . Analogously, he coho was conside ed a se o di e en ex docu-
men s (i.e. ex co pus), whe e he deg ee con en o e lap be ween he di e en documen s di e .
The idea is hen o g oup pa ien s based on he simila i y in diagnosis codes, jus like ex docu-
men s can be so ed wi h espec o hei heme based on he con en o wo ds. In his analogy,
di e en mani es a ions o mul i-mo bidi y among he pa ien s in he coho co espond o di e -
en ex hemes. This s a egy o s udying mul i-mo bidi y o igina es om in o ma ion e ie al
and has was p e iously adop ed in s udying diabe es subg oups[77].
In b oad e ms, in o ma ion e ie al is a discipline de ined by he p ocess o inding ma e ial
wi h a speci ic piece o in o ma ion om a la ge se o in o ma ion such as a da a se . Classical
in o ma ion e ie al asks e e o he ask o inding speci ic documen s om a la ge collec ion
which is usually s o ed on compu e s[78]. In he s udy o mul i-mo bidi y ha was based on hese
p inciples, he ini ial ask was o iden i y pa ien s (classically documen s) in a coho (classically a
ex co pus) ha sha ed ce ain cha ac e is ics. The idea was ha he esul ing pa ien subg oups
20

CHAPTER 3. METHODS ON DIAGNOSTIC CODES
could o m he basis o a mo e p ecise seg ega ion o high- and low- isk pa ien s. A ce ain isk
g oup would hen co espond o inding a speci ic piece o in o ma ion in classical in o ma ion
e ie al asks[78].
In classical in o ma ion e ie al asks a scaling ac o is o en in oduced o educe he e ec o
unspeci ic wo ds. Fo example, he e m "and" is an unspeci ic ea u e as i is ela i ely abundan
and is no e y app op ia e o disc imina ing be ween documen s. In he case o mul i-mo bidi y
in IHD pa ien s, a diagnosis code o hype ension could co espond o "and" in being e y equen
among IHD pa ien s bu no e y disc imina i e, i.e. la gely unin o ma i e.
The e m equency-in e se documen equency ( -id ) is an example o a scaling ac o ha
scales all ea u es acco ding o he numbe o occu ences (i.e., numbe o imes i is assigned o
a single pa ien ) and he o al numbe o occu ences in he en i e da a se . Thus, i dampens he
e ec o ela i ely unin o ma i e e ms[78]. Assuming ha he p inciples om classi ica ion o ex
documen s also apply o diagnosis codes in classi ica ion o pa ien s, he -id scaling ac o was
in oduced o educe he e ec o ela i ely unspeci ic diagnosis codes. Ma hema ically, he -id
scaling ac o is he p oduc o he e m equency and he in e se documen equency and can be
exp essed by
-id = ,d ·id (3.2)
whe e ,d is he no malized equency o a diagnosis code in he en i e coho and id is he in e se
o he a io be ween he numbe o pa ien s who a e assigned ha pa icula diagnosis code and he
o al numbe o pa ien s in he coho [77], [78]. Equa ions and u he de ails o compu a ion o
,d and id a e a ailable in equa ions A.2, A.3, and A.4 in appendix A.
3.3.2 A ma hema ical g aph ep esen a ion o ischemic hea disease pa ien s
A g aph is a ma hema ical objec ha can be used o ep esen a class o en i ies[79]. I is de ined
by a iple consis ing o a e ex se , an edge se and a ela ion ha associa es each edge wi h wo
e ices. I is no a equi emen ha he e ices a e dis inc as is he case in igu e 9, whe e all nodes
a e connec ed by exac ly wo e ices.
J
A B
C
D
E
F
G
H
I
Figu e 9: A g aph wi h e ices (ci cles), edges (lines) and nna u al g oups.
21
CHAPTER 3. METHODS ON DIAGNOSTIC CODES
To model mul i-mo bidi y in IHD, he Ma ko clus e (MCL) algo i hm was applied o diagnosis
codes om IHD pa ien s, ha concep ually co esponded o analyzing he wo ds o di e en ex
documen s in a ex co pus. Tha is, in his model IHD pa ien s co esponded o a class o en i ies.
The o e all aim was hen o es o na u al g oups wi hin his class, i.e, de ine mul i-mo bidi y
in a da a-d i en model by applica ion o an unsupe ised clus e ing algo i hm o he en i e se o
diagnosis codes assigned p io o IHD onse .
IHD pa ien s we e ep esen ed based on hei diagnosis codes and using a g aph whe e e ices
co esponded o pa ien s and edges ep esen ed pa ien s belonging o he same subg oup. Thus,
in he g aph ep esen ed in igu e 9, he e is one clus e as all he nodes a e ela ed ia a ini e
numbe o edges. Ye , i is also easonable o a gue ha some o he nodes a e mo e connec ed
han o he s, illus a ing ha he e may be na u al g oups wi hin he g aph. An example o such
a g oup could be he nodes labelled I, J and H as i appea s ela i ely de ached om he o he
nodes (Figu e 9). Many clus e ing algo i hms ha e been de eloped, likely e lec ing ha he no ion
o a na u al g oup is highly di e se; while also e lec ing ha clus e analysis is cu en ly being
applied in many di e en ields, including biology, communica ion, and compu e science[80]. In
sum and in b oad e ms, clus e analysis is conce ned wi h inding ways o look a he da a ha
enable a be e unde s anding o mechanisms a he han inding he co ec answe [81]. In he
case o mul i-mo bidi i y in IHD, clus e analysis is abou inding ways o look a he da a ha , o
example, allow o unde s and be e why symp om bu den and disease p og ession a y be ween
pa ien s.
3.3.3 A clus e algo i hm minimizes a cos unc ion
The ope a ional equi emen s o mos clus e ing algo i hms, including he MCL algo i hm a e (i)
ha he da a a e s uc u ed in a ma ix whe e da a poin s (obse a ions) a e ela ed o da a ea u es
( a iables) and (ii) ha he e is a measu e o simila i y (quali a i e ea u es) o dis ance (quan i-
a i e ea u es) be ween all da a poin s[82]. When clus e ing algo i hms a e applied o heal hca e
da a a ma ix is usually ha ing he indi idual pa ien s (obse a ions) as ows and di e en ea u es
(e.g. diagnosis codes) as columns [83], [84]. Be o e he ma ix can be used as inpu o a clus e ing
algo i hm, i is ypically op imized o be e ep esen he da a. In he s udy p esen ed in manusc ip
2, each pa ien was ep esen ed as a ec o o he -id scaled diagnosis codes.
The second ope a ional equi emen o a clus e ing algo i hm is ha he simila i y be ween ob-
se a ionsin he ma ixisquan i ied ei he usinga simila i yo dis ance unc ion, suchas he cosine,
Jacca d, and Hamming unc ions[80]. The cosine simila i y was compu ed o quan i y he simila -
i y be ween he -id scaled ec o s. I is de ined by an angle be ween wo ec o s which is equal o
he do p oduc o he wo ec o s di ided by he p oduc o hei leng hs was used
cos(θ) = xi•xj
||xi||·||xj|| (3.3)
whe e θis he angle be ween he wo ec o s xiand xj. Finally, he MCL clus e ing algo i hm was
applied o he ma ix ha ha ep esen ed he cosine simila i ies be ween he -id scaled pa ien
ec o s.
22
CHAPTER 3. METHODS ON DIAGNOSTIC CODES
Ma hema ically, he objec i e o he MCL algo i hm is o pa i ion a g aph by minimizing a cos
unc ion, ha is associa ed wi h he weigh o he edges. The algo i hm hen wo ks by iden i y-
ing he g aph pa i ion ha minimizes he cos unc ion. Concep ually i wo ks by s ochas ically
simula ing a sequence o andom walks o leng h kin a g aph, so simila egions in he g aph will
s ay connec ed, while dissimila egions will be pa i ioned in di e en clus e s. The algo i hm op-
imizes such ha he leng h o kwill minimize he cos unc ion. Tha is, he sequence o andom
walks will change he dis ibu ion o edges in a g aph, such ha e ices in dense egions (i.e. e-
gions wi h many edges) will s ay connec ed and e ices in spa se egions (i.e. egions wi h ewe
edges) will no s ay connec ed. The basic p emise is ha dense egions in he g aph will con ain el-
a i ely many pa hs o leng h k. Consequen ly, he e will be ela i ely many walks in dense egions
o he g aph and ewe in spa se egions. Acco dingly, ollowing simula ion and hose egions will
be conside ed simila [85].
Gene ally, addi ional da a sou ces a e necessa y in he subsequen clus e cha ac e iza ion, as a
compa ison be ween subg oups ob ained om clus e analysis comes wi h an ex eme Type I e o
in la ion a e, i.e. he isk o ejec ing a ue null-hypo hesis is high because he seemingly di e ence
is due o g aph pa i ion and hus a ci cula a gumen [86]. In e ec , pa ien subg oups ob ained by
applica ion o he MCL algo i hm we e cha ac e ized using biochemical and gene ic da a o e alua e
he biological ele ance o he pa e ns ecognized by he algo i hm. Ob iously, pa ien s who sha e
disease pa e ns do no necessa ily need o be gene ically simila . Tha will depend on he disease
in ques ion, e.g. polygenic and mul i ac o ial diseases.
3.3.4 Cha ac e iza ion o clus e s using su i al analysis
The aim wi h clus e analysis in he se ing o clinical esea ch is o en o iden i y dis inc subg oups
possibly cha ac e ized by di e en e iologies. In ou clus e -based s udy o mul i-mo bidi y in IHD,
he physiological ele ance o he clus e s was assessed by di e en means. One o he s a egies
was o assess i he e we e di e ences in isk o disease p og ession be ween pa ien s in he di e en
clus e s. Thus, clus e s we e used as co a ia es in a su i al model whe e seconda y ischaemic
e en s and dea h om non-IHD causes we e de ined as wo ou comes o in e es . In b oad e ms,
su i al analysis e e s o he ype o analysis whe e he ou come o in e es is ime- o-e en and is
ep esen ed using su i al cu es, ha desc ibe he ela ion be ween numbe o su i ed subjec s
a di e en ime-poin s (Figu e 10). When he e is mo e han one ou come o in e es , he ou comes
may be conside ed compe ing isks[87]. In he s udy p esen ed in manusc ip 2 he wo ou comes
we e ea ed as compe ing isks as only one o he ou comes could occu in a single pa ien . In he
s udy, clus e s we e also cha ac e ized using PGSs ha we e p epa ed om he aw gene ic da a in
CHB by collabo a o s in he au ho g oup as well as he esul s om blood es s. Since I was no
di ec ly in ol ed in hese analyses, hey a e no desc ibed in his sec ion.
One o he s eng hs wi h su i al analysis i ha in o ma ion om censo ed s udy pa icipan s
is included in he model. Censo ing e e s o ins ances whe e he exac su i al ime o a s udy
pa icipan is unknown. Classical examples o censo ing a e ha s udy pa icipan s a e ei he di-
agnosed be o e he obse a ion pe iod began o a e s ill ali e when he obse a ion pe iod ended.
Le -censo ing applies o si ua ions whe e he ue su i al ime canno be longe han he obse ed
23
CHAPTER 3. METHODS ON DIAGNOSTIC CODES
Figu e 10: Two su i al cu es and a isk able. Su i al cu es o pa ien subg oup A and B.
X-axis: Time om s a o ollow-up o e en , e.g. seconda y ischemic e en s. Y-axis: Su i al
p obabili y. Risk able: Numbe o pa ien s in each subg oup a isk a ime-poin s 0 h ough 5. The
numbe o pa ien s a isk a a gi en imepoin is he numbe o pa ien s whe e nei he he ou come
has occu ed no ha e hey been censo ed a he imepoin ha is being e alua ed. Unde lying da a
om s udy p esen ed in manusc ip 2.
su i al ime. Fo example, he e is le -censo ing i he ou come o in e es is ime un il disease on-
se and s udy pa icipan s a e no sc eened con inuously. In he case o NPR, mos diagnoses a e
le -censo ed i ime o diagnosis is used as a p oxy o disease onse . Explici ly, a pa ien who has
been assigned a diagnosis code o e.g. hype ension in NPR will mos likely ha e de eloped hype -
ension be o e he poin in ime whe e i is egis e ed in NPR. Con e sely, igh -censo ing applies
o cases whe e he ue su i al ime is minimum as long as he obse ed su i al ime. Hence, i
he ou come o in e es is dea h, da a will be igh -censo ed unless all s udy subjec s die du ing he
ollow-up pe iod. Righ -censo ing is he mos ypical ype o censo ing in su i al analysis[87].
The Cox p opo ional haza d (Cox PH) model is a semi-pa ame e ized model and among he
mos widely used models o analysis o su i al da a[88]. This was also he model o choice in
cha ac e izing he pa ien subg oups ob ained om applica ion o he MCL algo i hm. A cen al
componen o Cox PH is he baseline haza d unc ion, also e e ed o as he condi ional ailu e a e.
Cox PH is semi-pa ame e ized because he baseline haza d unc ion is ne e speci ied[89]. Ins ead
i is assumed ha he baseline haza d is independen o he co a ia es also coined he p opo ional
haza d assump ion[87]. The Cox PH unc ion is ep esen ed by he p oduc o wo e ms, whe e
24
CHAPTER 3. METHODS ON DIAGNOSTIC CODES
he i s e m is he baseline haza d unc ion and he second e m ep esen s a linea combina ion
o he explana o y a iables. Ma hema ically i is gi en by he o mula
h( , X) = h0( )·e
p
P
i=1
βiXi(3.4)
whe e h0is he baseline haza d unc ion and Xi ep esen s he explana o y a iables. The baseline
haza d unc ion can be ep esen ed by a su i al cu e (Figu e 10). The common in e p e a ion is
ha he haza d unc ion yields he isk which is quan i ied as he haza d in his analy ical ame-
wo k[87]. The haza d a a ce ain poin in ime (e.g. n) is equal o he chance o su i al a 0
gi en he s udy pa icipan has su i ed he du a ion o ime n. Haza ds a e commonly exp essed
as haza ds a ios (HRs) in compa ing o ins ance a ea men and a placebo g oup[89].
3.4 Da a-d i en p edic ions o mo ali y in ischemic hea disease
This chap e explains he p inciples o ANNs, which a e in oduced in his hesis as a me hod o
analyzing ime- o-e en da a based on hund es o inpu ea u es. Sec ion 3.4.1 in oduces he concep
o ANNs and sec ion 3.4.2 p esen s a model o explainabili y ha esponds o some gene al c i ique
agains he usage o ANNs and o he machine lea ning me hods in clinical esea ch.
3.4.1 A concep ual in oduc ion o a i icial neu al ne wo ks
O iginally, ANNs and he uni s hey con ain we e de eloped o model in o ma ion p ocessing in
he b ain[90]. Acco dingly, ANNs a e buil o neu al uni s called a i icial neu ons. Neu ons a e
ypically desc ibed using he pa ame e s weigh s and a bias e m. In his con ex a bias is a h esh-
old and de e mines how easy i is o he neu on o p oduce a ce ain ou pu gi en i s inpu [91].
Fo mally, he o al, weigh ed inpu o a neu on in laye ixican be exp essed as he weigh ed sum
o incoming ou pu s om he p e ious laye
xi=X
j∈N−(i)
wij ·yj+wi(3.5)
whe e wij ·yjco esponds o weigh ed sum o he inpu s om he p e ious laye and wiis he bias
e m o a neu on in laye i. The weigh ed sum o inpu s om he p e ious laye is hen passed
h ough a non-linea ac i a ion unc ion, such as he sigmoid shaped logis ic unc ion,
1
1 + e−xi (3.6)
o p oduce he ou pu o a neu on in lay e i ha se es as inpu o he nex laye [92].
Mul i-laye pe cep ons a e a class o ANNs ha s uc u ally a e desc ibed by e e ing o h ee
ypes o laye s: he inpu laye ha ecei es in o ma ion o be p ocessed, he ou pu laye ha de-
li e s he compu a ional esul , and in be ween, one o mo e hidden laye s (Figu e 11). The hidden
25

CHAPTER 3. METHODS ON DIAGNOSTIC CODES
and ou pu laye s a e all comp ised o a se o a i icial neu ons ha p ocess he da a om he p e-
ceding laye . The model a chi ec u e o an ANN e e s o e.g. he numbe o hidden laye s in a ANN
e med he dep h o he ANN and also he ela ion be ween he neu ons in he di e en laye s[92].
Inpu
Ou pu
Hidden laye s
Figu e 11: Schema ic ep esen a ion o an ANN wi h wo hidden laye s. A i icial neu ons a e ep-
esen ed as boxes and colo s indica e i he neu ons belong o he inpu laye , ou pu laye , o hidden
laye s. Inpu neu ons a e ligh blue (bo om). Ou pu neu ons a e g ey ( op). ANN: A i icial neu-
al ne wo k. Adap ed om manusc ip 3.
A de ining p ope y o ANNs is ha hey lea n om examples du ing de elopmen as weigh s
and bias e ms a e uned g adually in esponse o ex e nal s imuli (i.e. inpu s). Tha is, ANNs o
his kind lea n om examples[92]. Be o e de eloping an ANN, a da a se is ypically spli in o a
aining se and a es se used o in e nal and ex e nal alida ion, espec i ely[92]. The o me is
used o in e nal alida ion o quan i y and es ima e he pe o mance o he model. The es se is
used o assess i he ne wo k wo ks well on no el da a o has been o e - i ed on he aining se (c .
sec ion 3.4.2).
One o he mos widely applied lea ning algo i hms is s ochas ic g adien descen (SGD), which
is an algo i hm ha minimizes he cos unc ion which is e med loss o objec i e unc ion. In supe -
ised lea ning, he cos unc ion measu es he di e ence be ween he ac ual ou pu and he desi ed
ou pu [91], [93]. In p inciple, SGD sol es he cos unc ion by iden i ying he global minimum. In
o he wo ds, he SGD wo ks by minimizing he cos unc ion. To a oid o e - i ing, egula iza ion
o he da a is o en pe o med p io o model aining. Regula iza ion wo ks by adding a penal y
e m o he cos unc ion which ensu es ha i is no oo complica ed and hence p one o emembe
he da a a he han lea ning some gene alizable pa e ns in he da a[94].
26
CHAPTER 3. METHODS ON DIAGNOSTIC CODES
Technically, SGD wo ks by epea edly applying he so-called upda e ule o weigh s and biases
in a gi en ne wo k, which is exempli ied o he weigh wkhe e
wk→w0
k=wk−ηδC
δwk
(3.7)
whe e C deno es he cos unc ion and he change in he alue o a weigh wk o w0
k e lec s es ing i
he cos unc ion dec eases when changing a weigh . The upda e ule hen es s i he cos unc ion
C dec eases when changing a weigh wkas indica ed by wk→w0
k. The e m η e e s o he lea ning
a e and is o en explained as a "s ep-size". S ep-size and ANN dep h a e ye o he examples o so-
called hype pa ame e s and ANN de elopmen is in pa abou op imizing he hype pa ame e s o
educe he cos unc ion.
3.4.2 Ranking inpu ea u es based on le el o in o ma ion
A common c i ique agains he usage o ANN in mode n medicine is ha hey appea as black boxes
in he sense ha an ANN does no p o ide an explana ion o why a ce ain inpu leads o a ce ain
ou pu , o example why he isk o dea h o a pa ien is es ima ed o be a ce ain alue[95]. An
ob ious coun e a gumen is ha clinical decisions simila ly appea as black boxes as hey do no
always gene alize be ween cases. Ano he a gumen in a o o ANNs is ha se e al me hods o
explaining he ou pu o any machine lea ning model, including ANNs now exis [96]. This aspec
o machine lea ning is called in e p e abili y o explainabili y. An example o such me hod is he
so-called SHapley Addi i e exPlana ions alue (SHAP alue) named a e he Nobel P ize-winning
economis Lloyd S owell Shapley[97].
The unde lying idea o he SHAP alue is bo owed om coo po a i e game heo y. He e he
Shapley alue is a me hod ha assigns payou s o playe s depending on hei con ibu ion o he
o al payou [98]. Analogously, he SHAP alue can be used o measu e he indi idual ea u e im-
po ance in an ANN, based on he in o ma i e alue o ha ea u e. The idea is ha changing an
in o ma i e ea u e will also change he ou pu conside ably; whe eas he opposi e is ue o less
in o ma i e ea u es[99].
SHAP alues e e o he esul s o di e ences be ween he ou pu om he ac ual (baseline)
model (i.e. an ANN), and he ou pu o a new, explana o y model. The SHAP alue o a gi en
inpu ea u e measu es how much ha ea u e con ibu ed o he model ou pu in he con ex o i s
in e ac ionwi h heo he ea u es. So, he SHAP alue o a gi en ea u e iscalcula ed by compa ing
he ou pu o he ac ual model wi h he ou pu o a model wi hou ha speci ic ea u e. The SHAP
alue is he di e ence be ween hese wo ou pu s. Thus, SHAP alues a e ela i e o he inpu
da a and can be used o in e p e he model globally as well as in e p e a ion o he indi idual
p edic ions[99]. An impo an p ope y o SHAP alues is ha hey a e addi i e, meaning ha he
impo ance o se e al inpu ea u es can be pooled[100].
27
CHAPTER 3. METHODS ON DIAGNOSTIC CODES
Aside om his kind o explainabili y analysis, model pe o mance quan i ied using calib a ion
and disc imina ion a e also ele an in e alua ing a su i al model based on an ANNs[101], [102].
Model calib a ion e e s o he abili y o he model o p edic u u e ou comes compa ed o he
obse ed ou comes, whe eas model disc imina ion e e s o a measu e o he abili y o he model
o sepa a e ou comes. Tha is, a model can be well-calib a ed and a he same ime ha e a poo
disc imina ion i i p edic s eigh ailu es wi hin a ime pe iod bu he p edic ed subjec s who ail
do no co espond o he obse ed subjec s who ail[103]. The classi ica ion pe o mance, i.e. dis-
c imina ion o he algo i hm de eloped in manusc ip 3, was e alua ed using he ime-dependen
a ea unde he ecei e ope a ing cha ac e is ic cu e ( dAUC). The ecei e ope a ing cha ac e -
is ic cu e desc ibes he adeo s be ween bene i s and cos , i.e. ue posi i es and alse posi i es,
whe e he goal is o maximize he o me and minimize he la e [101], [104].
3.5 A p obabilis ic map o in-hospi al d ug dosage al e a ions
Bayesian hie a chical logis ic eg ession was applied o es ablish a model o polypha macy, whe e
es ima ed likelihoods o d ug-dosage al e a ions we e used as p oxies o po en ial d ug-d ug in-
e ac ions. Since he me hod ep esen s a s a is ical pa adigm ha is undamen ally di e en om
he equen is pa adigm cen e ed on he null hypo hesis, his chap e is dedica ed o in oducing
he undamen als o Bayesian s a is ics. Sec ion 3.5.1 desc ibes he componen s ha comp ise a
Bayesian eg ession model and sec ion 3.5.2 gi es an o e iew o he applica ion o such models in
he con ex o EHRs.
3.5.1 Bayesian models can be decomposed in o h ee dis ibu ions
A de ining a ibu e o Bayesian s a is ics is ha a join p obabili y dis ibu ion is gi en o all ob-
se ed andunobse ed a iables[105]. This con as s o a equen is pa adigm, whe ei is assumed
ha he e ec o unobse ed a iables is cons an . Consequen ly, Bayesian models a e lexible and
addi ion o da a o he model may change he model pa ame e s. Gene ally, Bayesian models a e
gaining a en ion due o an inc easing amoun o so wa e ha has been de eloped o pe o m he
eg ession[106]. Concep ually, a Bayesian model quan i ies a deg ee o (un-)ce ain y, which can be
upda ed upon addi ion o new da a, i.e. e idence[107]. Bayesian models can be decomposed in o
h ee pa s ha may be p obabili y dis ibu ions[105]. These a e a p io dis ibu ion, a likelihood
dis ibu ion and a pos e io dis ibu ion (Figu e 12).
Bayesian s a is ics is oo ed in Bayes’ heo em om which a condi ional p obabili y can be de i ed
wi hou knowledge o he join p obabili y. Ma hema ically Bayes’ heo em is exp essed by
p(B|A) = p(A|B)p(A)
p(B)(3.8)
whe e p(B|A)exp esses he p obabili y o B, gi en A. The heo em s a es ha his p obabili y is
equal o he p oduc o he p obabili y o A gi en B and he p obabili y o A di ided by he p oba-
bili y o B. In o he wo ds, Bayes’ heo em s a es ha he p obabili y o B gi en A is p opo ional o
28
CHAPTER 3. METHODS ON DIAGNOSTIC CODES
Da a
Likelihood P io s
Bayes’
heo em
Pos e io
∼
Figu e 12: A schema ic ep esen a ion o a Bayesian logis ic eg ession model. Uppe ow ep e-
sen s he h ee di e en dis ibu ions. Middle ow ep esen s a Bayesian logis ic eg ession, ha is
he ounda ion o ob aining he join dis ibu ion also coined he pos e io dis ibu ion, which is
ep esen ed in he bo om ow. Adap ed om [107].
he p obabili y o A gi en B, i.e.
p(B|A)∝p(A|B)p(A)(3.9)
As indica ed in igu e 12, a Bayesian logis ic eg ession model can be es ablished by applica ion o
Bayes’ heo em o a da a se , whe e he likelihood p(A|B)is he binomial (n, θ)and θis equal o
he sum o he co a ia es. The p io e lec s he le el o in o ma ion ha is assumed o be in he
da a. I co esponds o p(A)in Bayes’ heo em and ep esen s he p io belie s abou he ela ion
be ween a se o a iables and he ou come o in e es . I he da a is ep esen a i e o he popula-
ion ha is being modelled, he p io will ha e a small a iance. Con e sely, he p io will ha e a
la ge a iance in cases whe e o example he da a se is ela i ely small and hus no likely o be
s ongly ela ed o he ou come o in e es . Gene ally, p io s a e cha ac e ized based on hei le el
o in o ma ion anging om in o ma i e (small a iance) o di use o weakly in o ma i e (la ge
a iance). Impo an ly, he p io s con ain model subjec i i y and a e selec ed by he in es iga o
al hough di e en p io s can be compa ed[105], [108].
In a Bayesian logis ic eg ession model, he goal is o model a pos e io dis ibu ion ha is as close
o he ue pos e io as possible. The op imiza ion can be pe o med using di e en algo i hms,
e.g., Ma ko Chain Mon e Ca lo sampling (MCMC)[109]. The algo i hm es ima es he pos e io
29
CHAPTER 5. RESULTS ON DIAGNOSTIC CODES
Table 2: Summa y s a is ics o selec ed disease ajec o ies.
Popula ion ICD-10 codes Mean age in yea s Coun s Adj. P RR
D1 D2 D3 D1 D2 D3
All
E78 I21 60.8 62.9 76 334 <0.001 3.69
E11 I20 61.0 63.0 61 333 <0.001 1.56
E11 I50 65.6 70.4 42 023 <0.001 2.82
I20 I50 68.0 71.4 84 955 <0.001 1.94
E78 I20 I21 61.1 61.6 62.5 48 444
E11 I20 I50 64.2 65.6 69.3 23 431
I21 I21 I48 73.3 76.3 11 871 <0.001 1.09
I20, I21, I25 I48 I21 69.4 71.2 26 273 <0.001 1.09
F om manusc ip 1.
5.2 A s a egy o de ine mo e homogeneous ischemic hea disease
pa ien subg oups
T adi ionally, mul i-mo bidi y has been quan i ied wi h e e ence o a ew selec ed diseases and
summa ized in an index such as CCI. Some e en a gue ha he e does no exis any eplicable p o-
iles o mul i-mo bidi y[27], [123]. To accommoda e hese limi a ions in u u e mul i-mo bidi y
classi ica ion sys ems, i has been a gued ha i is necessa y o s udy i comp ehensi ely, a he
han pu ely assessing he p esence o indi idual diseases in a bina y ashion[9]. S udies ha clas-
si y ischemic hea disease pa ien s as well as hea ailu e pa ien s based on applica ion o clus e
analyses o EHRs da a ha e al eady been conduc ed[77], [124], [125]. Ye , clus e analysis in his
domain is s ill ela i ely unexplo ed whe e he analyses a e o en limi ed o ew diagnoses and hus
only analyze a ac ion o he a ailable heal hca e da a olumes. Thus, he e is a need o explo e he
po en ial o EHRs in his domain.
To his end, he s udy p esen ed in manusc ip 2 add esses mul i-mo bidi y in IHD based on a
lexible clus e ing analysis o mul i-mo bidi y by applica ion o he MCL algo i hm o he diagnosis
codes o 74 249 IHD pa ien s iden i ied in NPR (Figu e 15). The physiological ele ance o he
clus e s (i.e. pa ien subg oups) ob ained by applica ion o he MCL algo i hm was assessed based
on a an en ichmen analysis wi h espec o he diagnosis codes, a compa ison o he subg oups
ha included es ima es o isk o disease p og ession; and cha ac e iza ion de i ed om blood
es esul s and PGS ob ained om EHRs and CHB, espec i ely.6
Applica ion o he MCL algo i hm o he en i e se o diagnosis codes assigned p io o IHD onse
esul ed in iden i ica ion o high- and low- isk pa ien subg oups. Pa ien subg oups we e classi ied
as high- o low- isk based on su i al analysis. Fo example, among h ee o he male subg oups
6Fo an in-dep h desc ip ion o he en ichmen analyses as well as analyses o he biochemical and gene ic da a please
e e o he ull-leng h manusc ip a ailable in chap e 8.
36

CHAPTER 5. RESULTS ON DIAGNOSTIC CODES
Figu e 15: Concep ual igu e o he pa ien ep esen a ion in he clus e analysis. The en i e coho
is illus a ed o he le and he ec o s o i e pa ien s a e illus a ed on he igh . A sample o
diagnosis codes a e lis ed ho izon ally in he uppe igh . Black ci cles indica e ha diagnoses we e
assigned be o e IHD onse (included in clus e analysis). Hollow ci cles indica e diagnoses which
we e assigned a e IHD onse (no included in clus e analysis). 0: IHD onse . IHD: Ischemic
hea disease. Adap ed om manusc ip 2.
( o al o 31), mean age a IHD onse anged be ween 54.9and 62.4yea s. These h ee pa ien sub-
g oups we e also cha ac e ized by di e en isks o disease p og ession, as he pa ien s in one sub-
g oup associa ed wi h educed isk o p og ession (i.e. p ima y ou come) when compa ed o he
he isk o he pa ien s who we e no in his subg oup. In con as , he wo o he subg oups asso-
cia ed wi h inc eased isk o p og ession (Table 3). The esul s o he en ichmen analysis showed
ha one o he subg oups cha ac e ized by inc eased isk o disease p og ession was en iched o
diabe es whe eas he o he was en iched o mo e unspeci ic diagnoses such as dias asis o he mus-
cle. In e es ingly, he subg oup cha ac e ized by educed isk o disease p og ession was en iched
o acu e MI. The dis ibu ions o PGS o o al choles e ol (TC) and LDL choles e eol le els we e
signi ican ly di e en in his subg oup wi h posi i e e ec di ec ions (Table 3). Tha is, pa ien s in
his subg oup gene ally had highe PGS o TC and LDL le els han he pa ien s who we e no in
his subg oup.
In sum, he s udy p esen ed in manusc ip 2 exempli ied ha applica ion o he MCL algo i hm o
he en i e mul i-mo bidi y spec um o IHD pa ien s can be used o disc imina e be ween high- and
low- isk IHD pa ien s. Fu he , he s udy illus a es how, po en ially, mo e homogeneous pa ien
subg oups can be used o iden i y subg oups wi h gene ic associa ions whe e PGS ha e posi i e
o nega i e e ec sizes (Table 3). In compa ison o o he clus e analyses o ca dio ascula phe-
no ypes, we de eloped a lexible app oach ha analyze inpu da a in a da a-d i en manne based
on he analogy o ex documen s[125]. Fu he , he cha ac e iza ion o he clus e s in ol ed addi-
ion o da a sou ces and hus p esen s a p agma ic solu ion o educe he impac o an in la ed ype
I e o a e[86], [124]. Po en ially he me hods p esen ed in his s udy can be u he op imized
and ou pe o m adi ional mul i-mo bidi y sco es. In i s cu en o m, he disc imina ion o he
subg oups is equally as good as CCI o e all and in some subg oups, i is be e . The esul s a e
p esen ed mo e elabo a ely in he ull-leng h manusc ip , which is a ailable in chap e 8.
37
CHAPTER 5. RESULTS ON DIAGNOSTIC CODES
Table 3: Summa y s a is ics and cha ac e is ics o selec ed male clus e s.
Clus e Mean age HR, p ima y ou come Main en ichmen PGS
Yea s a index HR 95%CI Adj. P (e ec di ec ion)
4.M01 58.5 0.93 0.87 o 0.99 0.030 Acu e myoca dial TC(+), I25(-)
in a c ion LDL(+)
4.M02 62.4 1.51 1.40 o 1.62 <0.001 Diabe es E11(+)
4.M25 54.9 1.38 1.15 o 1.65 0.030 Dias asis o muscle
F om manusc ip 2.
5.3 Mo ali y in ischemic hea disease om a ime- o-e en ma-
chine lea ning algo i hm
Today, IHD pa ien p ognos ica ion is being ca ied ou wi h isk p edic ion ools ha la gely ha e
emained unchanged o decades[83]. While he olumes and a ie y o da a a e inc easing in com-
plexi y, isk sco ing algo i hms a e commonly ex apola ed om classical eg essions ha model
linea ela ionships be ween a limi ed se o well-cha ac e ized p edic i e ea u es[11]. Machine
lea ning me hods, such as ANNs comp ise a po en ial me hodological solu ion o he ac ha he
inc easingly complex heal hca e da a equi e analy ical capabili ies ha o en exceed ha o he
human mind[95]. The e a e al eady p omising esul s o pa ien isk p edic ion models de eloped
using machine lea ning me hods wi hin he ca dio ascula domain[126], [127]. Howe e , he e a e
s ill no examples o isk p edic ions models in IHD whe e inpu ea u es a e being included based
on a ailabili y a he han assump ions ega ding hei p edic i e alue.
The e o e, he s udy p esen ed in manusc ip 3 in oduces PMHne -alpha, an ANN-based su -
i al model de eloped by applica ion o a gene ic app oach ha was o iginally desc ibed by Gen-
sheime e al[128]. Based on a de i a ion coho iden i ied in EDHR, PMHne -alpha was de e-
loped op edic all-cause mo ali y in IHD pa ien s om 596 inpu ea u es (Figu e16). The p emise
o he s udy was ha he summa ized p edic i e alue o many ea u es wi h a ying p edic i e
alue could comp ise a p edic ion model wi h highe pe o mance han he summa ized p edic i e
alue o ewe , highly p edic i e ea u es.
A e model de elopmen , PMHne -alpha achie ed excellen pe o mance and ou pe o med
he GRACE 2.0 isk model in alida ion wi h a dAUC o 0.869 a he one-yea e alua ion ime poin
(Table 4). Using SHAP alues, we we e able o quan i y he p edic i e impo ance o all 595 inpu
ea u es ob ained om a combina ion o he di e en da a ypes (Figu e 16). The SHAP alues
enabled us o disc imina e be ween he impo ance o di e en inpu ea u es a a global and a
a speci ic le el. In a model based solely on diagnosis codes (ICD-10 codes), he diagnoses wi h
he la ges p edic i e alues globally we e diabe es (E11), acu e myoca dial in a c ion (I21), a ial
ib illa ion (I48), hea ailu e (I50) and ch onic obs uc i e pulmona y disease (J44). All o hem
educed he chances o su i al. In con as , he p esence o angina pec o is (I20), diso de s o
lipop o ein me abolism and o he lipidaemias (E78), open wound o w is o hand (S61) and ou-
ine examina ion o speci ic sys ems (Z01) inc eased he chances o su i al. In he model ha was
ained on all 595 inpu ea u es, age was, no su p isingly, he mos p edic i e a iable. In e es -
38
CHAPTER 5. RESULTS ON DIAGNOSTIC CODES
Figu e 16: O e iew o he de elopmen o PMHne -alpha. A: De i a ion coho whe e he inpu
ea u es o ou pa ien s a e illus a ed as indica ed by he ou dashed lines. B: The di e en da a
ypes ha we e used as inpu ea u es. C: A schema ic ep esen a ion o he PMHne -alpha a chi-
ec u e. D: PMHne -alpha ou pu he e illus a ed by su i al cu es o ou indi idual pa ien s.
F om manusc ip 3.
ingly, only oponins we e among he mos p edic i e ea u es in e alua ion o he pe o mance a
six mon hs a e index, whe eas o he biochemical es s aken be o e IHD onse we e only among
he mos p edic i e ea u es a he one-yea and h ee-yea s e alua ion imepoin . PMHne -alpha
is cu en ly being alida ed a pa ne hospi als in No way and Iceland, espec i ely.
Table 4: Valida ion and benchma king o PMHne -alpha.
Ou come Wi hou impu a ion Wi h impu a ion
PMHne -alpha GRACE 2.0 PMHne -alpha GRACE 2.0
Six-mon hs all-cause mo ali y 0.876 0.770 0.883 0.770
One-yea all-cause mo ali y 0.869 0.782 0.872 0.756
Th ee-yea s all-cause mo ali y 0.843 0.738 0.831 0.727
F om manusc ip 3.
Al oge he , PMHne -alpha exempli ies how da a-d i en s udies can imp o e he pe o mance
o isk p edic ion bo h by analyzing non-linea ly in e ac ing ea u es and also by including hun-
d eds o ea u es. Mo eo e , he s udy p esen s a solu ion o he common c i ique agains machine
39
CHAPTER 5. RESULTS ON DIAGNOSTIC CODES
lea ning me hods as PMHne -alpha is an explainable model. Fo example, he SHAP analysis
p o ided a amewo k o unde s anding why PMHne -alpha based on diagnosis codes only had
a simila pe o mance o ha o he GRACE 2.0 isk sco e. In con as o o he machine lea ning
models wi hin he ca dio ascula domain, PMHne -alpha was de eloped wi h essen ially no a
p io assump ions ega ding po en ial p edic i e alue[126]. We a gue ha his is an ad an age in
a ou o PMHne -alpha as he a ibu e makes i b oadly applicable and p esumably adap able
in o many di e en heal hca e sys ems whe e da a ypes and a ailabili y a y. This a gumen is
p esen ed mo e ho oughly in he ull-leng h manusc ip a ailable in chap e 9, which is cu en ly
being p epa ed o submission.
5.4 E idence o d ug-d ug in e ac ions in elec onic heal h eco ds
Mul i-mo bid pa ien s, including IHD pa ien s, a e commonly subjec ed o polypha macy which is
o en oo ed in he simul aneous p esc ip ions o li e-long ea men egimens[31]. Polypha macy
pu s pa ien s a isk o ad e se d ug eac ions and educes he likelihood o compliance. T adi ion-
ally, i has p ima ily been quan i ied wi h e e ence o he numbe o d ugs only and d ugs a e a ely
p esc ibed o pa ien s who a e diagnosed wi h only one ch onic disease[30], [32]. In e ec , e idence
ega ding he ue he apeu ic e ec o d ugs in mul i-mo bid pa ien s is lacking as hey a e ypi-
cally excluded om RCTs[9], [30]. This is pa icula ly impo an in he con ex o IHD, whe e he e
a e a weal h o di e en op ions o d ugs used in p ima y and seconda y ea men [129].
Figu e 17: Concep ual igu e illus a ing concomi an ea men episodes du ing a hypo he ical hos-
pi al admission. Th ee di e en d ugs a e depic ed. D ug A (blue) is p esc ibed in wo di e en
egimens. B ug B (pink) is p esc ibed in h ee di e en egimens. D ug C (g een) is p esc ibed
one ime. Th ee di e en modi ica ions we e conside ed: D ug addi ion, dosage adjus men , and
d ug discon inua ion. ADD: A e age daily dose. D/C: Discon inua ion. F om manusc ip 4.
40
CHAPTER 5. RESULTS ON DIAGNOSTIC CODES
The objec i e o he s udy p esen ed in manusc ip 4 was o analyze polypha macy wi h e e ence
o d ug dosage al e a ions based on in-hospi al p esc ip ions om 1 069 873 hospi alized pa ien s
in he pe iod yea s 2008 o 2016. D ug dosage al e a ions we e analyzed sys ema ically by iden i-
ying concomi an ea men episodes, de ined by a poin in ime whe e a pa ien was p esc ibed
wo di e en d ugs. Concomi an ea men episodes wi h a signi ican likelihood o dosage ad-
jus men s we e de ined as co-medica ion pai s comp ised o an index-d ug and a co-d ug (Figu e
17). Co-medica ion pai s we e cha ac e ized speci ically wi h e e ence o selec ed examples and
also collec i ely by c oss- e e encing he esul s wi h da a om 15 publicly a ailable d ug-d ug
in e ac ion da abases[130].
A o al o 920 dis inc d ugs adminis e ed o e a pe iod o eigh yea s ga e ise o iden i ica ion
o 77 494 concomi an ea men episodes. In he da a, he e we e a o al o 3993 co-medica ion
pai s. Index d ugs we e domina ed by d ugs belonging o he ana omical classes ne ous sys em
(ATC 1s le el: N), ca dio ascula sys em (ATC 1s C), and alimen a y ac and me abolism sys em
(ATC 1s A). D ugs classi ied as an i h ombo ic agen s (ATC 2nd le el: B01) and ACE inhibi o s
(ATC 2nd le el: C09) we e among he d ugs ha mos equen ly appea ed as index d ugs. They
we e ep esen ed in 316 (7.9%) and 155 (3.8%) o he co-medica ion pai s, espec i ely (Table 5).
As expec ed, ace ylsalicylic acid and clopidog el comp ised an example o a co-medica ion pai .
Thus, when ace ylsalicylic acid was co-adminis e ed wi h clopidog el a dosage change in ace ylsal-
icylic acid was 2.24 imes mo e likely han in cases whe e ace ylsalicyclic acid was adminis e ed as
mono he apy. Simila ly, clopidog el was 1.52 imes mo e likely o be dosages adjus ed when co-
adminis e ed wi h ace ylsalicylic acid. Mo eo e , a o as a in had an odds a io o 1.31 o a dosage
al e a ion when co-adminis e ed wi h clopidog el. Thus he co-medica ion pai s ep esen ed ex-
pec ed as well as mo e su p ising examples o dosage adjus men s (Table 5).
Table 5: Examples o co-medica ion pai s om highly ep esen ed he apeu ic subg oups.
O e iew Selec ed co-medica ion pai s
ATC Index Co-d ug ATC is ana omical le el o index d ug.
2nd N N Index d ug Co-d ug Odds a io 95% HDI
A02 105 176 Pan op azole Te lip essin 2.26 1.88 o 2.79
B01 316 314
Ace ylsalicylic acid Clopidog el 2.24 2.03 o 2.49
Clopidog el Ace ylsalicylic acid 1.52 1.39 o 1.67
Enoxapa in Li aglu ide 1.78 1.23 o 2.61
C07 116 137 Me op olol Ri ampicin 1.45 1.09 o 1.94
C09 155 150 Losa an Que iapine 1.32 1.10 o 1.57
C10 111 129 A o as ain Clopidog el 1.31 1.18 o 1.44
N07 56 52 Disul i am Pan op azole 1.20 1.05 o 1.37
F om manusc ip 4.
O he 3993 co-medica ion pai s, he e we e 3297 pai s ha we e al eady en ies in a d ug-d ug
in e ac ion da abase. We a gue ha he emaining 696 d ugs ep esen po en ial d ug-d ug in e -
ac ions, gi en ha he dosage adjus men s a e no oo ed in clinical ou ines o guideline ecom-
menda ions wi h o he ea ionals han pu e d ug-d ug in e ac ions. O e all, he s udy p esen ed
in manusc ip 4 exempli ies he po en ial impo ance o da a-d i en s udies in his domain. The
s udy adds o he li e a u e by (i) p esen ing a model o ac ual in-hospi al ea men pa e ns in
41

CHAPTER 5. RESULTS ON DIAGNOSTIC CODES
a popula ion ha is ep esen a i e o pa ien s hospi alized in Denma k and (ii) by cap u ing a e
he apeu ic d ug combina ions ha a e no necessa ily cap u ed in smalle s udies o polypha macy.
In his espec , ca dio ascula d ugs gene ally se ed as good posi i e con ols du ing model in e -
p e a ion bo h wi h espec o he quan i y and quali y, co esponding o he numbe con en o he
co-medica ions pai s. The esul s o he s udy a e p esen ed and discussed in g ea e de ail in he
ull-leng h manusc ip which is a ailable in chap e 10.
5.5 Es ablishing a map o longi udinal, na ion-wide p esc ip ion
pa e ns
Mos o he d ugs ha a e p esc ibed as p ima y o seconda y p e en ion in ea men o IHD and
ela edmul i-mo bidi ya ebeing p esc ibed li e-long[30]. Al hough he numbe o pa ien sneeded
o ea o expec o p e en one ad e se e en a e high; he p ocess o dep esc ip ion is no s anda d
in clinical ca e[131], [132]. Ye , he e is only limi ed knowledge o he long- e m consequences o
polypha macy, which is becoming e en mo e p e alen in he ageing popula ion[75]. Albei hese
ac s, e y ew s udies ha e analyzed he long- e m consequences o d ugs used o ea mul i-
mo bidi y, including IHD[30], [32]. Based on p esc ip ion ajec o ies es ablished om 7 255 919
people in DNPR, he s udy p esen ed in manusc ip 5 mapped longi udinal p esc ip ion pa e ns
in an en i e Danish popula ion o e he yea s 1995 o 2019.
This s udy ans o med he indi idual le el p esc ip ions in DNPR in o 38 607 p esc ip ion ajec-
o ies o wo dis inc d ugs ha we e mos likely o be p esc ibed in a speci ic o de . Fo example,
he RR o edeeming digi alis glycoside a e a be a blocking agen was 3.55 (95% CI: 2.30 o 5.47,
P<0.001) compa ed o pa ien s wi h no p io edemp ion o a be a-blocking agen . This obse a-
ion is consis en wi h in ensi ica ion o a hy hmia ea men and hus se ed as a posi i e con ol.
We also ound ha edemp ion o di ec ac o Xa inhibi o s associa ed wi h a educed isk o sub-
sequen edemp ion o o he an i-demen ia d ugs (RR: 0.41,95% CI: 0.37 o 0.46, P <0.001). This
associa ion be ween an i-coagula i e he apy exempli ies ha da a-d i en s udies can assis in iden-
i ying associa ions ha a e unlikely o be iden i ied in pu ely hypo hesis-d i en s udies. Po en ially
his speci ic p esc ip ion ajec o y e en poin s o a ole o an i h ombo ic agen s in p ima y p e en-
ion o s oke. P esc ip ion ajec o ies ha con ained wo p esc ip ions we e also pieced oge he
o o m p esc ip ion ajec o ies o up o se en di e en p esc ip ions. We used his s a egy o map
he o de o an i-hype ensi e ea men s o pa ien s being p esc ibed up o se en di e en d ugs
wi hin his d ug g oup (i.e. an i-hype ension d ugs). Fo example, only in e y ew cases did peo-
ple who we e p esc ibed an aldos e one an agonis s (C03DA) edeem a p esc ip ion on ano he
an i-hype ensi e d ug a e wa ds. And his end seemed o be i espec i e o he p e ious e-
deemed p esc ip ions as he e we e ew p esc ip ion pai s on he igh side o boxes labeled C03DA
(Figu e 18).
Taken oge he , he p esc ip ion ajec o ies modelled polypha macy in a no el manne and sup-
po ed analyses o up o six di e en an i-hype ensi e d ugs p esc ibed o he same pe son o e
he cou se o 25 yea s (Figu e 18). The esul s indica e ha he e is a conside able oom o im-
p o emen when i comes o decisions ega ding ea men egimens. I is o cou se impo an o
42
CHAPTER 5. RESULTS ON DIAGNOSTIC CODES
Figu e 18: Allu ial plo illus a ing he complexi y o an i-hype ensi e ea men . Each e ical ba
ep esen s an ATC le el 4code con aining an i-hype ensi e d ugs. G ey ibbons ha connec he
e ical ba s ep esen a p esc ip ion ajec o y comp ised o wo p esc ip ion, whe e he RR isk o
edeeming he d ug on he le i s is highe han edeemed he d ug on he igh i s . The heigh
o he e ical ba s co esponds o he numbe o people who ollow all he p esc ip ion ajec o ies
going om one e ical ba o ano he (le o igh ). ATC: Ana omical The apeu ic and Chemical
Classi ica ion. RR: ela i e isk. F om manusc ip 5.
in e p e he esul s wi h cau ion. Fo example aldos e one an agonis s a e o en conside ed an
"add-on" d ug in ea men o hype ension[133]. Also, any disease p og ession ha could ha e
been he a ionale o changing ea men egimens we e no included in he analysis. We a gue
ha , gi en u he model e inemen , his concep o p esc ip ion ajec o ies can assis clinical de-
cision making by iden i ica ion o people whe e i unlikely ha hey will espond su icien ly o
i s -line he apy. The ull-leng h manusc ip is p esen ed in chap e 11.
5.6 In e ne applica ions display pe spec i es wi h da a-d i en e-
sea ch
Mos exis ing so wa e de eloped o s udy heal hca e da a do ei he equi e use s o upload hei
own da a o limi use s o que y wi hin a speci ic domain[134], [135]. I is pa amoun o da a-
d i en esea ch ha he esul s o he analyses a e being e alua ed con inuously as hei use ulness
o en depend on p ope examples [63]. The esul s desc ibed in sec ions abo e a e p esen ed wi h
e e ence o ela i ely ew examples ha assis in in e p e a ion o he model lea ing ou , po en ially,
43
CHAPTER 5. RESULTS ON DIAGNOSTIC CODES
ele an in o ma ion. To his end, we ha e de eloped The Danish Disease T ajec o y B owse (DTB)
which is an example o a so wa e ool whe e use s can decide which example o analyze.
The s udy p esen ed in manusc ip 6 in oduces DTB which is a web applica ion ha makes he
disease ajec o y p og am a ailable o anybody, including people wi hou coding expe ience, gi en
in e ne access, and a he same ime es ablishing a pla o m o model c i ique and sugges ions. The
b owse was de eloped as a ool ha enables use s o cha ac e ize a pheno ype o in e es using
disease ajec o ies. A se o unc ionali ies we e added o he b owse , including sea ch il e s, ha
o example enable use s o sea ch o disease ajec o ies compu ed om ei he males o emales.
As he b owse only con ains summa y s a is ics i does no p o ide any access o pe sonal da a.
Figu e 19: Sc eensho om he DTB home page. The home page has h ee cen al elemen s: a
sea ch ba in he middle, an uppe ho izon ibbon and a e ical side panel o he igh . In he
sea ch ba use s can ype in an ICD-10 code o in e es , e.g. I25. Use s can decide i he esul s o
he que y should be ep esen ed as a disease ajec o y ne wo k o linea disease ajec o ies. The
uppe ho izon al ba has a a ie y o di e en bu ons ha can be used o na iga e he in e ac i e
phase ha will eplace he a ea wi h he sea ch ba , when a que y has been placed. The e ical side
panel has he di e en sea ch il e s and op ions ega ding layou and downloads. DTB: Danish
Disease T ajec o y B owse F om manusc ip 6.
DTB is made a ailable a h p://d b.cp .ku.dk/ and le s use s cha ac e ize a pheno ype by
p o iding access o nea ly 25 yea s o da a om mo e han se en million people. Use s can que y
44
CHAPTER 5. RESULTS ON DIAGNOSTIC CODES
DTB by speci ying a case popula ion in he o m o le el 3ICD-10 codes, e.g. I25 o ch onic IHD
(Figu e 19). The main esul o he s udy was ha he esul s o he disease ajec o y p og am
became publicly a ailable wi h an in e ac i e in e ace. Po en ially, his may lead o u u e de el-
opmen s o he p og am ha a e based on inpu om a b oad ange o esea ch communi ies. DTB
has a use - iendly in e ace ha was de eloped o suppo a g aphical ep esen a ion o empo al
associa ions be ween diagnoses in NPR da a. This is becoming mo e impo an han e e as la ge
heal hca e da a a e becoming inc easingly di icul o manage and analyse because da a olumes
and a ie y a e g owing as e han e e . Sea ch esul s a e a ailable o download as g aphical
ep esen a ions o da a ables wi h summa y s a is ics. Mo eo e , a ange o sea ch il e s can by
applied o cus omize he que y. Fil e s include numbe o diagnoses pe ajec o y, ange o num-
be o pa ien s ha should ollow a ajec o y o i o be included in he sea ch, ange o RRs and
sex (males only, emales only o bo h). The s udy exempli ies ha he digi al de elopmen and ech-
nologies p o ide new oppo uni ies o sha ing esea ch esul s. This will ob iously also make hem
mo e p one o ex e nal c i ique. We a gue ha in he long un, bo h esea che s, clinicians, and
pa ien s will bene i om his le el o anspa ency which hand-on usage and que ies acili a es. To
his end, DTB and o he simila p ojec s a e only he beginning. The b owse unc ionali ies and
in e p e a ion o he disease ajec o ies a e desc ibed in g ea e de ail in he ull-leng h manusc ip ,
which is p esen ed in chap e 12.
45
REFERENCES ON DIAGNOSTIC CODES
[23] K. Nicholson, T. T. Mako ski, L. E. G i i h, P. Raina, S. S anges, and M. an den Akke ,
“Mul imo bidi y and como bidi y e isi ed: Re ining he concep s o in e na ional heal h
esea ch,”Jou nal o clinical epidemiology, ol.105, pp. 142–146,2019.doi:10.1016/j.jclinepi.
2018.09.008.
[24] S. P. Bell and A. A. Sa a , “Epidemiology o mul imo bidi y in olde adul s wi h ca dio as-
cula disease,” Clinics in ge ia ic medicine, ol. 32, no. 2, p. 215, 2016. doi:10.1016/j.jacc.
2018.03.022.
[25] A. R. Feins ein, “The p e- he apeu ic classi ica ion o co-mo bidi y in ch onic disease,” Jou -
nal o ch onic diseases, ol. 23, no. 7, pp. 455–468, 1970. doi:10.1016/0021-9681(70)90054-8.
[26] V. De G oo ,H. Becke man,G.J.Lankho s , andL.M.Bou e , “How omeasu e como bidi y:
A c i ical e iew o a ailable me hods,” Jou nal o clinical epidemiology, ol. 56, no. 3, pp. 221–
229, 2003. doi:10.1016/s0895-4356(02)00585-1.
[27] M. E. Cha lson, P. Pompei, K. L. Ales, and C. R. MacKenzie, “A new me hod o classi y-
ing p ognos ic como bidi y in longi udinal s udies: De elopmen and alida ion,” Jou nal o
ch onic diseases, ol. 40, no. 5, pp. 373–383, 1987. doi:10.1016/0021-9681(87)90171-8.
[28] F. Zhang, A. Bha adwaj, M. O. Mohamed, J. Enso , G. Pea , and M. A. Mamas, “Impac
o cha lson co-mo bidi y index sco e on managemen and ou comes a e acu e co ona y
synd ome,” The Ame ican Jou nal o Ca diology, ol. 130, pp. 15–23, 2020. doi:10.1016/j.
amjca d.2020.06.022.
[29] R. Gup a and D. A. Wood, “P ima y p e en ion o ischaemic hea disease: Popula ions,
indi iduals, and heal h p o essionals,” The Lance , ol. 394, no. 10199, pp. 685–696, 2019.
doi:10.1016/S0140-6736(19)31893-8.
[30] X. Rossello, S. J. Pocock, and D. G. Julian, “Long- e m use o ca dio ascula d ugs: Chal-
lenges o esea ch and o pa ien ca e,” Jou nal o he Ame ican College o Ca diology, ol. 66,
no. 11, pp. 1273–1285, 2015. doi:10.1016/j.jacc.2015.07.018.
[31] M. W. Rich, D. A. Chyun, A. H. Skolnick, K. P. Alexande , D. E. Fo man, D. W. Ki zman, M. S.
Mau e , J. B. McClu ken, B. M. Resnick, W. K. Shen, e al., “Knowledge gaps in ca dio ascula
ca e o he olde adul popula ion: A scien i ic s a emen om he ame ican hea associa-
ion, ame ican college o ca diology, and ame ican ge ia ics socie y,” Ci cula ion, ol. 133,
no. 21, pp. 2103–2122, 2016. doi:10.1161/CIR.0000000000000380.
[32] S. M. G undy, “D ug he apy o he me abolic synd ome: Minimizing he eme ging c isis in
polypha macy,” Na u e e iews D ug disco e y, ol. 5, no. 4, pp. 295–309, 2006. doi:10.1038/
n d2005.
[33] D. C. Angus, B. M. Alexande , S. Be y, M. Bux on, R. Lewis, M. Paoloni, S. A. Webb, S.
A nold, A. Ba ke , D. A. Be y, e al., “Adap i e pla o m ials: De ini ion, design, conduc
and epo ing conside a ions.,” Na u e Re iews D ug Disco e y, ol. 18, no. 10, pp. 797–808,
2019. doi:10.1038/s41573-019-0034-3.
[34] T. Boe ma, J. Ha ison, R. Jakob, C. Ma he s, A. Schmide , and S. Webe , “Re ising he icd:
Explaining he who app oach,” The Lance , ol. 388, no. 10059, pp. 2476–2477, 2016. doi:
10.1016/S0140-6736(16)31851-7.
52

REFERENCES ON DIAGNOSTIC CODES
[35] T. Siggaa d, R. Reguan , I. F. Jø gensen, A. D. Haue, M. Lademann, A. Aguayo-O ozco, J. X.
Hjal elin, A. B. Jensen, K. Banasik, and S. B unak, “Disease ajec o y b owse o explo -
ing empo al, popula ion-wide disease p og ession pa e ns in 7.2 million danish pa ien s,”
Na u e communica ions, ol. 11, no. 1, pp. 1–10, 2020. doi:10.1038/s41467-020-18682-4.
[36] M. Schmid , S. A. J. Schmid , J. L. Sandegaa d, V. Eh ens ein, L. Pede sen, and H. T. Sø ensen,
“The danish na ional pa ien egis y: A e iew o con en , da a quali y, and esea ch po en-
ial,” Clinical epidemiology, ol. 7, p. 449, 2015. doi:10.2147/CLEP.S91125.
[37] W.H.O ganiza ione al.,“Whoin e na ional wo kingg oup o d ug s a is icsme hodology,
who collabo a ing cen e o d ug s a is ics me hodology, who collabo a ing cen e o d ug
u iliza ion esea ch and clinical pha macological se ices: In oduc ion o d ug u iliza ion
esea ch,” In oduc ion o d ug u iliza ion esea ch. Gene a: Wo ld Heal h O ganiza ion, 2003.
[38] Ana omical he apeu ic chemical (a c) classi ica ion,h ps://www.who.in / ools/a c-ddd-
oolki /a c-classi ica ion, (Accessed on 05/10/2021).
[39] U. M. Pe e sen, R. Dybkæ , and H. Olesen, “P ope ies and uni s in he clinical labo a o y
sciences. pa xxiii. he npu e minology, p inciples, and implemen a ion: A use ’s guide
(iupac echnical epo ),” Pu e and Applied Chemis y, ol. 84, no. 1, pp. 137–165, 2011. doi:
10.1515/cclm.2011.735.
[40] N. NOMESCO, Nomesco classi ica ion o su gical p ocedu es (ncsp), e sion 1.14,h p://no den.
di a-po al.o g/smash/ge /di a2:968721/FULLTEXT01.pd , (Accessed on 05/17/2021),
2011.
[41] K. Adelbo g, J. Sundbøll, T. Munch, T. F øsle , H. T. Sø ensen, H. E. Bø ke , and M. Schmid ,
“Posi i e p edic i e alue o ca diac examina ion, p ocedu e and su ge y codes in he dan-
ish na ional pa ien egis y: A popula ion-based alida ion s udy,” BMJ open, ol. 6, no. 12,
2016. doi:10.1136/bmjopen-2016-012817.
[42] J. C. Hong and A. J. Bu e, “Assessing Clinical Ou comes in a Da a-Rich Wo ld—A Real-
i y Check on Real-Wo ld Da a,” en, JAMA Ne wo k Open, ol. 4, no. 7, e2117826, Jul. 2021,
issn: 2574-3805. doi:10.1001/jamane wo kopen.2021.17826. [Online]. A ailable: h ps:
/ / jamane wo k . com / jou nals / jamane wo kopen / ulla icle / 2782343 ( isi ed on
07/30/2021).
[43] E. S. Be ne , G. D. Webs e , A. A. Shuge man, J. R. Jackson, J. Algina, A. L. Bake , E. V.
Ball, C. G. Cobbs, V. W. Dennis, E. P. F enkel, e al., “Pe o mance o ou compu e -based
diagnos ic sys ems,” New England Jou nal o Medicine, ol. 330, no. 25, pp. 1792–1796, 1994.
doi:10.1056/NEJM199406233302506.
[44] K.-H. Yu, A. L. Beam, and I. S. Kohane, “A i icial in elligence in heal hca e,” Na u e biomed-
ical enginee ing, ol. 2, no. 10, pp. 719–731, 2018. doi:10.1038/s41551-018-0305-z.
[45] P. R. Njøls ad, O. A. And eassen, S. B unak, A. D. Bø glum, J. Dillne , T. Esko, P. W. F anks,
N. F eime , L. G oop, H. Heime , e al., “Roadmap o a p ecision-medicine ini ia i e in he
no dic egion,” Na u e gene ics, ol. 51, no. 6, pp. 924–930, 2019. doi:10.1038/s41588-019-
0391-1.
[46] S. W. Tu, O. Boden eide , C. Çelik, C. G. Chu e, S. Hea d, R. Jakob, G. Jiang, S. Kim, E.
Mille , M. M. Musen, e al., “A con en model o he icd-11 e ision,” Technical Repo BMIR,
pp. 2010–1405, 2010.
53
REFERENCES ON DIAGNOSTIC CODES
[47] R. H. de Jong, “Compu e c u ches,” JAMA, ol. 240, no. 11, pp. 1176–1176, 1978. doi:10.
1001/jama.240.11.1176b.
[48] M. A. Haendel, C. G. Chu e, and P. N. Robinson, “Classi ica ion, on ology, and p ecision
medicine,” New England Jou nal o Medicine, ol. 379, no. 15, pp. 1452–1462, 2018. doi:10.
1056/NEJM a1615014.
[49] S. Köhle , N. A. Vasile sky, M. Engels ad, E. Fos e , J. McMu y, S. Aymé, G. Baynam, S. M.
Bello, C. F. Boe koel, K. M. Boyco , e al., “The human pheno ype on ology in 2017,” Nucleic
acids esea ch, ol. 45, no. D1, pp. D865–D876, 2017. doi:10.1093/na /gkw1039.
[50] M. Schmid , S. A. J. Schmid , K. Adelbo g, J. Sundbøll, K. Laugesen, V. Eh ens ein, and H. T.
Sø ensen, “The danish heal h ca e sys em and epidemiological esea ch: F om heal h ca e
con ac s o da abase eco ds,” Clinical epidemiology, ol. 11, p. 563, 2019. doi:10.2147/CLEP.
S179083.
[51] M. Schmid , L. Pede sen, and H. T. Sø ensen, “The danish ci il egis a ion sys em as a ool
in epidemiology,” Eu opean jou nal o epidemiology, ol. 29, no. 8, pp. 541–549, 2014. doi:10.
1007/s10654-014-9930-3.
[52] Pa ien egis e ing - Fællesindhold - Sundhedsda as y elsen,h ps://sundhedsda as y elsen.
dk/da/ amme -og- e ningslinje /om-pa ien egis e ing/pa ien egis e ing-
eallesindhod, (Accessed on 31/05/2021), Ap . 2021.
[53] E. Lynge, J. L. Sandegaa d, and M. Rebolj, “The danish na ional pa ien egis e ,” Scandina-
ian jou nal o public heal h, ol.39,no. 7_suppl, pp.30–33,2011.doi:10.1177/1403494811401482.
[54] K. Helweg-La sen, “The danish egis e o causes o dea h,” Scandina ian jou nal o public
heal h, ol. 39, no. 7_suppl, pp. 26–29, 2011. doi:10.1177/1403494811399958.
[55] H. Wallach Kildemoes, H. To Sø ensen, and J. Hallas, “The danish na ional p esc ip ion
egis y,” Scandina ian jou nal o public heal h, ol. 39, no. 7_suppl, pp. 38–41, 2011. doi:10.
1177/1403494810394717.
[56] C. M. Delude, “Deep pheno yping: The de ails o disease,” Na u e, ol. 527, no. 7576, S14–
S15, 2015. doi:10.1038/527S14a.
[57] C. Özcan, K. Juel, J. F. Lassen, L. M. on Kappelgaa d, P. E. Mo ensen, and G. Gislason, “The
danish hea egis y,” Clinical epidemiology, ol. 8, p. 503, 2016. doi:10.2147/CLEP.S99475.
[58] P. B. Jensen, L. J. Jensen, and S. B unak, “Mining elec onic heal h eco ds: Towa ds be e
esea ch applica ions and clinical ca e,” Na u e Re iews Gene ics, ol. 13, no. 6, pp. 395–405,
2012. doi:10.1038/n g3208.
[59] A. F. G ann, R. E ichsen, A. G. Nielsen, T. F øsle , and R. W. Thomsen, “Exis ing da a sou ces
o clinicalepidemiology:Theclinicallabo a o yin o ma ionsys em(labka) esea chda abase
a aa hus uni e si y, denma k,” Clinical epidemiology, ol. 3, p. 133, 2011. doi:10.2147/CLEP.
S17901.
[60] E. Sø ensen, L. Ch is iansen, B. Wilkowski, M. H. La sen, K. S. Bu gdo , L. W. Thø ne ,
J. Nissen, O. B. Pede sen, K. Banasik, S. B unak, e al., “Da a esou ce p o ile: The copen-
hagen hospi al biobank (chb),” In e na ional Jou nal o Epidemiology, 2020. doi:10 . 1093 /
ije/dyaa157.
54
REFERENCES ON DIAGNOSTIC CODES
[61] Decode gene ics | a global leade in human gene ics,h ps://www.decode.com/, (Accessed on
08/15/2021).
[62] K. Poppe , The logic o scien i ic disco e y. Rou ledge, 2005.
[63] S. Ca ol and D. Goods ein, “De ining he scien i ic me hod,” Na Me hods, ol. 6, p. 237,
2009. doi:10.1038/nme h0409-237.
[64] D. A. G imes and K. F. Schulz, “An o e iew o clinical esea ch: The lay o he land,” The
lance , ol. 359, no. 9300, pp. 57–61, 2002. doi:10.1016/S0140-6736(02)07283-5.
[65] A. C. Fana o , R. M. Cali , R. A. Ha ing on, C. B. G ange , J. J. McMu ay, M. R. Pa el,
D. L. Bha , S. Windecke , A. F. He nandez, C. M. Gibson, e al., “Randomized ials e sus
commonsense and clinicalobse a ion:Jacc e iew opico he week,”Jou nal o he Ame ican
College o Ca diology, ol. 76, no. 5, pp. 580–589, 2020. doi:10.1016/j.jacc.2020.05.069.
[66] R. E zioni, S a is ics o Heal h Da a Science: An O ganic App oach. Sp inge Na u e, 2020.
[67] K. J. Ro hman, S. G eenland, and T. L. Lash, Mode n epidemiology. Lippinco Williams &
Wilkins, 2008.
[68] S.Z.Abilds om,S.Rasmussen, and M. Madsen,“Changesinhospi aliza ion a eandmo a-
li ya e acu emyoca dialin a c ionindenma k a e diagnos icc i e iaandme hods changed,”
Eu opean hea jou nal, ol. 26, no. 10, pp. 990–995, 2005. doi:10.1093/eu hea j/ehi039.
[69] L. Øs e gaa d, J. H. Bu , K. K agholm, M. Schou, M. Phelps, R. Sø ensen, M. Lambe s, G.
Gislason, C. To p-Pede sen, L. Købe , e al., “Incidence o acu e co ona y synd ome du -
ing na ional lock-down: Insigh s om na ionwide da a du ing he co ona i us disease 2019
(co id-19) pandemic,” Ame ican hea jou nal, ol. 232, pp. 146–153, 2020. doi:10.1016/j.
ahj.2020.11.004.
[70] P. N. Lau sen, L. Holm ang, J. Lønbo g, L. Købe , D. E. Hø s en, S. Helq is , P. Clemmensen,
H. Kelbæk, E. Jø gensen, J. F. Lassen, e al., “Compa ison be ween pa ien s included in an-
domizedcon olled ials o ischemichea disease and eal-wo ldda a.ana ionwides udy,”
Ame ican hea jou nal, ol. 204, pp. 128–138, 2018. doi:10.1016/j.ahj.2018.05.018.
[71] A. B. Jensen, P. L. Moseley, T. I. Op ea, S. G. Ellesøe, R. E iksson, H. Schmock, P. B. Jensen,
L. J. Jensen, and S. B unak, “Tempo al disease ajec o ies condensed om popula ion-wide
egis y da a co e ing 6.2 million pa ien s,” Na u e communica ions, ol. 5, no. 1, pp. 1–10,
2014. doi:10.1038/s ep36624.
[72] M. K. Beck, A. B. Jensen, A. B. Nielsen, A. Pe ne , P. L. Moseley, and S. B unak, “Diagnosis
ajec o ies o p io mul i-mo bidi y p edic sepsis mo ali y,” Scien i ic epo s, ol. 6, no. 1,
pp. 1–9, 2016. doi:10.1038/s ep36624.
[73] J. X. Hu, M. Hellebe g, A. B. Jensen, S. B unak, and J. Lundg en, “A la ge-coho , longi udi-
nal s udy de e mines p ecance disease ou es ac oss di e en cance ypes,” Cance esea ch,
ol. 79, no. 4, pp. 864–872, 2019. doi:10.1158/0008-5472.CAN-18-1677.
[74] I. F. Jø gensen, A. Aguayo-O ozco, M. Lademann, and S. B unak, “Age-s a i ied longi udi-
nal s udy o alzheime ’s and ascula demen ia pa ien s,” Alzheime ’s & Demen ia, ol. 16,
no. 6, pp. 908–917, 2020. doi:10.1002/alz.12091.
[75] C. López-O ín, M. A. Blasco, L. Pa idge, M. Se ano, and G. K oeme , “The hallma ks o
aging,” Cell, ol. 153, no. 6, pp. 1194–1217, 2013. doi:10.1016/j.cell.2013.05.039.
55
REFERENCES ON DIAGNOSTIC CODES
[76] E. L. F ome and H. Checkoway, “Use o poisson eg ession models in es ima ing incidence
a es and a ios,” Ame ican jou nal o epidemiology, ol. 121, no. 2, pp. 309–323, 1985. doi:
10.1093/ox o djou nals.aje.a114001.
[77] I. K. Ki k, C. Simon, K. Banasik, P. C. Holm, A. D. Haue, P. B. Jensen, L. J. Jensen, C. L.
Rod íguez, M. K. Pede sen, R. E iksson, e al., “Linking glycemic dys egula ion in diabe es
o symp oms, como bidi ies, and gene ics h ough eh da a mining,” Eli e, ol. 8, e44941,
2019. doi:10.7554/eLi e.44941.
[78] H. Schü ze, C. D. Manning, and P. Ragha an, In oduc ion o in o ma ion e ie al. Camb idge
Uni e si y P ess Camb idge, 2008, ol. 39.
[79] D. B. Wes e al.,In oduc ion o g aph heo y. P en ice hall Uppe Saddle Ri e , 2001, ol. 2.
[80] D.Xuand Y. Tian,“A comp ehensi esu eyo clus e ingalgo i hms,” Annals o Da a Science,
ol. 2, no. 2, pp. 165–193, 2015. doi:10.1007/s40745-015-0040-1.
[81] P. Spec o . (2011). “S a 133 class no es,” [Online]. A ailable: h ps://www.s a .be keley.
edu/~s133/all2011.pd ( isi ed on 03/18/2021).
[82] L. Kau man and P. J. Rousseeuw, Finding g oups in da a: an in oduc ion o clus e analysis. John
Wiley & Sons, 2009, ol. 344.
[83] K. W. Johnson, J. To es So o, B. S. Glicksbe g, K. Shamee , R. Mio o, M. Ali, E. Ashley, and
J. T. Dudley, “A i icial in elligence in ca diology,” Jou nal o he Ame ican College o Ca diology,
ol. 71, no. 23, pp. 2668–2679, 2018. doi:10.1016/j.jacc.2018.03.521.
[84] I. S. Kohane, B. J. A onow, P. A illach, B. K. Beaulieu-Jones, R. Bellazzi, R. L. B ad o d, G. A.
B a , M. Canna a o, J. J. Cimino, N. Ga cía-Ba io, e al., “Wha e e y eade should know
abou s udies using elec onic heal h eco d da a bu may be a aid o ask,” Jou nal o medical
In e ne esea ch, ol. 23, no. 3, e22219, 2021. doi:10.2196/22219.
[85] S. M. Van Dongen, “G aph clus e ing by low simula ion,” Ph.D. disse a ion, 2000.
[86] L. L. Gao, J. Bien, and D. Wi en, Selec i e in e ence o hie a chical clus e ing, 2020. a Xi : 2012.
02936 [s a .ME].
[87] D. G. Kleinbaum and M. Klein, Su i al analysis. Sp inge , 2012.
[88] D. R. Cox, “Reg ession models and li e- ables,” Jou nal o he Royal S a is ical Socie y: Se ies B
(Me hodological), ol. 34, no. 2, pp. 187–202, 1972.
[89] P. Roys on and M. K. Pa ma , “Res ic ed mean su i al ime: An al e na i e o he haza d
a io o he design and analysis o andomized ials wi h a ime- o-e en ou come,” BMC
medical esea ch me hodology, ol. 13, no. 1, pp. 1–15, 2013. doi:10.1186/1471-2288-13-152.
[90] W. S. McCulloch and W. Pi s, “A logical calculus o he ideas immanen in ne ous ac i i y,”
The bulle in o ma hema ical biophysics, ol. 5, no. 4, pp. 115–133, 1943.
[91] M. A. Nielsen, Neu al ne wo ks and deep lea ning. De e mina ion p ess San F ancisco, CA,
2015, ol. 25.
[92] P. Baldi, S. B unak, and F. Bach, Bioin o ma ics: he machine lea ning app oach. MIT p ess, 2001.
[93] P. Baldi, “G adien descen lea ning algo i hm o e iew: A gene al dynamical sys ems pe -
spec i e,” IEEE T ansac ions on Neu al Ne wo ks, ol. 6, no. 1, pp. 182–195, 1995. doi:10.1109/
72.363438.
56
REFERENCES ON DIAGNOSTIC CODES
[94] C. M. Bishop e al.,Neu al ne wo ks o pa e n ecogni ion. Ox o d uni e si y p ess, 1995.
[95] E. J. Topol, “High-pe o mance medicine: The con e gence o human and a i icial in elli-
gence,” Na u e medicine, ol. 25, no. 1, pp. 44–56, 2019. doi:10.1038/s41591-018-0300-7.
[96] G. Tseng, In e p e ing complex models wi h shap alues,h ps://medium.com/@gab iel seng/
in e p e ing-complex-models-wi h-shap- alues-1c187db6ec83,
(Accessed on 07/16/2021).
[97] L. S. Shapley, H. Kuhn, and A. Tucke , “Con ibu ions o he heo y o games,” Annals o
Ma hema ics s udies, ol. 28, no. 2, pp. 307–317, 1953.
[98] L. S. Shapley, “S ochas ic games,” P oceedings o he na ional academy o sciences, ol. 39, no. 10,
pp. 1095–1100, 1953.
[99] S. Lundbe g and S.-I. Lee, “A uni ied app oach o in e p e ing model p edic ions,” a Xi
p ep in a Xi :1705.07874, 2017.
[100] C. Molna , In e p e able machine lea ning. Lulu. com, 2020.
[101] N. R. Cook, “Useandmisuseo he ecei e ope a ingcha ac e is iccu e in iskp edic ion,”
Ci cula ion, ol. 115, no. 7, pp. 928–935, 2007. doi:10.1161/CIRCULATIONAHA.106.672402.
[102] G. S. Collins, J. B. Rei sma, D. G. Al man, and K. G. Moons, “T anspa en epo ing o a
mul i a iable p edic ion model o indi idual p ognosis o diagnosis ( ipod) he ipod
s a emen ,” Ci cula ion, ol. 131, no. 2, pp. 211–219, 2015. doi:10.1161/CIRCULATIONAHA.
114.014508.
[103] F. E. Ha ell J , Reg ession modeling s a egies: wi h applica ions o linea models, logis ic and o -
dinal eg ession, and su i al analysis. Sp inge , 2015.
[104] T. Fawce , “An in oduc ion o oc analysis,” Pa e n ecogni ion le e s, ol. 27, no. 8, pp. 861–
874, 2006. doi:10.1016/j.pa ec.2005.10.010.
[105] R. an de Schoo , S. Depaoli, R. King, B. K ame , K. Mä ens, M. G. Tadesse, M. Vannucci,
A. Gelman, D. Veen, J. Willemsen, e al., “Bayesian s a is ics and modelling,” Na u e Re iews
Me hods P ime s, ol. 1, no. 1, pp. 1–26, 2021. doi:10.1038/s43586-020-00001-2.
[106] P.-C. Bü kne , “B ms: An package o bayesian mul ile el models using s an,” Jou nal o
s a is ical so wa e, ol. 80, no. 1, pp. 1–28, 2017.
[107] T. M. Dono an and R. M. Mickey, Bayesian s a is ics o beginne s: A s ep-by-s ep app oach. Ox-
o d Uni e si y P ess, USA, 2019.
[108] A. Veh a i, A. Gelman, and J. Gab y, “P ac ical bayesian model e alua ion using lea e-one-
ou c oss- alida ion and waic,” S a is ics and compu ing, ol. 27, no. 5, pp. 1413–1432, 2017.
doi:10.1007/s11222-016-9696-4.
[109] M. Be ancou , “A concep ual in oduc ion o hamil onian mon e ca lo,” a Xi p ep in
a Xi :1701.02434, 2017.
[110] J. K uschke, “Doing bayesian da a analysis: A u o ial wi h , jags, and s an,” 2014.
[111] J. K. K uschke and T. M. Liddell, “Bayesian da a analysis o newcome s,” Psychonomic bul-
le in & e iew, ol. 25, no. 1, pp. 155–177, 2018.
57

REFERENCES ON DIAGNOSTIC CODES
[112] H. B. Ben zen and N. Høs mælingen, “Balancing p o ec ion and ee mo emen o pe sonal
da a:The neweu opeanunion gene alda a p o ec ion egula ion,”Annals o in e nal medicine,
ol. 170, no. 5, pp. 335–337, 2019. doi:10.7326/M18-2782.
[113] H. K. Beeche , “E hics and clinical esea ch,” Bulle in o he Wo ld Heal h O ganiza ion, ol. 79,
pp. 367–372, 2001.
[114] Wha o no i y? h ps://en.n k.dk/how- o- no i y/wha - o- no i y, (Accessed on
18/05/2021), May 2021.
[115] C. J. Haug, “Tu ning he ables— he new eu opean gene al da a p o ec ion egula ion,” New
England Jou nal o Medicine, ol. 379, no. 3, pp. 207–209, 2018. doi:10.1056/NEJMp1806637.
[116] M.D. Wilkinson,M.Dumon ie ,I. J.Aalbe sbe g,G.Apple on,M. Ax on,A.Baak,N. Blombe g,
J.-W. Boi en, L. B. da Sil a San os, P. E. Bou ne, e al., “The ai guiding p inciples o sci-
en i ic da a managemen and s ewa dship,” Scien i ic da a, ol. 3, no. 1, pp. 1–9, 2016. doi:
10.1038/sda a.2016.18.
[117] K. K. Jensen, L. Whi eley, and P. Sandøe, RCR-A Danish ex book o cou ses in Responsible Con-
duc o Resea ch. Uni e si y o Copenhagen, Depa men o Food and Resou ce Economics,
2020.
[118] F. Mölde , K. P. Jablonski, B. Le che , M. B. Hall, C. H. Tomkins-Tinch, V. Socha , J. Fo s e ,
S. Lee, S. O. Twa dziok, A. Kani z, e al., “Sus ainable da a analysis wi h snakemake,”
F1000Resea ch, ol. 10, p. 33, 2021. doi:10.12688/ 1000 esea ch.29032.2.
[119] Mee he gi lab eam | gi lab,h ps://abou .gi lab.com/company/ eam/, (Accessed on
05/17/2021).
[120] M.Claussni ze ,J.H.Cho,R.Collins,N.J.Cox, E. T. De mi zakis, M. E. Hu les,S.Ka hi esan,
E. E. Kenny, C. M. Lindg en, D. G. MacA hu , e al., “A b ie his o y o human disease
gene ics,” Na u e, ol. 577, no. 7789, pp. 179–189, 2020. doi:10.1038/s41586-019-1879-7.
[121] M.Gheo ghiade,G.Sopko, L.DeLuca, E. J.Velazquez,J.D.Pa ke ,P. F. Binkley,Z. Sadowski,
K. S. Golba, D. L. P io , J. L. Rouleau, e al., “Na iga ing he c oss oads o co ona y a e y
disease and hea ailu e,” Ci cula ion, ol. 114, no. 11, pp. 1202–1213, 2006. doi:10.1161/
CIRCULATIONAHA.106.623199.
[122] R. R. Kalyani, “Glucose-lowe ing d ugs o educe ca dio ascula isk in ype 2 diabe es,”
New England Jou nal o Medicine, ol. 384, no.13,pp. 1248–1260, 2021.doi:10.1056/NEJMcp2000280.
[123] L. Busija, K. Lim, C. Szoeke, K. M. Sande s, and M. P. McCabe, “Do eplicable p o iles o mul-
imo bidi y exis ? sys ema ic e iew and syn hesis,” Eu opean jou nal o epidemiology, ol. 34,
no. 11, pp. 1025–1053, 2019. doi:doi.o g/10.1007/s10654-019-00568-5.
[124] T. Nagamine, B. Gille e, A. Pakhomo , J. Kahoun, H. Maye , R. Bu ghaus, J. Lippe , and M.
Saxena, “Mul iscale classi ica ion o hea ailu e pheno ypes by unsupe ised clus e ing o
uns uc u ed elec onic medical eco d da a,” Scien i ic epo s, ol. 10, no. 1, pp. 1–13, 2020.
doi:10.1038/s41598-020-77286-6.
[125] F. C owe, D. T. Zemedikun, K. Oko h, N. J. Adde ley, G. Rudge, M. Sheldon, K. Ni an-
ha akuma , and T. Ma shall, “Como bidi y pheno ypes and isk o mo ali y in pa ien s
wi h ischaemic hea disease in he uk,” Hea , ol. 106, no. 11, pp. 810–816, 2020. doi:10.
1136/hea jnl-2019-316091.
58
REFERENCES ON DIAGNOSTIC CODES
[126] F. D’Ascenzo, O. De Filippo, G. Gallone, G. Mi one, M. A. De iu, M. Iannaccone, A. A iza-
Solé, C. Liebe au, S. Manzano-Fe nández, G. Quad i, e al., “Machine lea ning-based p e-
dic iono ad e se e en s ollowing an acu e co ona y synd ome (p aise):Amodellings udy
o pooled da ase s,” The Lance , ol. 397, no. 10270, pp. 199–207, 2021. doi:10.1016/S0140-
6736(20)32519-8.
[127] M. Mo wani, D. Dey, D. S. Be man, G. Ge mano, S. Achenbach, M. H. Al-Mallah, D. An-
d eini, M. J. Budo , F. Cadema i i, T. Q. Callis e , e al., “Machine lea ning o p edic ion o
all-cause mo ali y in pa ien s wi h suspec ed co ona y a e y disease: A 5-yea mul icen e
p ospec i e egis y analysis,” Eu opean hea jou nal, ol. 38, no. 7, pp. 500–507, 2017. doi:
10.1093/eu hea j/ehw188.
[128] M. F. Gensheime and B. Na asimhan, “A scalable disc e e- ime su i al model o neu al
ne wo ks,” Pee J, ol. 7, e6257, Jan. 2019, issn: 2167-8359. doi:10.7717/pee j.6257.
[129] J.L.Mega,N. O. S i ziel,J. G.Smi h,D.I. Chasman, M. J. Caul ield, J.J.De lin,F.No dio,C. L.
Hyde, C. P. Cannon, F. M. Sacks, e al., “Gene ic isk, co ona y hea disease e en s, and he
clinical bene i o s a in he apy: An analysis o p ima y and seconda y p e en ion ials,”
The Lance , ol. 385, no. 9984, pp. 2264–2271, 2015. doi:10.1016/S0140-6736(14)61730-X.
[130] D. S. Wisha , C. Knox, A. C. Guo, S. Sh i as a a, M. Hassanali, P. S o ha d, Z. Chang, and
J. Woolsey, “D ugbank: A comp ehensi e esou ce o in silico d ug disco e y and explo-
a ion,” Nucleic acids esea ch, ol. 34, no. suppl_1, pp. D668–D672, 2006. doi:10.1093/na /
gkj067.
[131] A. Laupacis, D. L. Sacke , and R. S. Robe s, “An assessmen o clinically use ul measu es
o he consequences o ea men ,” New England jou nal o medicine, ol. 318, no. 26, pp. 1728–
1733, 1988. doi:10.1056/NEJM198806303182605.
[132] I. A. Sco , S. N. Hilme , E. Ree e, K. Po e , D. Le Cou eu , D. Rigby, D. Gnjidic, C. B. Del
Ma , E. E. Roughead, A. Page, e al., “Reducing inapp op ia e polypha macy: The p ocess
o dep esc ibing,” JAMA in e nal medicine, ol. 175, no. 5, pp. 827–834, 2015. doi:10.1001/
jamain e nmed.2015.0324.
[133] C. Rosendo , D. T. Lackland, M. Allison, W. S. A onow, H. R. Black, R. S. Blumen hal, C. P.
Cannon,J.A.DeLemos, W. J. Ellio ,L. Findeiss,e al., “T ea men o hype ensioninpa ien s
wi h co ona y a e y disease: A scien i ic s a emen om he ame ican hea associa ion,
ame ican college o ca diology, and ame ican socie y o hype ension,” Ci cula ion, ol. 131,
no. 19, e435–e470, 2015. doi:10.1016/j.jacc.2015.02.038.
[134] M. Schmid , L. V. Ande sen, S. F iis, K. Juel, and G. Gislason, “Da a esou ce p o ile: Danish
hea s a is ics,” In e na ional jou nal o epidemiology, ol. 46, no. 5, 1368–1369g, 2017. doi:10.
1093/ije/dyx108.
[135] A. Aguado, F. Mo a alla-Na a o, F. López-Sima o, and V. Mo eno, “Mo bine : Mul imo -
bidi y ne wo ks in adul gene al popula ion. analysis o ype 2 diabe es melli us como bid-
i y,” Scien i ic epo s, ol. 10, no. 1, pp. 1–12, 2020. doi:10.1038/s41598-020-59336-1.
[136] J. Sundbøll, K. Adelbo g, T. Munch, T. F øsle , H. T. Sø ensen, H. E. Bø ke , and M. Schmid ,
“Posi i e p edic i e alueo ca dio ascula diagnosesin hedanishna ionalpa ien egis y:
A alida ion s udy,” BMJ open, ol. 6, no. 11, 2016.
59
REFERENCES ON DIAGNOSTIC CODES
[137] D. S. Cha , N. H. Shah, and D. Magnus, “Implemen ing machine lea ning in heal h ca e—
add essing e hical challenges,” The New England jou nal o medicine, ol. 378, no. 11, p. 981,
2018. doi:10.1056/NEJMp1714229.
[138] E. L. Fosbøl, J. H. Bu , L. Øs e gaa d, C. Ande sson, C. Selme , K. K agholm, M. Schou, M.
Phelps, G. H. Gislason, T. A. Ge ds, e al., “Associa ion o angio ensin-con e ing enzyme
inhibi o o angio ensin ecep o blocke use wi h co id-19 diagnosis and mo ali y,” Jama,
ol. 324, no. 2, pp. 168–177, 2020.
60
Lis o abb e ia ions
ANN A i icial neu al ne wo k
ATC Ana omical The apeu ic Chemical
BTH BigTempHeal h
CABG Co ona y a e y by-pass g a ing
CAG Co ona y a e iog aphy
CCI Cha lson co-mo bidi y index
CHB Copenhagen Hospi al Biobank
Cox PH Cox p opo ional haza d
CRS The Danish Ci il Regis a ion Sys em
DAR The Danish Regis y o Causes o Dea h
DNPR The Danish Na ional P esc ip ion Regis y
DTB The Danish Disease T ajec o y B owse
ECG Elec oca diog am
EDHR The Eas e n Danish Hea Regis y
GDPR Gene al Da a P o ec ion Regula ion
GRACE Global Regis y o Acu e Co ona y E en s
HDI Highes densi y in e al
HDL High-densi y lipop o ein
HR Haza d a io
ICD In e na ional Classi ica ion o Diseases and Rela ed Heal h P oblems
IHD Ischemic hea disease
61
,
3
Me hods
Da a ounda ion and s udy popula ion
The main da a sou ce was he Danish Na ional Pa ien Regis y (NPR), whe e heal hca e da a om
all encoun e s wi h a Danish hospi al ha e been eco ded since 1977. The da a includes con ac ype
(i.e. in-pa ien , ou -pa ien and eme gency oom isi s), da e o con ac s a (e.g. admission), da e o
discha ge, diagnosis codes and diagnosis ype (e.g. p ima y codes ha bes desc ibe he con ac
eason and seconda y codes ha complemen he desc ip ion o he con ac )27. To ob ain
demog aphic da a on pa ien s such as da e o bi h, sex and s a us (dead o ali e), da a om NPR
was linked o he Danish Ci il Regis a ion Sys em (CRS) ia he pe sonal iden i ica ion numbe 28.
Since 1994, diagnoses in NPR ha e been epo ed using he In e na ional S a is ical Classi ica ion
o Disease and Heal h Rela ed P oblems 10 h Re ision (ICD-10), which has a hie a chical s uc u e
comp ising ICD-10 chap e s, code blocks, le el 3 and le el 4 codes29.
The NPR da ase used in his s udy co e s he pe iod 1994-2018 and con ains da a om 7 179 538
people co esponding o mo e han 142 million admissions. I con ains 4 565 dis inc le el 3 ICD-
10 codes, which we he e e e o as ICD-10 codes. P io o analysis, le el 4 codes we e unca ed o
le el 3 codes30. Pa ien s ha we e deceased by he end o he s udy pe iod we e assigned he code
Y99 and da e o dea h was ob ained om CRS28. To de ine he case popula ion, all pa ien s in NPR
who had been assigned an ICD-10 code o angina pec o is (ICD-10 code: I20), acu e myoca dial
in a c ion (ICD-10 code: I21) o ch onic ischemic hea disease (ICD-10 code I25) in he pe iod
1994-2018 we e i s iden i ied. Nex , pa ien s who we e assigned ei he o he diagnosis codes I20,
I21, o I25 be o e he age o 18 yea s we e excluded. Emig an s and ou is s we e also excluded, as
hei con ac s wi h he Danish heal h-ca e sys em a e likely o be spo adic and hus, no app op ia e
o modeling he empo ali y o IHD co-mo bidi ies. All ICD-10 codes om chap e s I-XIV
assigned as a p ima y o seconda y code (i.e., diagnosis ypes A, B o G) we e included in he
analysis. Codes assigned o less han 25 pa ien s we e excluded (Figu e 1). Da e o discha ge in
NPR was used o es ima e age a i s diagnosis ( ia linkage o CRS), iden i y pai s o diagnosis
codes ha we e assigned o he same pa ien (i.e., diagnos ic co-occu ences), de ine ime-o de ed
di ec ional diagnosis pai s and build disease ajec o ies.
Expe imen al model and iden i ica ion o diagnos ic co- occu ences
To s udy he empo al o de o he en i e spec um o mul i-mo bidi ies in IHD pa ien s, di ec ional
diagnosis pai s we e compu ed and pieced oge he o o m longe disease ajec o ies comp ising
h ee diagnoses. Disease ajec o ies we e ob ained by applica ion o a modi ied e sion o he

,
4
disease ajec o y p og am21. The wo main s eps in he p og am a e (i) quan i ica ion o he
o e ep esen a ion o diagnos ic co-occu ences be ween wo diagnosis using ela i e isk (RR) and
(ii) iden i ica ion o di ec ional diagnos ic co-occu ences whe e he empo al o de in which he
diagnoses a e assigned is s a is ically signi ican (i.e., di ec ional diagnosis pai s).
The i s s ep (i) consis s o a binomial es p ocedu e ha was de eloped in he o iginal e sion o
he disease ajec o y p og am. He e all pa ien s discha ged wi h assignmen o a ce ain diagnosis
(e.g. D1) a e conside ed exposed21. Following a p e- il e ing s ep ha uses he mean p obabili y o
assignmen o any diagnosis om all discha ges wi h diagnosis D1 (i.e., mean p obabili y
pa ame e ), RRs o he diagnosis pai s (D1, D2) we e ob ained by conside ing each discha ge as a
Be noulli sample. The algo i hm hen es s o s a is ical signi icance o he diagnosis pai s (D1,
D2) being assigned o he same pa ien wi hin i e yea s compa ed o he mean p obabili y
pa ame e o D1. Fo each pa ien ha we e assigned diagnosis D1 a discha ge (i.e., exposed), 10
000 compa ison g oups we e o med by sampling om g oups ha we e ma ched by sex, yea o
bi h and week o discha ge o co ec o e.g., seasonal a ia ion in diagnosis codes and changes in
coding p ac ices. RRs we e hen calcula ed as he a io o exposed pa ien s who we e assigned
diagnosis D2 compa ed o he a io o unexposed pa ien s who we e assigned diagnosis D2. The
le el o signi icance was se o 0.001 o gua d agains alse posi i es due o he binomial es
p ocedu e. The algo i hm was un using R . 3.4.0, Py hon . 2.7, Py hon . 3 and C++ . 1121,22.
De ining di ec ional diagnosis pai s and building disease ajec o ies
The second s ep (ii) in he p og am es ablishes he di ec ionali y o diagnos ic co-occu ences21. In
con as o p e ious e sions o he disease ajec o y p og am, a se ies o linea eg ession models
(LRs) was in oduced in his s ep. LRs we e in oduced o iden i y diagnos ic co-occu ences (D1,
D2) wi h a s a is ically signi ican di e ence be ween age a diagnoses D1 and D2, espec i ely. In
cases whe e a pa ien was assigned he same ICD-10 code mul iple imes, only he ea lies eco ded
diagnosis (wi h e e ence o end o con ac ) was included. In cases whe e a pa ien was assigned
mo e han one diagnosis o he i s ime a he same con ac , hese ins ances would be included in
he model.
The dependen a iable o he LRs was age a diagnosis and he independen a iables we e he
diagnosis pai om s ep (i) (D1, D2), he ype o diagnosis (A o B/G co esponding o p ima y o
non-p ima y diagnosis), discha ge da e, ype o pa ien (in-pa ien o ou -pa ien ), yea o bi h, and
sex. These co a ia es we e included o accoun o he possibili y o di e ences in baseline
,
5
cha ac e is ics a diagnosis due o ac o s no ela ed o he pa hogenesis, e.g., changes in coding
p ac ice o e he yea s. The P- alue o he main e ec o he diagnosis pai a iable was used o
de e mine he signi icance o di e ence in age be ween D1 and D2, whe e he p edic ed age a
diagnoses D1 and D2 de ined disease di ec ionali y. P- alues we e adjus ed using he Bon e oni
me hod. Signi icance le el was se o 0.05. The LRs we e applied o all diagnos ic co-occu ences
iden i ied in s ep (i) and i ed using S a smodels in Py hon 3.6.1031,32.
The empo al o de o he diagnos ic co-occu ences (e.g., D1 and D2) wi h s a is ically signi ican
age di e ence was de e mined by he age a diagnoses D1 and D2 ha we e compu ed om he
LRs. Due o he numbe o co a ia es, i would be cumbe some o ob ain a p edic ed age o each
subg oup, e.g., p ima y diagnosis, emales, and ou pa ien s o each discha ge yea . The e o e, he
p edic ed age was calcula ed using only he coe icien s o diagnosis pai s and ype o diagnosis, as
we obse ed ha he co a ia e o he diagnosis ype gene ally had he highes impac (smalles P-
alues) on he age di e ence (Supplemen a y igu e 1). The p edic ed age was calcula ed o D1
and D2 when assigned as p ima y diagnoses and he diagnosis ha was p edic ed o be assigned a
he younges age was de ined as he i s diagnosis in he di ec ional diagnosis (equi alen o leng h
wo ajec o ies) pai . In cases whe e he p edic ed age o D1 was less han ha o D2 he leng h wo
ajec o y would be D1 à D2. Pa ien s who we e assigned diagnoses D1 and D2 we e cha ac e ized
by ollowing he ajec o y and he o de would be he one de e mined by he LR.
To de e mine he di ec ionali y o disease ajec o ies con aining h ee diagnoses, a simila app oach
o he di ec ional diagnosis pai s was applied o se s o pa ien s, who all we e assigned diagnoses
D1, D2 and D3. Diagnosis pai s wi h a signi ican di ec ionali y whe e diagnosis D2 o one pai was
equal o diagnosis D1 o ano he pai we e pieced oge he in o a leng h h ee ajec o y. The
di ec ionali y o he disease ajec o y was de e mined by ex ac ing all pa ien s wi h diagnoses D1,
D2 and D3 and calcula ing he age o he h ee diagnoses. The age a diagnosis was calcula ed using
he se o he h ee diagnoses (i.e., D1, D2 and D3) and ype o diagnosis when assigned as p ima y
diagnosis. The diagnoses we e o de ed by es ima ed age, om younges o oldes age, e.g., he
leng h h ee ajec o y D1 à D2 à D3. As he di ec ionali y was ecalcula ed o he disease
ajec o ies, he di ec ionali y wi hin a leng h wo ajec o y could change when combined wi h
o he diagnoses. In addi ion, leng h h ee ajec o ies could es ablish new di ec ional associa ions, as
hei assemblage did no equi e he i s and he las diagnoses o be a di ec ional diagnosis pai
(co esponding o diagnosis D1 and D3 in he example abo e). In he inal analysis, only
,
6
ajec o ies compu ed based on se s o mo e han 50 pa ien s we e included; and pa ien s wi h
diagnoses D1 and D2 o D1, D2 and D3 we e hen said o ollow he esul ing ajec o y.
Using disease ajec o ies o cha ac e ize di e en IHD popula ions
To assess i he empo al o de o diagnoses was di e en be ween pa ien s wi h di e en IHD
mani es a ions, he coho was spli in o se en g oups de ined by which IHD codes hey we e
assigned. Tha is, one g oup was de ined by ha ing only I20, I21, o I25. Ano he g oup was
de ined by ha ing wo o he h ee codes, e.g. I20 and I21. A inal g oup was de ined by pa ien s
ha we e assigned I20, I21 and I25. These g oups comp ised a o al o se en dis inc index g oups.
Di ec ional diagnoses pai s we e now compu ed o each o hese g oups sepa a ely, ollowing he
same s eps as desc ibed abo e (Figu e 1).
Resul s
Coho cha ac e iza ion and o e iew o he di ec ional diagnosis pai s
The case popula ion was comp ised o a o al o 552 216 pa ien s (57.2% males) diagnosed wi h
IHD (ICD-10 codes I20, I21 o I25) in he 1994-2018 pe iod. Mean age a i s IHD diagnosis was
65.9 yea s o males and 70.8 yea s o emales (Table 1). The size o he se en index g oups
anged om 65 302 o 117 493 pa ien s. Using he en i e se o diagnosis codes assigned o pa ien s
in he case popula ion, a o al o 16 712 diagnos ic co-occu ences (D1, D2) we e iden i ied. Among
all he diagnos ic co-occu ences, he e we e 1 459 pai s wi h a s a is ically signi ican di e ence in
mean age a diagnosis D1 and D2, espec i ely. Those di ec ional diagnosis pai s we e ollowed by
546 704 pa ien s (Figu e 1).
The di ec ional diagnosis pai s ha cha ac e ized he empo ali y o mul i-mo bidi y in IHD we e
o med by a o al o 408 dis inc ICD-10 codes and gene ally, he empo al associa ions be ween
IHD co-mo bidi ies a e highly di e se and domina ed by o he diseases o he ca dio ascula
sys em as well as me abolic diseases (Figu e 2A and B). Among he 1 459 di ec ional diagnosis
pai s wi h a diagnosis code o IHD, 210 pai s con ained a diagnosis code o ei he angina pec o is
(ICD-10 code: I20), acu e myoca dial in a c ion (ICD-10 code: I21) o ch onic IHD (ICD-10 code:
I25). In 149 o hese pai s (71.0%), he codes appea ed as D2 consis en wi h he ac ha IHD is a
mul i- ac o ial disease wi h a wide spec um o e iologies and a ely he i s mani es a ion o mul i-
mo bidi y (Table 2).
,
7
Di ec ional diagnosis pai s add in o ma ion abou he coho
As expec ed, essen ial hype ension (I10), non-insulin-dependen diabe es (E11), insulin-dependen
diabe es (E10), a ial ib illa ion (I48), hea ailu e (I50) and dyslipidemia (E78), we e among he
co-mo bidi ies wi h he highes p e alence in he IHD coho (Table 1). No su p isingly, hese
diagnoses we e also among he diagnoses ha appea ed in mos dis inc di ec ional diagnosis pai s,
indica ing ha hey a e diagnosed in a wide ange o pa ien popula ions. Diagnoses ha we e
p e alen in he coho ended o be ch onic condi ions o mani es a ions o ch onic disease, e.g.,
olume deple ion (E86) (Table 2).
A quan i a i e summa y o he di ec ional diagnosis pai s e ealed addi ional cha ac e is ics
ega ding IHD co-mo bidi ies ha we e no cap u ed by he c ude co-mo bidi y coun s. Fo
example, insulin-dependen diabe es (E10), obesi y (E66), and diso de s due o he use o alcohol
(F10) we e among he mos equen ly occu ing diagnoses in he di ec ional diagnosis pai s,
al hough hey we e no among he diagnoses ha we e assigned o mos pa ien s in he popula ion.
Simila ly, al hough a ely desc ibed in he con ex o IHD, di e icula disease o he in es ine
(K57) and os eopo osis wi hou pa hological ac u e (M81) we e among he co-mo bidi ies ha
associa ed wi h mos diagnoses in a di ec ional manne . Con e sely, a condi ion ha is no ch onic
such as pneumonia (J18) was among he diagnoses assigned o mos pa ien s, ye i was no among
he mos equen ly occu ing diagnoses in he di ec ional diagnosis pai s (Tables 1 and 2).
Quali i e di e ences be ween isk ac o s, complica ions and diseases ha can be bo h
Gene ally, IHD isk ac o s appea ed as D1 in mos di ec ional diagnosis pai s, e.g., hype ension
(I10) and non-insulin dependen diabe es (E11) whe eas condi ions ha may be ela ed o bo h he
e iology o IHD and ep esen a complica ion we e equally likely o appea as ei he D1 o D2, e.g.,
hea ailu e (I50). In e es ingly, a ial ib illa ion (I48) mos o en appea ed as diagnosis D1,
indica ing ha in his popula ion in his con ex , a ial ib illa ion was o en es ablished a one o he
ea lies hospi al encoun e s. Howe e , he es ima ed age o a diagnosis o a ial ib illa ion was no
olde han ha o IHD, e.g. I20 à I48 (n = 80 560, P < 0.001). Pa ien s ollowing his di ec ional
diagnosis pai we e also among he pa ien s ha we e oldes when diagnosed wi h angina pec o is
(ICD-10: I20). Among he en mos equen di ec ional diagnosis pai s con aining I20, I21 o I25,
he e we e conside able di e ences in he age dis ibu ions o age a IHD diagnosis. Fo example,
he mean age a diagnosis o angina pec o is (I20) in he di ec ional diagnosis pai E78 à I20 (n =
114 074, P < 0.001) was 6.6 yea s below he p edic ed age a angina pec o is o pa ien s ollowing
he ajec o y I20 à I48 (n = 80 560, P < 0.001) (Table 3).
,
8
In e es ingly, we obse ed ha he empo al ela ion be ween IHD and a ial ib illa ion was no
consis en ac oss he di e en IHD subpopula ions. Fo pa ien s wi h a diagnosis code o acu e
myoca dial in a c ion and no code o angina pec o is and ch onic IHD I48 appea ed as D1 in he
di ec ional diagnosis pai , I48 à I21 (n = 11 871, P < 0.001). In con as , I21 à I48 appea ed as a
di ec ional diagnosis pai when compu ed o he popula ion who we e indexed wi h diagnosis
codes I20, I21 and I25 (n = 26 273, P < 0.001). This is consis en wi h he dual na u e o a ial
ib illa ion ha may be an age phenomenon as well as a disease complica ion. When he coho was
analyzed in i s en i e y, he diagnoses I21 and I48 did no comp ise a di ec ional diagnosis pai
u he poin ing o he complexi y o mul i-mo bidi y in his domain (Table 4).
In con as o he di e ing end be ween acu e myoca dial in a c ion and a ial ib illa ion, he age
o he diagnoses o hea ailu e (I50) and acu e espi a o y ailu e (J96) was simila in he wo
index g oups whe e i appea ed wi h a signi ican age di e ence be ween age a diagnosis o hea
ailu e and espi a o y ailu e. Among pa ien s indexed wi h I20 and I25, es ima ed age a I50 was
70.5 yea s and 75.0 yea s o acu e espi a o y ailu e (n = 3 882, p < 0.0001). Fo pa ien s indexed
wi h I20, I21, and I25 i was 70.0 yea s and 74.8 yea s, espec i ely (n = 5 161, P < 0.001) (Table
4). Taken oge he , hese obse a ions sugges a mo e indispu able empo al associa ion be ween
I50 and J96, han be ween I21 and I48 among IHD pa ien s.
Compa ison o leng h wo and leng h h ee ajec o ies
Nex , he 1 459 di ec ional diagnosis pai s we e combined in o 4 729 leng h h ee ajec o ies, i.e.,
disease ajec o ies comp ised o h ee diagnoses. Selec ed leng h wo and h ee ajec o ies wi h
sha ed diagnoses we e hen compa ed. Gene ally, he p edic ed ages o IHD in ajec o ies
con aining IHD isk ac o s, e.g., dyslipidemia (E78) and hype ension (I10) we e younge han
ajec o ies con aining a diagnosis code o IHD and no isk ac o s. In con as , among diseased
pa ien s, he p edic ed age a IHD was olde p ima ily e lec ing ha olde people a e mo e likely o
be diseased. Howe e , in ajec o ies ha con ained dea h (Y99) and a code o common IHD isk
ac o s e.g., E78 o I10, age a dea h was gene ally young (Figu e 3). Mo eo e , o some diagnoses
he p edic ed age a one diagnosis a ied conside ably be ween ajec o ies. Fo example, he
p edic ed age a angina pec o is (I20) was below 60.8 yea s o pa ien s diagnosed wi h
dyslipidemia and angina pec o is (i.e., E78 à I20). In con as , he p edic ed age a angina pec o is
was 68.0 o pa ien s diagnosed wi h angina pec o is and hea ailu e (i.e., I20 à I50). When
combined in o a leng h h ee ajec o y, i.e., E78 à I20 à I50 i was p ima ily he age a diagnosis

,
9
o hea ailu e ha changed. The p edic ed age a diagnosis o hea ailu e among pa ien s
diagnosed wi h angina was 70.2 whe eas i was 68.3 o pa ien s diagnosed wi h dyslipidemia,
angina pec o is and hea ailu e (Table 3).
In e es ingly, in some cases he leng h h ee ajec o ies e ealed ha he o de o he diagnoses was
no cons an when compa ing leng h wo ajec o ies and leng h h ee ajec o ies, e.g., D1 à D2
and D3 àD2 à D1, espec i ely. Fo example, in he popula ion o pa ien s who had been assigned
he diagnosis code o coxa h osis (M16) and acu e myoca dial in a c ion (I21), he p edic ed age
o coxa h osis was younge han ha o acu e myoca dial in a c ion (p edic ed ages 70.9 and 71.4
o M16 and I21, P < 0.001). Howe e , o pa ien s who had been assigned he diagnoses code o
dyslipidemia (E78), coxa h osis (M16) and acu e myoca dial in a c ion (I21) he o de was
e e sed, i.e., he p edic ed age o I21 was younge han ha o M16 (p edic ed ages 68.6 and 67.7
o M16 and I21). Again, his indica es he dual na u e o coxa h osis ha migh be a componen o
he me abolic synd ome o simply age- ela ed degene a ion (Table 3). In addi ion, he leng h h ee
ajec o ies cap u ed empo al ends in he coho ha leng h wo ajec o ies did no iden i y. Fo
example, albei he leng h h ee ajec o y I10 à I20 à I25 was ollowed by 105 652 pa ien s, he
diagnos ic co-occu ence o I10 and I25 was no among he di ec ional diagnosis pai s. The numbe
o pa ien s ollowing he ajec o y I10 à I20 à I25 co esponded o 66.1% o he pa ien s
ollowing he ajec o y I10 à I20 (n = 162 376, P < 0.001). A simila end was obse ed when
compa ing he diagnoses E78 and I48 (no diagnos ic co-occu ence) wi h pa ien s ollowing he
ajec o y E78 à I20 à I48 (n = 32 248), co esponding o 28.3% o pa ien s who ollowed he
ajec o y E78 à I20 (n = 114 074) (Table 3).
Discussion
We ha e p esen ed a s a egy o analyzing he empo al o de o IHD co-mo bidi ies based on
ends in na ion-wide egis y da a o mo e han 500 000 IHD pa ien s obse ed o e a pe iod o 24
yea s. By i s es ablishing di ec ional diagnosis pai s om diagnos ic co-occu ences and hen
piecing oge he disease ajec o ies, we p esen a da a-d i en cha ac e iza ion o mul i-mo bidi y
based on na ionwide heal h egis y da a. The di ec ional diagnosis pai s cap u ed empo al
associa ions ela ed o mul i-mo bidi y ha a e usually omi ed in a pu ely hypo hesis-d i en model.
Fo example, he disease ajec o y p og am iden i ied os eopo osis as one o he co-mo bidi ies ha
appea ed in mos di ec ional diagnosis pai s. Simila ly, he e we e cases whe e a ial ib illa ion
was diagnosed be o e IHD as well as a e , consis en wi h he ac ha i can ep esen an age
phenomenon as well as a disease complica ion. Finally, he disease ajec o ies se ed as a ool o
,
10
iden i y cases whe e coxa h osis was mo e likely o be a componen o me abolic synd ome as
opposed o an age- ela ed degene a ion in a single o gan sys em (Table 2).
The s udy has se e al s eng hs and limi a ions. Gene ally, he disease ajec o y app oach o e s a
no el s a egy o iden i y empo al ends in he complex mul i-mo bidi y landscape o IHD. We
a gue ha his s a egy can complemen adi ional s udies, whe e mul i-mo bidi y is assessed in a
bina y ashion, depending on hei p esence o absence11. An inhe en limi a ion wi h NPR is ha
disease his o y o he indi idual pa ien is no comple e, meaning ha all da a is condi ioned on he
ac ha he pa ien wen o he hospi al. Simila ly, he e will o cou se be cases whe e he ue age
a i s diagnosis was ea lie han 1994 and hence will no be cap u ed in he analysis. We assume
ha o pa ien s wi h many con ac s be o e 1994 his will apply o mos diagnoses ha a e hen
likely o be egis e ed a he same con ac in he obse a ion pe iod. Fo pa ien s wi h only ew
con ac s be o e he s a o he s udy ha a e close in ime o 1994 his will only ha e limi ed impac
on he esul s. Fu he , due o di e ences in yea o bi h o s udy pa icipan s combined wi h he
b oad inclusion c i e ia, di e ences in disease di ec ionali y may be con ounded by ac o s no
ela ed o e iological di e ences. Ul ima ely, u u e s udies will also include da a om he Danish
ICD-8 pe iod, i.e., be o e 1994. In con as o p e ious e sions o he disease ajec o y algo i hm,
we p esen a s a egy whe e hese con ounde s a e adjus ed o . In i s cu en o m he me hod can
only de ine a di ec ion in a p ede ined popula ion in he sense ha he di ec ion is de e mined using
he en i e dis ibu ion o he popula ion and no o he indi idual pa ien . Fu he s udies a e
needed o de e mine he ac ual o de o diagnoses in he indi idual pa ien and he eby disen angle
he ue mechanism in cases whe e one condi ion may bo h appea as a isk ac o and a
complica ion.
In sum, we belie e ha he po en ial o conduc la ge-scale s udies in na ionwide egis y da a goes
a beyond he mo e es ic i e, classical usage. Howe e , in i s cu en s a e, he Danish disease
ajec o ies a e in luenced by he inhe en limi a ions wi hin NPR and o he heal h egis ies oo ed
in he uni e sal heal hca e in o he No dic egions, such as medical su eillance bias. So, addi ional
s udies ha inco po a e o he da a ypes a e needed o iden i y he disease ajec o ies ha e lec
di e ences in disease p og ession pa e ns a ibu able o e iological di e ence. We belie e ha he
alue o s udying na ionwide heal h egis y da a comp ehensi ely ou weighs he limi a ions and
calls o u u e collabo a ions be ween basic and physician scien is s.
,
11
Re e ences
1. Timmis, A., Townsend, N., Gale, C. P., e al. Eu opean Socie y o Ca diology:
Ca dio ascula Disease S a is ics 2019. Eu . Hea J. 41, 12–85 (2020)
doi:10.1093/eu hea j/ehz859.
2. Fo man, D. E., Mau e , M. S., Boyd, C., e al. Mul imo bidi y in Olde Adul s Wi h
Ca dio ascula Disease. J. Am. Coll. Ca diol. 71, 2149–2161 (2018)
doi:10.1016/j.jacc.2018.03.022.
3. Busija, L., Lim, K., Szoeke, C., Sande s, K. M. & McCabe, M. P. Do eplicable p o iles o
mul imo bidi y exis ? Sys ema ic e iew and syn hesis. Eu . J. Epidemiol. 34, 1025–1053
(2019) doi:10.1007/s10654-019-00568-5.
4. Gup a, R. & Wood, D. A. P ima y p e en ion o ischaemic hea disease: popula ions,
indi iduals, and heal h p o essionals. The Lance 394, 685–696 (2019) doi:10.1016/S0140-
6736(19)31893-8.
5. Ades, P. A. Ca diac Rehabili a ion and Seconda y P e en ion o Co ona y Hea Disease. N.
Engl. J. Med. 345, 892–902 (2001) doi:10.1056/NEJM a001529.
6. An man, E. M. & B aunwald, E. Managing S able Ischemic Hea Disease. N. Engl. J. Med.
382, 1468–1470 (2020) doi:10.1056/NEJMe2000239.
7. T an, J., No on, R., Con ad, N., e al. Pa e ns and empo al ends o como bidi y among
adul pa ien s wi h inciden ca dio ascula disease in he UK be ween 2000 and 2014: A
popula ion-based coho s udy. PLOS Med. 15, e1002513 (2018)
doi:10.1371/jou nal.pmed.1002513.
8. Moseley, P. L. & B unak, S. Iden i ying Sepsis Pheno ypes. JAMA 322, 1416–1417 (2019)
doi:10.1001/jama.2019.12591.
9. López-O ín, C., Blasco, M. A., Pa idge, L., Se ano, M. & K oeme , G. The Hallma ks o
Aging. Cell 153, 1194–1217 (2013) doi:10.1016/j.cell.2013.05.039.
10. Ba ne , K., Me ce , S. W., No bu y, M., e al. Epidemiology o mul imo bidi y and
implica ions o heal h ca e, esea ch, and medical educa ion: a c oss-sec ional s udy. Lance
Lond. Engl. 380, 37–43 (2012) doi:10.1016/S0140-6736(12)60240-2.
11. Cosen ino, F., G an , P. J., Aboyans, V., e al. 2019 ESC Guidelines on diabe es, p e-diabe es,
and ca dio ascula diseases de eloped in collabo a ion wi h he EASD. Eu . Hea J. 41,
255–323 (2020) doi:10.1093/eu hea j/ehz486.
12. Købe , L., Swedbe g, K., McMu ay, J. J. V., e al. P e iously known and newly diagnosed
a ial ib illa ion: a majo isk indica o a e a myoca dial in a c ion complica ed by hea
,
12
ailu e o le en icula dys unc ion. Eu . J. Hea Fail. 8, 591–598 (2006)
doi:10.1016/j.ejhea .2005.11.007.
13. Benjamin, E. J., Le y, D., Vazi i, S. M., e al. Independen isk ac o s o a ial ib illa ion in
a popula ion-based coho . The F amingham Hea S udy. JAMA 271, 840–844 (1994).
14. Gheo ghiade, M., Sopko, G., De Luca, L., e al. Na iga ing he C oss oads o Co ona y
A e y Disease and Hea Failu e. Ci cula ion 114, 1202–1213 (2006)
doi:10.1161/CIRCULATIONAHA.106.623199.
15. Adelbo g, K., Széplige i, S. K., Holland-Bill, L., e al. Mig aine and isk o ca dio ascula
diseases: Danish popula ion based ma ched coho s udy. BMJ 360, k96 (2018)
doi:10.1136/bmj.k96.
16. Kalyani, R. R. Glucose-Lowe ing D ugs o Reduce Ca dio ascula Risk in Type 2 Diabe es.
N. Engl. J. Med. 384, 1248–1260 (2021) doi:10.1056/NEJMcp2000280.
17. McMu ay, J. J. V., Solomon, S. D., Inzucchi, S. E., e al. Dapagli lozin in Pa ien s wi h Hea
Failu e and Reduced Ejec ion F ac ion. N. Engl. J. Med. 381, 1995–2008 (2019)
doi:10.1056/NEJMoa1911303.
18. Lau sen, P. N., Holm ang, L., Lønbo g, J., e al. Compa ison be ween pa ien s included in
andomized con olled ials o ischemic hea disease and eal-wo ld da a. A na ionwide
s udy. Am. Hea J. 204, 128–138 (2018) doi:10.1016/j.ahj.2018.05.018.
19. Rasmussen-To ik Lau a J., Shay Ch is ina M., Ab amson Judi h G., e al. Ideal
Ca dio ascula Heal h Is In e sely Associa ed Wi h Inciden Cance . Ci cula ion 127, 1270–
1275 (2013) doi:10.1161/CIRCULATIONAHA.112.001183.
20. Fana o , A. C., Cali , R. M., Ha ing on, R. A., e al. Randomized T ials Ve sus
Common Sense and Clinical Obse a ion: JACC Re iew Topic o he Week. J. Am. Coll.
Ca diol. 76, 580–589 (2020) doi:10.1016/j.jacc.2020.05.069.
21. Jensen, A. B., Moseley, P. L., Op ea, T. I., e al. Tempo al disease ajec o ies condensed
om popula ion-wide egis y da a co e ing 6.2 million pa ien s. Na . Commun. 5, 4022
(2014) doi:10.1038/ncomms5022.
22. Siggaa d, T., Reguan , R., Jø gensen, I. F., e al. Disease ajec o y b owse o explo ing
empo al, popula ion-wide disease p og ession pa e ns in 7.2 million Danish pa ien s. Na .
Commun. 11, 4952 (2020) doi:10.1038/s41467-020-18682-4.
23. Wes e gaa d, D., Moseley, P., Sø up, F. K. H., Baldi, P. & B unak, S. Popula ion-wide
analysis o di e ences in disease p og ession pa e ns in men and women. Na . Commun. 10,
666 (2019) doi:10.1038/s41467-019-08475-9.
,
19
The Danish Na ional Pa ien Regis y in pe iod 1994-2018
n pa ien s = 7 179 538
Case popula ion de ined by ICD-10 codes I20, I21, I25
n pa ien s = 552 216
——– ——– ——–
Index g oups by ICD-10 codes (n pa ien s):
I20 (117 493), I21 (65 302), I25 (105 924), I20 I21 (14 346),
I20 I25 (96 133), I21 I25 (66 293), I20 I21 I25 (86 725)
ICD-10 codes excluded i no :
·assigned as an A, B o G code
·in chap e I-XIV
·assigned o >25 pa ien s
Pa ien s ollowing di ec ional diagnosis pai s
n pa ien s = 546 704
1
Figu e 1

,
20
,
21
,
22
IV
VI
III
I
II
V
VII
Ce ain in ec ious and pa asi ic diseases
Diseases o he blood and blood- o ming
o gans and ce ain diso de s in ol ing he
immune mechanism
Neoplasms
Endoc ine, nu i ional and me abolic diseases
Diseases o he eye and adnexa
Diseases o he ne ous sys em
Men al and beha iou al diso de s
XI
IX
VIII
X
XII
XIII
XIV
Diseases o he diges i e sys em
Diseases o he espi a o y sys em
Diseases o he ci cula o y sys em
Diseases o he ea and mas oid p ocess
Diseases o he geni ou ina y sys em
Diseases o he musculoskele al sys em and
connec i e issue
Diseases o he skin and subcu aneous issue
In e na ional S a is ical Classi ica ion o Disease and Heal h
Rela ed P oblems 10 h Re ision chap e I h ough XIV
,
23
!
!"##$%&%'()*+,-./"*%,0,
Supplemen a y igu e 1
8|Risk s a i ica ion o 72,249 pa ien s
wi h ischemic hea disease: a e -
ospec i e s udy linking p io mul i-
mo bidi y, biochemical da a, and
gene ics
89

,
1
Risk s a i ica ion o 72,249 pa ien s wi h
ischemic hea disease: a e ospec i e
s udy linking p io mul i-mo bidi y,
biochemical da a, and gene ics
Amalie D. Haue, MD1,2,*, Pe e C. Holm, MSc1,*, Ka ina Banasik, PhD1, Agne e T oen Lundgaa d, MSc1, Vic o ine P.
Muse, MEng1, Timo Röde , MSc1, Da id Wes e gaa d, PhD1, Pio J. Chmu a, MSc1, T oels Siggaa d, MSc1, Alex H.
Ch is ensen, PhD2,3, Pe e E. Weeke, PhD2, E ik Sø ensen, PhD4, Sisse R. Os owski, DMSc4,5, Daníel F. Gudbja sson,
PhD6, Hilma Hólm, PhD6, Kaspe K. I e sen, DMSc3, La s V. Købe , DMSc2,5, Hen ik Ullum, PhD7, Henning
Bundgaa d, DMSc2,5†, Sø en B unak, PhD1,8†
1No o No disk Founda ion Cen e o P o ein Resea ch, Facul y o Heal h and Medical Sciences, Uni e si y o
Copenhagen, Blegdams ej 3B, DK-2200 Copenhagen, Denma k
2Depa men o Ca diology, The Hea Cen e , Rigshospi ale , Blegdams ej 9, DK-2100 Copenhagen, Denma k
3Depa men o Ca diology, Copenhagen Uni e si y Hospi al, He le Hospi al, Bo gmes e Ib Juuls Vej 1, DK-2730
He le
4Depa men o Clinical Immunology, Copenhagen Uni e si y Hospi al, Rigshospi ale , Blegdams ej 9, DK-2100
Copenhagen, Denma k
5Depa men o Clinical Medicine, Uni e si y o Copenhagen, Rigshospi ale , Blegdams ej 3B, DK-2200 Copenhagen,
Denma k
6deCODE Gene ics-Amgen, S u luga a 8, 101 Reykja ik, Iceland
7S a ens Se um Ins i u , A ille i ej 5, DK-2300 Copenhagen, Denma k
8Copenhagen Uni e si y Hospi al, Rigshospi ale , Blegdams ej 9, DK-2100 Copenhagen, Denma k
*Deno es equal con ibu ion.
†To whom co espondence should be add essed: [email p o ec ed], No o No disk Founda ion Cen e o P o ein
Resea ch, Facul y o Heal h and Medical Sciences, Uni e si y o Copenhagen, DK-2200 Copenhagen, Denma k
Wo d coun (excl. abs ac , e e ences, - ables, igu es, and cap ions): 3,442
,
2
Key poin s
Ques ion: How can da a on p io mul i-mo bidi ies be used o isk s a i ica ion and subg ouping
o pa ien s wi h ischemic hea disease (IHD)?
Findings: Unsupe ised clus e ing iden i ied IHD pa ien subg oups de ined by he en i e mul i-
mo bidi y spec um could be s a i ied acco ding o di e en isks o seconda y ischemic e en s and
dea h om non-IHD causes. Subg ouping was suppo ed biologically by biochemical and gene ic
da a.
Meaning: Unsupe ised clus e ing disc imina ed be ween pa ien s a high- and low isk o
seconda y ischemic e en s pa ing he way o di e en ia ed managemen o pa ien s wi h IHD, i.e.,
de elopmen o p ecision medicine in his domain.
,
3
Abs ac
Impo ance: The clinical use in isk s a i ica ion o p io mul i-mo bidi y da a spanning decades;
and biochemical da a and gene ics needs o be ealized.
Objec i e: Explo e how unsupe ised clus e ing o mul i-mo bidi y om elec onic heal h eco ds
(EHRs) can be used o isk s a i y pa ien s wi h ischemic hea disease (IHD).
Design: Re ospec i e s udy ha used da a om he Danish Na ional Pa ien Regis y, EHRs om
Eas e n Denma k and gene ic da a om Copenhagen Hospi al Biobank. Unsupe ised clus e ing o
iden i ica ion o IHD subg oups based on mul i-mo bidi ies in male and emale pa ien s was
pe o med sepa a ely. Cox-p opo ional haza d models we e used o compu e haza d a ios (HRs)
o p ima y and seconda y ou comes in subg oups. Fo a subse o pa ien s, subg oups we e
cha ac e ized by a ailable biochemical da a and polygenic sco es (PGS) o 10 di e en ai s
associa ed wi h IHD. Disc imina ion by clus e -based isk s a i ica ion was es ed agains he
Cha lson and Elixhause co-mo bidi y indices (CMIs) and e alua ed using he C-index.
Se ing: A popula ion-based s udy using na ion-wide heal hca e da a. Mul i-mo bidi y was assessed
using spec um-wide ICD-10 codes assigned p io o IHD.
Pa icipan s: Pa ien s diagnosed wi h IHD in he yea s 2004-2016.
Exposu es: None.
Main Ou comes: P ima y ou come was a composi e ou come comp ised o acu e myoca dial
in a c ion, uns able angina pec o is, e ascula iza ion (no ela ed o ini ial IHD diagnosis) and
dea h caused by IHD. Seconda y ou come was any dea h no caused by IHD.
Resul s: In a coho o 72,249 pa ien s, we iden i ied 31 male and 25 emale subg oups. In o al, 11
clus e s had HR > 1 o he p ima y ou come only, and wo clus e s in each sex had HR > 1 o bo h
he p ima y and seconda y ou come. We also iden i ied wo subg oups ha had HR < 1 o he
p ima y ou come. Di e ences in biochemical da a la gely ep oduced he subg ouping. The PGS o
o al choles e ol, LDL choles e ol, diabe es, a ial ib illa ion had signi ican e ec sizes in a leas
one subg oup. Clus e C-indices we e gene ally highe o emale han o male clus e s and
compa able o CMI C-indices.
Conclusions and Rele ance: Unsupe ised clus e ing can be used o disc imina e be ween high-
and low- isk subg oups based on p io mul i-mo bidi ies ha a e consis en wi h well-known
pheno ypes. Subg oup-speci ic biochemical and PGS cha ac e is ics we e iden i ied. O e all, he
disc imina ion ob ained by he clus e s was compa able o CMIs. Fo pa ien s who we e no a
inc eased isk i was be e .
,
4
In oduc ion
Wo ldwide, ischemic hea disease (IHD) a ec s mo e han 125 million people. O e he las
decades, he e has been a decline in mo ali y a es1,2. Ye , IHD pa ien s emain a high isk o
complica ions and ha e an age-adjus ed mo ali y ha is wo o six imes highe han in people
wi hou IHD3. Nea ly 85% o IHD pa ien s a e mul i-mo bid making i a complex pheno ype, wi h
disease cou ses ha a y be ween pa ien s and be ween sexes4–6. Fu he , mul i-mo bidi y is a
c ucial componen o eliable ou come p edic ion, including suscep ibili y o disease
p og ession4,7,8. I is cu en ly being assessed using only ew diagnoses and consequen ly o en
dis ega d he ue complexi y9. As biomedical esea ch is being ans o med by an exponen ial
g ow h in pa ien p o iling da a and analy ical capabili ies, condi ions o conduc s udies ha
add ess he complexi y o mul i-mo bidi y a e now concei able8,10. Ad anced analysis o ou inely
ob ained biochemical da a is expec ed o supplemen da a on mul i-mo bidi ies10,11.
Polygenic sco es (PGS) a e becoming a means o pa ien -p o iling al hough he ue po en ial is s ill
o be con i med12. To da e PGS a e no inco po a ed in clinical ca e sys ema ically and mo e s udies
using gene ic da a ha model he isk o seconda y ischemic e en s among IHD pa ien s a e
needed13–17. This includes explo a i e s udies ha es how da a om na ion-wide heal h egis ies,
EHRs (i.e. ou inely collec ed da a ha cap u e mos aspec s o heal h ca e) and geno ypes can be
combined18–20. Agg ega ed and possibly weake phenome-wide in o ma ion ha appea s onge in
mo e homogenous subg oups may inc ease he s a is ical powe o u u e ials21,22. In u n, his may
e ine isk s a i ica ion o IHD pa ien s, allowing di e en ia ion o ea men be ween pa ien s10.
He e, we p esen a s udy o 72,249 IHD pa ien s demons a ing how unsupe ised clus e ing based
on mul i-mo bidi ies can disc imina e be ween pa ien s a high- and low- isk o seconda y ischemic
e en s and dea h o non-IHD causes (i.e., high- and low- isk pa ien s). Among low- isk pa ien s, he
pe o mance o ou model is be e han ha o o he exis ing co-mo bidi y indices. By in eg a ing
biochemical and gene ic da a, we assessed i he unsupe ised clus e ing co ela e wi h biology. In
sum, he s udy p esen s a s a egy o de ining and using mul i-mo bidi y, ha ad ances p ecision
medicine o p ac ice.
,
11
Finally, he MCL algo i hm also iden i ied clus e s whe e he seconda y ou come was inc eased,
while isks o he p ima y ou come we e no signi ican . These included 4.M04, 4.M06, 4.M07,
4.M08, 4.M10, 4.M13, 4.M29, 4.F06, 4.F11, 4.F12 and 4.F21 (Figu es 2 and 3); encompassing
dis inc clus e s cha ac e ized by cance (e.g. C44.3/C50.9), diso de s due o he use o alcohol (e.g.
F10.2), a ial ib illa ion/hea ailu e (e.g. I48.9/I50.9) and ch onic obs uc i e pulmona y diseases
(COPD) (e.g. J44.9) (Table 2). In bo h sexes, he PGS o a ial ib illa ion was signi ican in
clus e s en iched o a ial ib illa ion/hea ailu e (4.M04 and 4.F11) (Table 2). In hese clus e s,
C-indices o he CMIs we e highe han he clus e C-indices. Howe e , in some clus e s wi hou
signi ican ly al e ed isk o he p ima y o seconda y ou come, a clus e -based isk s a i ica ion
pe o med be e han CMIs, e.g., 0.76, 0.73 and 0.72 o he seconda y ou come o clus e 4.M14
(eTable 6 in he Supplemen ).
Discussion
The complexi y and scale o mul i-mo bidi y calls o de elopmen o isk s a i ica ion me hods
ha acknowledge pa ien he e ogenei y9. By applica ion o he MCL algo i hm o spec um-wide
mul i-mo bidi ies om 72,249 IHD pa ien s, we iden i ied pa ien subg oups a di e en isk o
seconda y ischemic e en s as well as dea h om non-IHD causes. Fo clus e s a inc eased isk o
seconda y ischemic e en s, HRs anged om 0.73 o 1.65 o males and 0.64 o 1.76 o emales
(Table 2). The s udy demons a ed ha he MCL algo i hm cap u ed biologically ele an pa e ns
wi hin he ICD-10 e minology, which had suppo in biochemical and gene ic da a.
The ac ha only modes e ec s on HRs we e obse ed when changing he in la ion pa ame e
sugges s o e all clus e s abili y (Figu e 2). Impo an ly, disc imina ion o su i al models i ed
using clus e s as co- a ia es was compa able he CMIs. In some subg oups a clus e -based
es ima ion e en ou pe o med CMIs (eTable 6 in he Supplemen ). In ou analyses only one clus e
(4.M01) a dec eased isk o seconda y ischemic e en s, associa ed wi h PGS dis ibu ions in
compa ison o ha o he o he clus e s (Table 2). This adds o he g owing body o li e a u e
ega ding sex-speci ic mani es a ions o IHD5. In compa ison o o he clus e ing s a egies, ou
model succeeded in including a e pheno ypes by applica ion o a lexible algo i hm ha only
depends on a limi ed se o a p io i assump ions ega ding explana o y a iables6,43. Explici ly, we
applied a da a-d i en s a egy ha disc imina ed be ween pheno ypes ha a e a isk o bo h
seconda y ischemic e en s and dea h om non-IHD causes. Fo example, in bo h sexes, clus e s

,
12
en iched o a ial ib illa ion only associa ed wi h inc eased isk o dea h om non-IHD causes and
co ela ed wi h he PGS o a ial ib illa ion. The obse a ion is consis en wi h a p e ious inding
ha an i-coagula i e he apy should no be in ensi ied in pa ien s wi h a ial ib illa ion and a ecen
pe cu aneous co ona y in e en ion7. In addi ion we show ha he associa ion be ween IHD
subg oups and PGS allows o adding gene ics and possibly he eby enhancing he clinical u ili y o
gene ic da a14,15.
The s udy has se e al s eng hs, bu also limi a ions. By pe o ming an en ichmen analysis based
on pas mul i-mo bidi ies and suppo i wi h biochemical and gene ic da a, we showcase how
in eg a ion o complemen a y da a ypes may lead o a mo e obus analysis. The s udy con i med
p io indings, while i po en ially also leads o new disco e ies. By s udying pa ien s aligned
empo ally acco ding o IHD onse and hei pas spec um-wide mul i-mo bidi ies, ou wo k
unde sco es he complexi y o mul i-mo bidi y in a model ha in p inciple can be applied o any
disease o in e es . The complemen a y analysis o biochemical da a and gene ics p o ide e idence
ha his is indeed easible.
As ou modelling s a egy allows o analysis o clus e g anula i y (as a unc ion o he in la ion
pa ame e ), he complexi y o mul i-mo bidi y is an in eg al pa o he model. E iden ly, his is also
a limi a ion. Howe e , we a gue ha he amewo k adds alue as his limi a ion is no unique o
ou model and o en p e- ixed6,19. Ye , mul i-mo bidi y is o en s udied bina ily whe e nei he he
se e i y no he o de wi h which hey we e diagnosed is aken in o accoun 9. We asse ha models
allowing o such complexi y will inc ease he likelihood o iden i ying ele an , ye weake
pheno ypic and gene ic in o ma ion in subg oups. In o ma ion ha po en ially has he capaci y o
accoun o di e ences in disease isk and bu den ha a e cu en ly unknown. Fu he , he model
equi es da a ha a e a ailable in mos heal h-ca e sys ems (i.e., ICD-10 codes). Fo success ul
ansla ion o ou indings in o ac ionable isk s a i ica ion p inciples o alue in clinical decisions
making, u he s udies ha apply he me hod a he N = 1 le el a e needed.
,
13
Re e ences
1. Nabel, E. G. & B aunwald, E. A ale o co ona y a e y disease and myoca dial in a c ion.
N. Engl. J. Med. 366, 54–63 (2012) doi:10.1056/NEJM a1112570.
2. An i h ombo ic T ialis s’ (ATT) Collabo a ion, Baigen , C., Blackwell, L., e al. Aspi in in
he p ima y and seconda y p e en ion o ascula disease: collabo a i e me a-analysis o
indi idual pa icipan da a om andomised ials. Lance 373, 1849–1860 (2009)
doi:10.1016/S0140-6736(09)60503-1.
3. WHO | P e en ion o Recu ences o Myoca dial In a c ion and S oke S udy. WHO
h ps://www.who.in /ca dio ascula _diseases/p io i ies/seconda y_p e en ion/coun y/en/in
dex1.h ml (2020).
4. Fo man, D. E., Mau e , M. S., Boyd, C., e al. Mul imo bidi y in Olde Adul s Wi h
Ca dio ascula Disease. Jou nal o he Ame ican College o Ca diology 71, 2149–2161
(2018) doi:10.1016/j.jacc.2018.03.022.
5. Pagidipa i, N. J. & Pe e son, E. D. Acu e co ona y synd omes in women and men. Na u e
Re iews Ca diology 13, 471–480 (2016) doi:10.1038/n ca dio.2016.89.
6. C owe, F., Zemedikun, D. T., Oko h, K., e al. Como bidi y pheno ypes and isk o
mo ali y in pa ien s wi h ischaemic hea disease in he UK. Hea 106, 810–816 (2020)
doi:10.1136/hea jnl-2019-316091.
7. Lopes, R. D., Heize , G., A onson, R., e al. An i h ombo ic The apy a e Acu e Co ona y
Synd ome o PCI in A ial Fib illa ion. New England Jou nal o Medicine 380, 1509–1524
(2019) doi:10.1056/NEJMoa1817083.
8. Joshi, A., Rienks, M., Theo ila os, K. & May , M. Sys ems biology in ca dio ascula
disease: a mul iomics app oach. Na Re Ca diol 18, 313–330 (2021) doi:10.1038/s41569-
020-00477-1.
9. Glynn, L. G. Mul imo bidi y: ano he key issue o ca dio ascula medicine. The Lance
374, 1421–1422 (2009) doi:10.1016/S0140-6736(09)61863-8.
10. Hemingway, H., Asselbe gs, F. W., Danesh, J., e al. Big da a om elec onic heal h eco ds
o ea ly and la e ansla ional ca dio ascula esea ch: challenges and po en ial. Eu . Hea
J. 39, 1481–1495 (2018) doi:10.1093/eu hea j/ehx487.
11. Agniel, D., Kohane, I. S. & Webe , G. M. Biases in elec onic heal h eco d da a due o
p ocesses wi hin he heal hca e sys em: e ospec i e obse a ional s udy. BMJ 361, k1479
(2018) doi:10.1136/bmj.k1479.
,
14
12. Mosley, J. D., Gup a, D. K., Tan, J., e al. P edic i e Accu acy o a Polygenic Risk Sco e
Compa ed Wi h a Clinical Risk Sco e o Inciden Co ona y Hea Disease. JAMA 323,
627–635 (2020) doi:10.1001/jama.2019.21782.
13. Buysschae , I., Ca u he s, K. F., Dunba , D. R., e al. A a ian a ch omosome 9p21 is
associa ed wi h ecu en myoca dial in a c ion and ca diac dea h a e acu e co ona y
synd ome: The GRACE Gene ics S udy. Eu Hea J 31, 1132–1141 (2010)
doi:10.1093/eu hea j/ehq053.
14. Mega, J. L., S i ziel, N. O., Smi h, J. G., e al. Gene ic isk, co ona y hea disease e en s,
and he clinical bene i o s a in he apy: an analysis o p ima y and seconda y p e en ion
ials. The Lance 385, 2264–2271 (2015) doi:10.1016/S0140-6736(14)61730-X.
15. Nikpay, M. & Mohammadzadeh, S. Phenome-wide sc eening o ai s causally associa ed
wi h he isk o co ona y a e y disease. J Hum Gene 65, 371–380 (2020)
doi:10.1038/s10038-019-0716-z.
16. Pa el, R. S., T agan e, V., Schmid , A. F., e al. Subsequen E en Risk in Indi iduals Wi h
Es ablished Co ona y Hea Disease. Ci c Genom P ecis Med 12, e002470 (2019)
doi:10.1161/CIRCGEN.119.002470.
17. Vaa a Sa u, Tikkanen Emmi, Pa kkonen Ola i, e al. Gene ic Risk Sco es P edic
Recu ence o Acu e Co ona y Synd ome. Ci cula ion: Ca dio ascula Gene ics 9, 172–178
(2016) doi:10.1161/CIRCGENETICS.115.001271.
18. Khe a Ami V. & Ka hi esan Seka . Is Co ona y A he oscle osis One Disease o Many?
Ci cula ion 135, 1005–1007 (2017) doi:10.1161/CIRCULATIONAHA.116.026479.
19. Ki k, I. K., Simon, C., Banasik, K., e al. Linking glycemic dys egula ion in diabe es o
symp oms, como bidi ies, and gene ics h ough EHR da a mining. eLi e 8, e44941 (2019)
doi:10.7554/eLi e.44941.
20. Jensen, P. B., Jensen, L. J. & B unak, S. Mining elec onic heal h eco ds: owa ds be e
esea ch applica ions and clinical ca e. Na u e Re iews Gene ics 13, 395–405 (2012)
doi:10.1038/n g3208.
21. Man el, N. & Haenszel, W. S a is ical Aspec s o he Analysis o Da a F om Re ospec i e
S udies o Disease. J Na l Cance Ins 22, 719–748 (1959) doi:10.1093/jnci/22.4.719.
22. Denny, J. C., Ri chie, M. D., Bas o d, M. A., e al. PheWAS: demons a ing he easibili y
o a phenome-wide scan o disco e gene–disease associa ions. Bioin o ma ics 26, 1205–
1210 (2010) doi:10.1093/bioin o ma ics/b q126.
,
15
23. Sø ensen, E., Ch is iansen, L., Wilkowski, B., e al. Da a Resou ce P o ile: The Copenhagen
Hospi al Biobank (CHB). In J Epidemiol dyaa157 (2020) doi:10.1093/ije/dyaa157.
24. Schmid , M., Schmid , S. A. J., Sandegaa d, J. L., e al. The Danish Na ional Pa ien
Regis y: a e iew o con en , da a quali y, and esea ch po en ial. Clin Epidemiol 7, 449–
490 (2015) doi:10.2147/CLEP.S91125.
25. Helweg-La sen, K. The Danish Regis e o Causes o Dea h. Scand J Public Heal h 39, 26–
29 (2011) doi:10.1177/1403494811399958.
26. Nielsen, A. B., Tho sen-Meye , H.-C., Belling, K., e al. Su i al p edic ion in in ensi e-
ca e uni s based on agg ega ion o long- e m disease his o y and acu e physiology: a
e ospec i e s udy o he Danish Na ional Pa ien Regis y and elec onic pa ien eco ds.
The Lance Digi al Heal h 1, e78–e89 (2019) doi:10.1016/S2589-7500(19)30024-X.
27. Schmid , M., Pede sen, L. & Sø ensen, H. T. The Danish Ci il Regis a ion Sys em as a ool
in epidemiology. Eu J Epidemiol 29, 541–549 (2014) doi:10.1007/s10654-014-9930-3.
28. Sundbøll, J., Adelbo g, K., Munch, T., e al. Posi i e p edic i e alue o ca dio ascula
diagnoses in he Danish Na ional Pa ien Regis y: a alida ion s udy. BMJ Open 6, e012832
(2016) doi:10.1136/bmjopen-2016-012832.
29. Adelbo g, K., Sundbøll, J., Munch, T., e al. Posi i e p edic i e alue o ca diac
examina ion, p ocedu e and su ge y codes in he Danish Na ional Pa ien Regis y: a
popula ion-based alida ion s udy. BMJ Open 6, e012817 (2016) doi:10.1136/bmjopen-
2016-012817.
30. En igh , A. J., Van Dongen, S. & Ouzounis, C. A. An e icien algo i hm o la ge-scale
de ec ion o p o ein amilies. Nucleic acids esea ch 30, 1575–84 (2002).
31. Acos a-He e a, M., Ke ick, M., González-Se na, D., e al. Genome-wide me a-analysis
e eals sha ed new loci in sys emic se oposi i e heuma ic diseases. Annals o he
Rheuma ic Diseases 78, 311–319 (2019) doi:10.1136/ann heumdis-2018-214127.
32. Wu ke, M. A ca alog o gene ic loci associa ed wi h kidney unc ion om analyses o a
million indi iduals. Na u e Gene ics 51, 22 (2019).
33. Malik, R., Chauhan, G., T aylo , M., e al. Mul iances y genome-wide associa ion s udy o
520,000 subjec s iden i ies 32 loci associa ed wi h s oke and s oke sub ypes. Na u e
Gene ics 50, 524–537 (2018) doi:10.1038/s41588-018-0058-3.
34. Nielsen, J. B., Tho ol sdo i , R. B., F i sche, L. G., e al. Biobank-d i en genomic
disco e y yields new insigh in o a ial ib illa ion biology. Na u e Gene ics 50, 1234–1239
,
16
(2018) doi:10.1038/s41588-018-0171-3.
35. Mahajan, A., Taliun, D., Thu ne , M., e al. Fine-mapping ype 2 diabe es loci o single-
a ian esolu ion using high-densi y impu a ion and isle -speci ic epigenome maps. Na u e
Gene ics 50, 1505–1513 (2018) doi:10.1038/s41588-018-0241-6.
36. an de Ha s Pim & Ve weij Niek. Iden i ica ion o 64 No el Gene ic Loci P o ides an
Expanded View on he Gene ic A chi ec u e o Co ona y A e y Disease. Ci cula ion
Resea ch 122, 433–443 (2018) doi:10.1161/CIRCRESAHA.117.312086.
37. Su akka, I., Ho ikoshi, M., Mägi, R., e al. The impac o low- equency and a e a ian s on
lipid le els. Na u e Gene ics 47, 589–597 (2015) doi:10.1038/ng.3300.
38. Bland, J. M. & Al man, D. G. Mul iple signi icance es s: he Bon e oni me hod. BMJ 310,
170 (1995) doi:10.1136/bmj.310.6973.170.
39. Violán, C., Roso-Llo ach, A., Fogue -Bo eu, Q., e al. Mul imo bidi y pa e ns wi h K-
means nonhie a chical clus e analysis. BMC Family P ac ice 19, 108 (2018)
doi:10.1186/s12875-018-0790-x.
40. R Co e Team. R: A Language and En i onmen o S a is ical Compu ing. (R Founda ion
o S a is ical Compu ing, 2019).
41. Mölde , F., Jablonski, K. P., Le che , B., e al. Sus ainable da a analysis wi h Snakemake.
F1000Resea ch 10, 33 (2021).
42. Edi o s, T. P. M. Obse a ional S udies: Ge ing Clea abou T anspa ency. PLOS Medicine
11, e1001711 (2014) doi:10.1371/jou nal.pmed.1001711.
43. MCL - a clus e algo i hm o g aphs. h ps://micans.o g/mcl/index.h ml.

,
17
Tables
Table 1: Coho demog aphics, ou comes, and mos p e alen non-IHD ICD-10 codes.
Coho demog aphics
To al
Males
Females
P- alue1
Numbe o pa ien s (%)
72,249
45,576 (63.1)
26,673 (36,1)
Mean age a index (SD)
63.9 (11.9)
62.9 (11.6)
65.6 (12.1)
< 0.001
Ou comes, numbe o cases
To al
Males
Females
P- alue2
Seconda y ischemic e en s (%)
11,476 (15.9)
7,891 (17.3)
3,585 (13.4)
< 0.001
n Myoca dial in a c ion
4,522
2,851
1,671
n Re ascula iza ion
4,910
3,660
1,250
n Dea h caused by IHD
1,943
1,379
664
Dea h om non-IHD causes (%)
6,177 (8.5)
3,902 (8.6)
2,275 (8.5)
Censo ed (%)
54,596 (75.6)
33,783 (74.1)
20,813 (78.0)
Ou comes, ime o e en
Mean ime o e en in yea s (SD)
P- alue1
Coun
Males
Females
Seconda y ischemic e en s
2.25 (1.91)
2.27 (1.92)
2.21 (1.89)
0.82
n Myoca dial in a c ion
1.65 (1.41)
1.63 (1.42)
1.67 (1.40)
0.31
n Re ascula iza ion
1.48 (1.34)
1.49 (1.36)
1.45 (1.31)
0.33
n Dea h caused by IHD
1.14 (1.46)
1.18 (1.47)
1.05 (1.44)
0.060
Dea h om non-IHD causes
3.36 (1.80)
3.34 (1.81)
3.39 (1.80)
0.19
Censo ed
4.27 (1.13)
4.25 (1.15)
4.30 (1.11)
0.002
To al
3.72 (1.65)
3.67 (1.67)
3.81 (1.60)
< 0.001
ICD-10
Desc ip ion
Males
Females
To al
I10
P ima y (essen ial) hype ension
14,519
10,316
24,835
E78.0
Hype choles e olemia
7,843
4,939
12,782
E11.9
Type 2 diabe es melli us: Wi hou
complica ions
4,895
2,667
7,562
I48.9
A ial ib illa ion and a ial lu e ,
unspeci ied
4,509
2,567
7,076
I50.9
Hea ailu e, unspeci ied
4,061
2,101
6,162
R07.9
Ches pain
3,450
2,423
5,873
H25.9
Senile ca a ac , unspeci ied
2,795
2.969
5,764
J18.9
Pneumonia, unspeci ied
3,329
2,265
5,504
E78.5
Hype lipidaemia, unspeci ied
3,306
1,696
5,002
J44.9
Ch onic obs uc i e pulmona y
disease, unspeci ied
2,452
2,174
4,626
1
S uden s T- es , wo-sided
2
χ2- es es ing dis ibu ion o ou comes among males s. emales. I alics no included.
,
18
Table 2: Demog aphics o selec ed clus e s. Lis o ICD-10 code desc ip ions in eTable 8.
ID
Mean age a index
HR, p ima y ou -
come (95% CI)
Adj. P-
alue
Main en ichmen
Signi ican PGS.
T ai (e ec )5
Yea s (SD)
P- alue4
ICD-10
O/E- a io
4.M01
58.5 (10.8)
< 0.001
0.93 (0.87 o 0.99)
0.030
I21.3
2.46
TC (+) I25 (-)
LDL (+)
4.M02
62.4 (10.8)
0.20
1.51 (1.40 o 1.62)
< 0.001
E10.9
28.8
E11 (+)
E11.9
11,7
H36.0
74.8
4.M19
67.1 (10.4)
< 0.001
1.28 (1.11 o 1.48)
0.04
I69.4
33.7
None
4.M20
67.6 (9.3)
< 0.001
1.65 (2.22 o 2.69)
< 0.001
I70.8
58.4
I48 (+)
4.M25
54.9 (10.9)
< 0.001
1.38 (1.15 o 1.65)
0.030
I30.9
2.6
None
I73.0
4.4
M62.6
26.4
4.F01
62.7 (11.0)
< 0.001
0.74 (0.66 o 0.84)
< 0.001
M75.4
2.1
None
I20.1
2.03
4.F02
64.0 (12.5)
< 0.001
1.73 (1.57 o 1.91)
< 0.001
E10.9
34.3
E11 (+)
E11.9
11.9
H36.0
95.0
4.F04
63.7 (12.5)
< 0.001
1.23 (1.10 o 1.38)
0.013
I21
2.52
None
4.F13
64.6 (11.6)
0.33
1.13 (0.93 o 1.38)
> 0.99
I63.9
21.1
None
4.F16
70.0 (10.5)
< 0.001
1.76 (1.48 o 2.10)
< 0.001
I73.9
43.0
I25 (-)
ID
Mean age a index
HR, seconda y ou -
come (95% CI)
Adj. P-
alue
Main en ichmen
Signi ican PGS.
T ai (e ec )
Yea s (SD)
P- alue4
ICD-10
O/E- a io
4.M04
65.8 (10.6)
< 0.001
1.51 (1.35 o 1.70)
< 0.001
I48.2
11.1
I48 (+) E11 (-)
I50.1
7.1
4.M08
66.2 (11.2)
< 0.001
3.10 (2.81 o 3.4)
< 0.001
J44.0
65.7
None
4.M10
57.9 (10.3)
< 0.001
3.14 (2.72 o 3.62)
< 0.001
F10.2
196.2
None
4.M23
71.4 (9.5)
< 0.001
1.86 (1.57 o 2.20)
<
0.001
C61
80.9
None
4.F06
67.4 (11.3)
< 0.001
3.93 (2.93 o 3.69)
< 0.001
J44.9
16.7
None
4.F11
70.1 (11.4)
< 0.001
1.42 (1.19 o 1.70)
< 0.001
I34.0
7.3
I48 (+)
I48.9
12.7
I50.0
4.0
4.F12
69.9 (9.7)
< 0.001
1.75 (1.47 o 2.08)
< 0.001
C50.9
291.7
E11 (-)
4.F21
59.0 (11.2)
< 0.001
3.13 (2.38 o 4.12)
< 0.001
F10.2
104.7
None
4Two-sided Wilcoxon es o di e ence in mean age o pa ien s in clus e agains mean age o he o he clus e s. P- alue Bon e oni adjus ed.
5ICD-10 code excep o LDL = LDL choles e ol and TC = o al choles e ol. O/E a io = Ra io o obse ed and expec ed e m equencies.
,
19
Figu e cap ions
Figu e 1: Flowcha : Da a sou ces, s udy popula ion and ou comes. G ay: Iden i ica ion.
Blue: Sc eening. Red: Eligibili y. G een: Inclusion and ou comes. NPR = The Danish
Na ional Pa ien Regis y. IHD: ischemic hea disease (ICD-10 codes I20-I25 o block R94).
CAG = co ona y a e iog aphy. CCTA: co ona y compu ed omog aphy angiog aphy. ICD-
10: In e na ional S a is ical Classi ica ion o Diseases and Rela ed Heal h P oblems 10 h
Re ision. SKS: Sundheds æsene s Klassi ika ionssys em (The Danish Heal h Au ho i y
Classi ica ion Sys em).
Figu e 2: Clus e s as a unc ion o in la ion pa ame e . Clus e s plo ed e ically a six
di e en g anula i y le els displaying hei b anching in esponse o inc eased in la ion going
le o igh . Numbe o male clus e s (A) anged om 24 o 35 and numbe o emale
clus e s (B) anged om 16 o 31 clus e s ac oss he six le els. Each black e ical ba
co esponds o one g anula i y le el (le el 1 h ough 6 going le o igh ). Leng h o he ba s
is p opo ional o he numbe o pa ien s in each clus e . Anno a ions inside black e ical
ba s deno e a single, sex-speci ic clus e uniquely pe g anula i y le el. G anula i y le el
indica ed abo e black e ical ba s. Colo key acco ding o he HR o a clus e . I a clus e is
pa ly colo ed (i.e., di e en om g ey) he en i e clus e has he HR indica ed by he colo .
Figu e 3: Risk o seconda y ischemic e en s and isk o dea h om non-IHD causes
s a i ied by clus e . Fo es plo s whe e clus e s a g anula i y le el ou a e shown agains
HR o isk p ima y (le ) and seconda y ( igh ) o males (A) and emales (B), espec i ely.
Table on he le wi h clus e , numbe o pa ien s (size) mean age a index (age) and clus e
cha ac e is ics based on O/E- a ios. “*”: No ex eme en ichmen in clus e , i.e., no e ms in
clus e wi h O/E- a io > 10. Boxes co espond o HR and ba s indica e 95% CI. A: Male
clus e s a g anula i y le el ou . B: Female clus e s a g anula i y le el ou . X-axis: HR o
a clus e ela i e o a e age HR o all o he clus e s. Y-axis: Clus e s a anged by isk o
seconda y ischemic e en s, inc easing om op o bo om. IHD: Ischemic hea disease.
GERD: Gas oesophageal e lux disease. GI- ac : Gas o-in es inal ac . HR: Hazza d a io.
CI: Con idence in e al. Fo a abula ed e sion o he esul s, see eTable 5 (su i al
analyses) and he Supplemen wi h O/E- a ios.
,
20
,
Fea u e selec ion and p e-p ocessing
Using exclusi ely da a egis e ed up un il, and including, he index da e, we ex ac ed a o al o 595 ea-
u es om NPR, EDHR, BTH and CHB. Fea u es included diagnosis codes, p ocedu e codes (e.g. imag-
ing examina ions and su gical p ocedu es), esul s om biochemical es s, clinical ea u es such as blood
p essu e, heigh , weigh , and smoking s a us, and co ona y pa hology a ime ze o. Mo eo e , a panel o
en di e en PGSs ob ained om CHB we e included and compu ed ollowing a p o ocol desc ibed in
p e ious wo k19. An o e iew o he ea u es can be ound in able 1.
Diagnosis codes (ICD-10) and p ocedu e codes (SKS/NOMESCO) we e ex ac ed om NPR and in-
cluded in he model as one-ho encoded ea u es.17 Diagnosis codes assigned mo e han 20 yea s p io
o he index da e and codes assigned o less han 1% o pa ien s in he coho we e excluded.
Biochemical es esul s and hei e e ence anges we e o iginally s o ed in he da abases Labka and BCC
and in his s udy ob ained om BTH.23 Tes s we e ei he anno a ed in acco dance wi h he Nomencla u e,
P ope ies and Uni s (NPU) o a ious local coding sys ems.24 Biochemical es s d awn mo e han i e
yea s p io o he index da e and es s pe o med on less han 5% o he coho we e excluded. Resul s
o biochemical es s we e included in he model as ca ego ical a iables. Each es could ei he ake he
alue -1, 0, o 1 indica ing i he alue was below, wi hin, o abo e he no mal e e ence ange. In cases
whe e one pa ien had mo e han one alue a ailable he mos ecen was used.
A o al o 23 clinical ea u es we e included o which 8 we e he same as hose used in he GRACE
Risk Sco e 2.0 (GRACE2.0) inpu ea u es (“Clinical cha ac e is ics 1”, Table 1). Blood p essu e and
hea a e we e ob ained om he uns uc u ed pa o he EHRs using an in-house de eloped in o ma ion
ex ac ion ool ha ecognizes ea u es om Danish clinical ex . The emaining clinical ea u es we e
ex ac ed om he s uc u ed da a. A ailable con inuous ea u es we e Z-sco e no malized p io o model
de elopmen and missing alues we e hen encoded wi h a alue o ze o. Fo he ca ego ical ea u es we
used one-ho encoding wi h an addi ional ca ego y o designa e when a iables we e missing.
Model a chi ec u e and de elopmen
To model ime- o-e en da a and allow o censo ing, we used he gene ic disc e e- ime su i al model
o neu al ne wo ks desc ibed in Gensheime 2019.25 In his model, ollow-up ime is di ided in o a ixed
numbe o in e als and he model es ima es a condi ional haza d o each in e al, i.e., he p obabili y o
ailu e gi en ha no e en has occu ed be o e ha pa icula in e al. PMHne -alpha uses 30 in e als
sepa a ed in ime such ha e en imes in he aining da a we e e enly dis ibu ed ac oss all in e als. The
implemen a ion applied he PyTo ch machine lea ning amewo k using he au ho s’ Ke as implemen-
a ion (Nne -su i al) as a e e ence.26 Fo benchma king agains GRACE2.0, he clinical ea u es we e
di ided in o he wo ca ego ies “clinical cha ac e is ics 1” ha was comp ised o he eigh GRACE2.0
inpu ea u es and “clinical cha ac e is ics 2” ha con ained he o he 15 clinical ea u es.
We used a simple mul ilaye pe cep on a chi ec u e wi h one o h ee hidden laye s and 10-200 ReLU-
ac i a ed neu ons in each o he hidden laye s. The inal ou pu laye was a ully connec ed sigmoid ac-
i a ed laye ha ou pu s condi ional haza ds o each o he 30 di e en ime poin s. We added d opou
o each o he hidden laye s o egula ize he ne wo k and p e en o e - i ing. Numbe o laye s, num-
be o neu ons, and d opou a e o each laye we e ine- uned h ough hype pa ame e op imiza ion.
In hype pa ame e -op imiza ion, we used i e- old c oss- alida ion o ob ain an op imism-co ec ed es i-
ma e o model pe o mance on he aining se . Ne wo ks we e ained by applica ion o s ochas ic g adi-
en descen , wi h he nega i e log-likelihood as he loss unc ion. P io o modelling inpu ea u es we e
g ouped in he ca ego ies “diagnosis codes”,“p ocedu e codes”,“clinical cha ac e is ics 1”,“clinical
4

,
cha ac e is ics 2”,“clinical labo a o y es s”, and “polygenic sco es” (Table 1). PMHne -alpha was
de eloped by sequen ially combining hese six da a ype ca ego ies.
Model pe o mance
PMHne -alpha was used o p edic su i al cu es on all indi iduals in he alida ion coho . Model
pe o mance was e alua ed h ough ca e ul assessmen o model disc imina ion and calib a ion based on
hese p edic ions.
To assess calib a ion and disc imina ion g aphically, he dis ibu ion o p edic ed su i al p obabili ies
a i e yea s was used o cons uc i e di e en isk s a a. Fo each o he isk g oups, he mean p e-
dic ed su i al cu e was plo ed oge he wi h a Kaplan-Meie es ima e o he obse ed su i al.27 Fo
model calib a ion, pa ien ou comes we e plo ed agains i e-yea mo ali y p edic ions and a locally
es ima ed sca e plo smoo he was used o ob ain a non-pa ame ic es ima e o he calib a ion cu e.28
The smoo he was es ima ed using he loess unc ion in R wi h he de aul bandwid h and he a ea 𝑎
be ween he eg ession line and he ideal line 𝑥 = 𝑦 was used as a measu e o model calib a ion.15 Using
boo s ap esampling, con idence in e als o he calib a ion sco e could be ob ained.29
To complemen he g aphical me hods, we also calcula ed ime-dependen AUC sco es as well as Ha el’s
C-index. Time-dependen AUC ( dAUC) sco es we e calcula ed using he dROC R-package.30 Ha el’s
C-index was calcula ed using he co .cens unc ion om he Hmisc R-package.31 Time-dependen
AUC sco es we e e alua ed a six mon hs, one yea , h ee yea s, and i e yea s a e ime ze o.
To u he benchma k PMHne -alpha we calcula ed he GRACE2.0 o all pa ien s in he alida ion
coho .6We ex ac ed he ja asc ip sou ce code o he GRACE2.0 web ool using Google Ch ome’s De-
elope Tools. The ja asc ip code was hen manually con e ed o an R-package o enable au oma ised
compu a ion o GRACE2.0 on he en i e coho . Since GRACE2.0 does no allow o missing ea u es,
we impu ed missing a iables using he missFo es R-package.32 Impu a ion was only pe o med o
acqui e a GRACE2.0 sco e o all pa ien s; missing alues we e le missing in PMHne -alpha. Explain-
abili y analysis was pe o med using SHapley Addi i e exPlana ions (SHAP) alues. The ela i e ea u e
impo ance o he di e en inpu ea u e ca ego ies was compu ed by summa izing he absolu e SHAP
alues o each ea u e ca ego y33.
E hics decla a ion
S udy design, me hods, and esul s we e epo ed in ag eemen wi h he TRIPOD s a emen .34,35 The
sc ip s and sou ce code a e a ailable upon easonable eques o he au ho s. The s udy was app o ed by
The Danish Na ional Commi ee on Heal h Resea ch E hics p ojec IDs 60833 and 1708829.
Resul s
Baseline cha ac e is ics o de i a ion coho
A o al o 39 746 pa ien s (67·3% males) we e included in he de i a ion coho . I was andomly spli
in o a aining (n = 34 746) and a es ing se (n = 5000) used o model de elopmen and alida ion,
espec i ely (Figu e 1). Table 2 summa izes he pa ien baseline cha ac e is ics and he da a composi ion
o he aining se , es ing se , and he de i a ion coho as a whole. Mean age a index da e was 66·1
5
,
yea s and 19·5% o he coho p esen ed wi h a STEMI. The es ic ed mean ollow up ime was 1651
days (SE: 2·58) and he p ima y ou come (all-cause mo ali y) was obse ed in 21·3% o he subjec s in
he aining se and 22·2% o hose in he es ing se (Figu e S1).
Pe o mance and benchma king o PMHne -alpha
PMHne -alpha was de eloped by sequen ially adding and combining he six di e en inpu ea u e ypes
lis ed in Table 2, which led o six di e en in e media e models labelled 1 h ough 6. Th ough hype -
pa ame e op imiza ion, he bes pe o ming combina ion o hype pa ame e s we e iden i ied and hen
ained on he en i e aining da a (n = 34746). The a ious PMHne models hen unde wen alida ion
using he unseen es ing da a (n = 5000). Figu e 2 displays he dAUC a six-mon hs, one-yea , and
h ee-yea s e alua ion ime poin s o he di e en e sions o PMHne -alpha de ined by hei espec i e
inpu ea u e combina ions. The p elimina y esul s ha e shown ha he pe o mance o PMHne -alpha
emains unchanged by simply adding he PGSs o model 6, indica ing ha he in o ma ion om he PGSs
may be included in he o he da a ypes.
Se ing as a e e ence o he obse ed pe o mances, he GRACE2.0 sco e o six-mon hs, one-yea , and
h ee-yea s was calcula ed on he en i e aining se . The eigh a iables used in GRACE2.0 we e a ail-
able o 51·4% o he coho . Impu a ion was pe o med o pa ien s wi h missing alues (see Me hods).
The dAUC o GRACE2.0 was 0·768, 0·768, and 0·729 a six-mon h, one-yea , and h ee-yea imepoin s,
espec i ely ( igu e S2). Thus, excep o model 1, PMHne -alpha ou pe o med GRACE2.0 in p edic ion
all-cause mo ali y a all he e alua ed imepoin s (Figu e 2).
Gene ally, he pe o mance o PMHne -alpha imp o ed as mo e ea u e ca ego ies we e added. The e -
sion o PMHne -alpha wi h diagnosis codes only, ou pe o med GRACE2.0 in p edic ion o h ee-yea
all-cause mo ali y bu had poo e disc imina ion a six mon hs and one yea . In e es ingly, he e sion
o PMHne -alpha ained using he ea u e ca ego y “clinical cha ac e is ics 1” (co esponding o he
GRACE2.0 inpu ea u es) ou pe o med GRACE2.0 a he six-mon hs and h ee-yea s imepoin s, likely
because non-linea co ela ions a e conside ed by PMHne -alpha. A he one-yea e alua ion imepoin
he pe o mance o his e sion o PMHne -alpha was almos iden ical o ha o GRACE2.0 o he pa-
ien s wi h missing alues. Howe e , he pe o mance o GRACE2.0 on he pa ien s wi hou any missing
alues was be e han ha o PMHne -alpha (Figu e 2). In con as o GRACE2.0, he p edic i e pe -
o mance o PMHne -alpha was he same o pa ien s wi h and wi hou missing alues (Table 4). The
bes pe o ming e sion o PMHne -alpha was model 6 ha based i s p edic ions on all ea u e ca ego ies
(Figu e 2). The model has excellen model disc imina ion (Table 3) and a good calib a ion wi h a sco e,
1−𝛼, o 0·944 (CI: [0·925;0·964]) (Figu e 3). In e ms o calib a ion, he model p o ided oo pessimis ic
p edic ions o pa ien s in he 50% isk ange bu o he wise looked good (Figu e S3).
Explainabili y analysis o PMHne -alpha
Nex , he p edic i e alues (i.e., ea u e impo ance) o he di e en ea u es was quan i ied using SHAP
alues o PMHne -alpha model 6 (in he ollowing e e ed o as he inal e sion). Globally, age was
he mos p edic i e ea u e and i s ea u e impo ance inc eased a inc easing imepoin s. Tha is, age
con ibu ed mo e o he p edic ion a he h ee-yea s han a he six-mon hs e alua ion imepoin . The
sequen ial addi ion o inpu ea u es ypes showed ha diagnosis codes we e an impo an addi ion o he
model ha only used GRACE2.0 ea u es. Howe e , in he inal e sion o PMHne -alpha, no diagnosis
codes we e among he op anking ea u es. This indica es ha he pe o mance o a model wi h many
6
,
ea u es wi h ela i ely li le p edic i e alue can be be e han one wi h ew highly p edic i e ones
(Figu e 4).
In he inal e sion o PMHne -alpha, clinical ea u es used o compu e he GRACE2.0, e.g. Killip class,
p esen a ion wi h STEMI (yes/no), and SBP we e among he mos p edic i e ones. Aside om age, he
wo mos p edic i e ea u es (la ges SHAP alues) we e co ona y pa hology and smoking s a us, which
a e no included in he GRACE2.0 sco e. Like age, co ona y pa hology was consis en ly impo an ac oss
he h ee e alua ion imepoin s. Ye , he impo ance dec eased sligh ly o e ime sugges ing ha mos
pa ien s wi h se e e co ona y pa hology died ela i ely soon a e index (i.e., ea lie ime poin s). The
opposi e was ue o age, whe e he ea u e impo ance inc eased a inc easing ime poin s (Figu e 4).
Fo he ea u e co ona y pa hology, DA and 1VD we e p edic i e ea u es d i ing owa ds dea h whe eas
2VD and 3VD we e p edic i e ea u es d i ing owa ds su i al. O he ou possible ca ego ies o
co ona y pa hology, 3VD had he la ges impac on he p edic ion. The impac on model ou pu o DA
and 1VD was abou he same. Unlike age and co ona y pa hology, he ea u e impo ance o smoking was
ela i ely cons an o e he h ee di e en imepoin s. SHAP alues o smoking we e nega i e, meaning
ha hey con ibu ed nega i ely o he ou come (Figu e 4). The p edic i e impo ance o he ea u e
ca ego ies was also assessed collec i ely. Al hough he biochemical es s we e gene ally no among he
highes anked ea u es, he inpu ea u e ca ego y “clinical labo a o y es s” had he highes o e all
ea u e impo ance (Table 5)
Explainabili y analysis o PMHne -alpha o one pa ien , N = 1
Finally, he ea u e impo ance was e alua ed o ou indi idual cases wi h a p edic ed h ee-yea s su -
i al chance be ween 0·75-1·00, 0·50-0·75, 0·25-0·50, and 0·00-0·25, espec i ely. Each case was ep-
esen ed by he linea sum o he SHAP alues o he p edic ions a he h ee e alua ion imepoin s.
P edic i e ea u es wi h he la ges absolu e SHAP alues we e anno a ed on igu e 5. No SHAP alues
below 0·02 we e anno a ed (Figu e 5).
Case 1 was cha ac e ized by e y ew SHAP alues abo e he h eshold o anno a ion indica ing ha
he pa ien had ew hospi aliza ions p io o index da e. The ICD-10 code N81 ( emale geni al p olapse)
was one o ew anno a ed alues and con ibu ed posi i ely o he es ima ed chance o su i al. In con-
as , cases 2,3, and 4had o e lapping op- anked ea u es. Fo example, he ICD-10 code J44 (ch onic
obs uc i e pulmona y disease) was among he op- anked p edic i e ea u es wi h he la ges absolu e
SHAP alue. In all h ee cases i con ibu ed nega i ely o p edic ions (Figu e 5). In con as , 1VD con-
ibu ed nega i ely o he p edic ion o case 3 and posi i ely o he p edic ion o case 4. Simila ly, age
con ibu ed nega i ely o he p edic ion o case 3 (82·7 yea s) and posi i ely o he p edic ion o case
4(57·9 yea s). Age was no among he op-sco ing p edic i e ea u es wi h he la ges absolu e SHAP
alue o case 2. Fo case 2 he mos p edic i e ea u es we e comp ised o diagnosis codes, p ocedu es
codes, and wo biochemical es s ha we e below e e ence ange, whe e low a e ial pa ial p essu e o
oxygen and low plasma glucose we e among he mos p edic i e biochemical ea u es. In con as , low
plasma lac a e was among he posi i e p edic i e ea u es o case 3. In e es ingly, missing in o ma ion
ega ding amilial IHD was among he mos p edic i e ea u es o case 4 (Figu e 5).
Gene ally, he ea u e a ibu ions o PMHne -alpha was consis en wi h physiological esponse mecha-
nisms, e.g. low oxygen sa u a ion and low plasma glucose s. plasma lac a e and he chances o su i al
(Figu e 5). Thus, he SHAP analysis se es as a p oo -o -p inciple o model de elopmen . In addi ion, he
obse a ions de i ed om he SHAP analysis demons a e ha he ea u e impo ance in PMHne -alpha
is con ex -dependen . Fo example, he analysis showed ha in case 1 he age o 75·2 yea s con ibu ed
7
,
nega i ely, whe eas he age o 82·7 con ibu ed posi i ely o he p edic ion o case 3; bu ha he o e -
all su i al p obabili y was highe o case 1 han o case 3. Las ly, he SHAP analysis illus a es ha
PMHne -alpha can in eg a e in o ma ion o an adminis a i e na u e as well as in o ma ion e lec ing
physiology, i.e., missing in o ma ion ega ding amilia IHD as well as biochemical es esul s.
Discussion
In his s udy we ha e de eloped PMNne -alpha, a neu al ne wo k-based model o p edic ion o all-cause
mo ali y in IHD pa ien s. PMHne -alpha ou pe o med GRACE2.0 when he p edic ions we e based on
same eigh inpu ea u es as hose used in GRACE2.0. By ex ending he olume and a ie y o inpu
ea u es using six di e en ea u e ca ego ies amoun ing o a o al o 595 dis inc inpu ea u es, he
dAUC o PMHne -alpha imp o ed om 0·75 o 0·876 (Figu e 2).
I is a de ining a ibu e o PMHne -alpha ha ea u es we e included wi h p ac ically no a p io i assump-
ions ega ding po en ial p edic i e impo ance. Simila ly, no assump ions ega ding linea dependency
be ween a iables we e made, which allowed us o include inpu ea u es based on a ailabili y. This con-
as s wi h mos p e iously published p edic ion models in his domain, whe e model ea u es a e o en
limi ed o he ew mos p edic i e ones8,36. PMHne -alpha is e idence ha i is indeed an ad an age o
include mo e han he mos p edic i e ea u es in a isk sco ing model by ou pe o ming GRACE2.0.
Using SHAP alues, we we e able o ank he le el o in o ma ion o he di e en inpu ea u es and
cha ac e ize he model sys ema ically.33 O e all, he SHAP analysis se ed as a p oo -o -p inciple and
also exempli ied ye ano he ad an age wi h PMHne -alpha o e adi ional isk sco ing algo i hms as
PMHne -alpha was able o es ima e he su i al a e o all pa ien s in he coho i espec i e o da a
a ailabili y (Figu e 5). Explici ly, GRACE2.0 could no p o ide a isk es ima e o case 1 wi hou ea u e
impu a ion, whe eas PMHne -alpha success ully es ablished a su i al cu e o his pa ien . Based on
he esul s om he SHAP analysis, we could assess ha he es ima e seemed easonable based on he
p edic i e a ibu ion o he a ailable ea u es (Figu e 2). Fu he , PMHne -alpha enabled us o iden i y
p edic i e ea u es ha a e no included in con en ional models and he SHAP analysis allowed us o
quan i y hei impo ance. Fo example, smoking, which is no included in GRACE2.0, was among he
mos in o ma i e ea u es a all h ee e alua ion imepoin s. I was among he mos p edic i e ea u es a
all h ee ime-poin s wi h a nea ly cons an SHAP alue (Figu e 4). This adds o he li e a u e ega ding
lack o imp o ed su i al in IHD pa ien s, wi h a his o y o smoking.4Taken oge he , he lexibili y o
PMHne -alpha wi h ega ds o inpu ea u es is highly aluable in mode n medicine, whe e da a a ail-
abili y and access a y be ween coun ies. Thus, PMHne -alpha can in p inciple be ained on a da a
se based on ea u e a ailabili y in ano he coun y o heal hca e sys em, independen o he ea u es ha
we e a ailable a he ime and place o de elopmen . Ongoing wo k is cen e ed on ex e nal alida ion and
in e p e a ion o he pe o mance when PGSs a e added. The ac ual ealized disease his o y may con ain
he same in o ma ion as he cu en PGSs could con ibu e. Howe e , he lack o imp o emen could also
be due o he cu en a chi ec u e o he neu al ne wo k. As we a e s ill se ing up expe imen s a ge ing
he analyses o he p edic i e pe o mance o he PGSs, hese esul s a e no discussed ho oughly in his
e sion o he manusc ip .
This s udy has se e al s eng hs bu also some limi a ions. Fi s , in his s udy PMHne -alpha was only
benchma ked agains one o he ex ensi ely alida ed isk sco es in his domain. Howe e , gene ally
o he isk sco es, e.g. TIMI, a e no supe io o GRACE2.0 and he e is e en e idence ha GRACE2.0
has he bes disc imina ion in some popula ions37. Fu he , PMHne -alpha was de eloped in a seconda y
p e en ion se ing and hus, he F amingham isk sco e is unsui able o benchma king.5Howe e , he
de i a ion coho was di e en om ha o GRACE2.0 as i included pa ien s wi h acu e mani es a ions
8
,
o IHD as well as ch onic IHD (Table 2). In con as o he F amingham, TIMI and GRACE2.0 sco es,
PMHne -alpha is no only based a limi ed se o isk ac o s only5. O e all we ake ou p edic ion model
se e al s eps u he han con en ional isk sco es, by de eloping a model ha is lexible wi h espec
o olumes and a ie y o inpu da a. We show ha he pe o mance o PMHne -alpha is supe io o
con en ional su i al models based on he same ea u es and ha he pe o mance o PMHne -alpha
inc eases g adually when he numbe o ea u es is inc eased (Figu e 2).
Ha ing de eloped and es ed PMHne -alpha, we a gue ha he e is a need o e hink he usage o clinical
da a wi hin medical esea ch o mode n medicine o ake ull ad an age o he inc ease in da a olumes,
a ie y, and analy ical capabili ies. Ano he ad an age o he s a egy ha we applied o isk assessmen
is ha i may o e come he exis ing ba ie be ween isk sco ing algo i hms and clinical p ac ice.38 Explic-
i ly, PMHne -alpha could po en ially be included as an applica ion in EHRs o eal- ime isk assessmen
o IHD pa ien s, as i compu es a isk o all indi idual pa ien s based on he a ailable da a (Figu e 5).
We ha e howe e no es ed sys ema ically a wha poin he missingness becomes a p oblem. Finally,
PMHne -alpha includes amewo k o model explainabili y, which is a common conce n agains he us-
age o ML models wi hin he medicine9. In his s udy, we showcase ha an explainable model is a c ucial
componen o success ul su i al modeling based on ML, such as neu al ne wo ks. We showcase ha
explainabili y analysis may bo h p o ide g ounds o model con idence as well as disco e y o impo an
p edic i e ea u es ha ha e no been included in adi ional su i al models.
9

,
Re e ences
1 James SL, Aba e D, Aba e KH, e al. Global, egional, and na ional incidence, p e alence, and
yea s li ed wi h disabili y o 354 diseases and inju ies o 195 coun ies and e i o ies, 1990–
2017: a sys ema ic analysis o he Global Bu den o Disease S udy 2017. The Lance 2018; 392:
1789–858.
2 Takahashi K, Se uys PW, Fus e V, e al. Rede elopmen and alida ion o he SYNTAX sco e
II o indi idualise decision making be ween pe cu aneous and su gical e ascula isa ion in pa-
ien s wi h complex co ona y a e y disease: seconda y analysis o he mul icen e andomised
con olled SYNTAXES ial wi h ex e nal coho alida ion. The Lance 2020; 396: 1399–412.
3 Gup a R, Wood DA. P ima y p e en ion o ischaemic hea disease: popula ions, indi iduals, and
heal h p o essionals. The Lance 2019; 394: 685–96.
4 A bel Y, Fi zGe ald G, Yan AT, e al. Tempo al ends in all-cause mo ali y acco ding o smoking
s a us: Insigh s om he Global Regis y o Acu e Co ona y E en s. In e na ional Jou nal o
Ca diology 2016; 218: 291–7.
5 Wilson PWF, D’Agos ino RB, Le y D, Belange AM, Silbe sha z H, Kannel WB. P edic ion o
Co ona y Hea Disease Using Risk Fac o Ca ego ies. Ci cula ion 1998; 97: 1837–47.
6 Fox KAA, Fi zGe ald G, Puymi a E, e al. Should pa ien s wi h acu e co ona y disease be s a i-
ied o managemen acco ding o hei isk? De i a ion, ex e nal alida ion and ou comes using
he upda ed GRACE isk sco e. BMJ Open 2014; 4: e004425.
7 Hung J, Roos A, Kadesjö E, e al. Pe o mance o he GRACE 2.0 sco e in pa ien s wi h ype 1
and ype 2 myoca dial in a c ion. Eu opean Hea Jou nal 2021; 42: 2552–61.
8 An man EM, Cohen M, Be nink PJLM, e al. The TIMI Risk Sco e o Uns able Angina/Non–ST
Ele a ion MI. JAMA 2000; 284: 835.
9 Rajkoma A, Dean J, Kohane I. Machine Lea ning in Medicine. New England Jou nal o
Medicine 2019; 380: 1347–58.
10 Topol EJ. High-pe o mance medicine: he con e gence o human and a i icial in elligence. Na-
u e Medicine 2019; 25: 44–56.
11 B aje N, Cozzi B, Gao M, e al. P ospec i e and Ex e nal E alua ion o a Machine Lea ning
Model o P edic In-Hospi al Mo ali y o Adul s a Time o Admission. JAMA Ne wo k Open
2020; 3: e1920733.
12 Gibson WJ, Na ee T, T a is R, e al. Machine lea ning e sus adi ional isk s a i ica ion me h-
ods in acu e co ona y synd ome: a pooled andomized clinical ial analysis. Jou nal o Th om-
bosis and Th ombolysis 2019; 49: 1–9.
13 D’Ascenzo F, De Filippo O, Gallone G, e al. Machine lea ning-based p edic ion o ad e se
e en s ollowing an acu e co ona y synd ome (PRAISE): a modelling s udy o pooled da ase s.
The Lance 2021; 397: 199–207.
10
,
14 Mo wani M, Dey D, Be man DS, e al. Machine lea ning o p edic ion o all-cause mo ali y in
pa ien s wi h suspec ed co ona y a e y disease: a 5-yea mul icen e p ospec i e egis y analysis.
Eu opean Hea Jou nal 2016; : ehw188.
15 S eele AJ, Denaxas SC, Shah AD, Hemingway H, Luscombe NM. Machine lea ning models
in elec onic heal h eco ds can ou pe o m con en ional su i al models o p edic ing pa ien
mo ali y in co ona y a e y disease. PLOS ONE 2018; 13: e0202344.
16 Özcan C, Juel K, Flens ed Lassen J, on Kappelgaa d LM, Mo ensen PE, Gislason G. The Danish
Hea Regis y. Clin Epidemiol 2016; 8: 503–8.
17 Schmid M, Schmid SAJ, Sandegaa d JL, Eh ens ein V, Pede sen L, Sø ensen HT. The Danish
Na ional Pa ien Regis y: a e iew o con en , da a quali y, and esea ch po en ial. Clinical
Epidemiology 2015; : 449.
18 Nielsen AB, Tho sen-Meye H-C, Belling K, e al. Su i al p edic ion in in ensi e-ca e uni s
based on agg ega ion o long- e m disease his o y and acu e physiology: a e ospec i e s udy o
he Danish Na ional Pa ien Regis y and elec onic pa ien eco ds. The Lance Digi al Heal h
2019; 1: e78–89.
19 Sø ensen E, Ch is iansen L, Wilkowski B, e al. Da a Resou ce P o ile: The Copenhagen Hospi al
Biobank (CHB). In e na ional Jou nal o Epidemiology 2021; 50: 719–720e.
20 Helweg-La sen K. The Danish Regis e o Causes o Dea h. Scandina ian Jou nal o Public
Heal h 2011; 39: 26–9.
21 Schmid M, Pede sen L, Sø ensen HT. The Danish Ci il Regis a ion Sys em as a ool in epidemi-
ology. Eu opean Jou nal o Epidemiology 2014; 29: 541–9.
22 A lo S, Celisse A. A su ey o c oss- alida ion p ocedu es o model selec ion. S a is ics Su eys
2010; 4. DOI:10.1214/09-ss054.
23 G ann, E ichsen R, Nielsen, F øsle , Thomsen R. Exis ing da a sou ces o clinical epidemiology:
The clinical labo a o y in o ma ion sys em (LABKA) esea ch da abase a Aa hus Uni e si y,
Denma k. Clinical Epidemiology 2011; : 133.
24 A end JFH, Hansen AT, Lade oged SA, Sø ensen HT, Pede sen L, Adelbo g K. Exis ing Da a
Sou ces in Clinical Epidemiology: Labo a o y In o ma ion Sys em Da abases in Denma k. Clin-
ical Epidemiology 2020; Volume 12: 469–75.
25 Gensheime MF, Na asimhan B. A scalable disc e e- ime su i al model o neu al ne wo ks.
Pee J 2019; 7: e6257.
26 Gi Hub - MGensheime /nne -su i al: Disc e e-Time Su i al Model o Neu al Ne wo ks.
Gi Hub. h ps://gi hub.com/MGensheime /nne -su i al (accessed Aug 18, 2021).
27 Roys on P, Al man DG. Ex e nal alida ion o a Cox p ognos ic model: p inciples and me hods.
BMC Medical Resea ch Me hodology 2013; 13: 33.
28 Ha ell FE, Lee KL, Ma k DB. Mul i a iable p ognos ic models: issues in de eloping models,
e alua ing assump ions and adequacy, and measu ing and educing e o s. S a Med 1996; 15:
361–87.
11
,
29 Hes e be g T. Boo s ap. Wiley In e disciplina y Re iews: Compu a ional S a is ics 2011; 3:
497–526.
30 Rod íguez-Ál a ez MX, Mei a-Machado L, Abu-Assi E, Raposei as-Roubín S. Nonpa ame ic
es ima ion o ime-dependen ROC cu es condi ional on a con inuous co a ia e. S a is ics in
Medicine 2016; 35: 1090–102.
31 Hmisc: Ha ell Miscellaneous. h ps://CRAN.R-p ojec .o g/package=Hmisc (accessed Aug 18,
2021).
32 S ekho en DJ, Buhlmann P. MissFo es --non-pa ame ic missing alue impu a ion o mixed- ype
da a. Bioin o ma ics 2011; 28: 112–8.
33 Lundbe g SM, E ion G, Chen H, e al. F om local explana ions o global unde s anding wi h
explainable AI o ees. Na u e Machine In elligence 2020; 2: 56–67.
34 Collins GS, Rei sma JB, Al man DG, Moons KGM. T anspa en Repo ing o a Mul i a iable
P edic ion Model o Indi idual P ognosis o Diagnosis (TRIPOD). Ci cula ion 2015; 131: 211–
9.
35 Collins GS, Moons KGM. Repo ing o a i icial in elligence p edic ion models. The Lance 2019;
393: 1577–9.
36 Boe sma E, Piepe KS, S eye be g EW, e al. P edic o s o Ou come in Pa ien s Wi h Acu e
Co ona y Synd omes Wi hou Pe sis en ST-Segmen Ele a ion. Ci cula ion 2000; 101: 2557–
67.
37 A agam KG, Tamhane UU, Kline-Roge s E, e al. Does Simplici y Comp omise Accu acy in
ACS Risk P edic ion? A Re ospec i e Analysis o he TIMI and GRACE Risk Sco es. PLoS
ONE 2009; 4: e7947.
38 Man ini O, Bugia dini R. Ba ie s o clinical isk sco es adop ion. Eu opean Hea Jou nal
2007; 28: 1045–6.
12
,
Tables
Table 1
Table 1: O e iew o he di e en inpu ea u e ca ego ies.
Ca ego y Fea u es Hype pa ame e s
Clinical cha ac e is ics 1 Age, pulse, sys olic blood
p essu e, ca diac a es a index,
abno mal ca diac enzymes
killip class, c ea inine,
s -segmen de ia ion
Clinical cha ac e is ics 2 Abno mal ECG, CCS class,
dias olic blood p essu e,
co ona y a e y dominance,
amilia y IHD, pa ien heigh ,
icd-de ice o pm, ischemia es ,
LVEF, NYHA class, sex,
smoking s a us, co ona y
pa hology, pa ien weigh
Diagnosis codes 422 di e en le el-4 ICD-10
diagnosis codes. Kep as-is o
con e ed o le el-3, block, o
chap e du ing model
de elopmen .
• icd10 le el
• ime cu o
P ocedu e codes 154 di e en NOMESCO
p ocedu e codes co esponding
o a ious adiological
examina ions and su gical
p ocedu es
• ime cu o
Clinical labo a o y es s 85 di e en lab es s wi h esul s
ca ego ised as below,wi hin, o
abo e he e e ence ange
• ime cu o
Polygenic sco es 10 di e en ai s • winso iza ion
13
,
Figu e 5: Speci ic model e alua ion o N = 1 o ou dis inc pa ien s. P edic ion su i al a e ep-
esen ed as SHAP alues o ou dis inc pa ien s in he coho ep esen ed as case 1,case
2,case 3, and case 4 eading le o igh , op o bo om. Cases a e om ou di e en isk
s a a wi h a p edic ed isk o 15.1 (case 1), 34.6 (case 2), 72.8 (case 3), and 95.6 (case 4) a
he h ee-yea s e alua ion imepoin . X-axis: The h ee e alua ion imepoin s, i.e. six-mon hs,
one-yea , and h ee-yea s. Y-axis: P edic ed su i al a e o he ou di e en cases a he
in e sec ion o he wo colo s blue and ed. Blue ep esen s ea u es ha con ibu e posi i ely
o he p edic ion. Red ep esen s ea u es ha con ibu e nega i ely o he p edic ion. Highes
anked ea u es we e anno a ed abo e a SHAP alue o 0.02 we e anno a ed. SHAP: SHapley
Addi i e exPlana ions.
20

,
Figu e S1: Kaplan-Meie es ima es o he PMHne de i a ion coho . Kaplan-Meie es ima es o
he aining (blue) and es ing ( ed) se o 34749 and 5000 pa ien s, espec i ely. X-axis:
Follow-up ime in yea s. Y-axis: Su i al pe cen age.
Figu e S2: Time-dependen ecei e ope a ing cha ac e is ics o he GRACE2.0 on he PMHne
es se . Th ee dROCs co espondong o he h ee e alua ion six-mon hs, one-yea s and
h ee-yea s e alua io imepoin s goind le o igh . Red: Pa ien s in he es se whe e all
GRACE2.0 inpu ea u es we e a ailable. Blue: Pa ien s whe e impu a ion was pe o med
p io o compu a ion o he GRACE2.0 due o missing ea u es.
21
,
Figu e S3: Calib a ion cu e o PMHne -alpha model 6 Calib a ion cu e showing he obse ed
s. p edic ed all-cause mo ali y a 5 yea s. The do ed line ep esen s he ideal calib a ion
line. The a ea be ween he non-pa ame ic loess cu e and he ideal line can be used as a
measu e o calib a ion.
22
10 |Polypha macy and d ug dosage
modi ica ions: a longi udinal anal-
ysis o 3.5 million elec onic heal h
eco ds
135
,
1
Polypha macy and d ug dosage modi ica ions:
1
A longi udinal analysis o 3.5 million elec onic
2
heal h eco ds
3
C is ina Leal Rod íguez1, Gianluca Mazzoni1, Amalie Dahl Haue1, Robe E iksson1,2, Jo ge
4
He nansanz Biel1, Lisa Can well1, Da id Wes e gaa d1,3, Ki s ine G. Belling1, Sø en
5
B unak1,4*
6
1No o No disk Founda ion Cen e o P o ein Resea ch, Facul y o Heal h and Medical
7
Sciences, Uni e si y o Copenhagen, DK-2200 Copenhagen, Denma k
8
2Depa men o Pulmona y and In ec ious Diseases, No dsjællands Hospi al, DK-3400
9
Hille ød, Denma k
10
3Depa men o Obs e ics and Gynaecology, Copenhagen Uni e si y Hospi al H ido e,
11
DK-2650 H ido e, Denma k
12
4Depa men o Heal h Technology, Technical Uni e si y o Denma k, DK-2800 Kongens
13
Lyngby, Denma k
14
* To whom co espondence should be add essed: [email p o ec ed]
15
,
2
Abs ac
16
Polypha macy con inues o g ow in impo ance because o mul imo bid, aging popula ions.
17
Howe e , a comp ehensi e, d ug-speci ic analysis o concomi an he apies and esponse by
18
means o dosage adjus men s has no been pe o med. In his longi udinal popula ion-wide
19
s udy o dosage adjus men s, we used elec onic heal h eco ds om 3.5 million inpa ien
20
admissions a Danish hospi als spanning he yea s 2008-2016 and iden i ied associa ions
21
be ween concomi an d ug pai s and dosages in close o 185 million ea men episodes. The
22
s udy sugges s ha dosage modi ica ions a e p oxies o d ug-d ug in e ac ions (DDIs) in 83%
23
o he cases. We ound ha d ug pai s wi h no e idence o DDIs o en had sha ed
24
me abolizing cy och ome enzyme and cellula anspo e s, and on a e age had a
25
signi ican ly highe odds a io o dosage modi ica ions han d ug pai s wi h es ablished DDIs
26
(p- alue < 0.001). Typically, hese high odds a io pai s we e p esc ibed o signi ican ly
27
ewe pa ien s and desc ibed signi ican ly less oge he in he li e a u e. This may a ionalize
28
he absence o e idence o DDIs among high odds a io pai s as hey ha e been ha de o
29
iden i y in smalle -scale s udies.
30

,
3
In oduc ion
31
In mos s udies, polypha macy has been in es iga ed using he numbe o d ugs as a bu den
32
indica o a he han he bu den e lec ing he ac ual combina ion o speci ic d ugs.
33
Polypha macy, b oadly de ined as he concomi an use o mul iple medica ions by a pa ien ,
34
makes p esc ip ion a non- i ial ask as polypha macy inc eases he isk o comp omising
35
d ug esponse h ough d ug-d ug in e ac ions (DDIs) and by ad e se d ug eac ions (ADRs)1-
36
5. DDIs may in luence he e icacy o he isk o oxici y o a d ug6, 7, while he isk o ADRs
37
has been shown o inc ease by he numbe o medica ions6. The igh d ug a he igh dose is
38
key in p ecision medicine o achie e op imal he apeu ic e ec . Al hough d ugs a e being
39
es ed epea edly in clinical ials, p e- and pos -ma ke ing, knowledge o he mechanisms
40
explaining he g ea a iabili y in pa ien s’ ea men esponse is limi ed8. Pha macogenomics
41
is an eme ging and cu ing-edge ield aiming o unde s and d ug esponse a iabili y9. The
42
polypha macy- ela ed dosage changes obse ed in eal wo ld se ings a e highly ele an in
43
his con ex as well10.
44
As li e expec ancy and p esc ip ion olumes a e inc easing, balancing pa ien sa e y among
45
he nume ous ea men op ions is becoming a majo challenge o mode n medicine11-17.
46
Mul iple, simul aneous ea men s a e po en ially ha m ul and di icul o manage and ha e
47
been associa ed wi h in-hospi al mo ali y and eadmissions independen ly o age18-20. DDIs
48
a e o en s udied in small coho s and ypically consis o s anda d e alua ion o po en ial
49
pha macokine ic inhibi o s and induc o s. F equen ly, only majo CYP450s and d ug
50
anspo e s a e es ed in ela ion o likely in e ac ing d ugs ha a e known o be p esc ibed
51
concomi an ly in selec ed a ge popula ions21-23. This is a p oblem pa icula ly in he in-
52
hospi al se ing whe e he ex en o polypha macy is highe han in he gene al popula ion24.
53
La ge-scale clinical ials o polypha macy cha ac e ising d ug esponse a e di icul o ca y
54
ou . In andomized con olled se ups, he olume o da a equi ed o ep oduce he ull
55
spec um o ea men egimens is o en no conside ed and concomi an ea men s a e
56
ypically add essed in he elde ly o in selec ed pheno ypes only25, 26.
57
As elec onic heal h eco ds (EHRs) a e b oadened in o a esea ch esou ce, s udies
58
modelling p esc ip ion pa e ns, including iden i ica ion o dose dependen ADRs, ha e been
59
conduc ed o en i e popula ions and o selec ed pa ien g oups27-35. Howe e , he ocus o
60
hese s udies has p ima ily been inapp op ia e p esc ip ion, imp o emen o pa ien
61
adhe ence, educ ion o eadmissions, ADR de ec ion, and examina ion o he po en ial o
62
,
4
dep esc ip ion36-38. Ye , e idence o he ela ionship be ween polypha macy and d ug
63
esponse by means o dosage adjus men s is spa se, and s udies modelling polypha macy a e
64
ypically limi ed o a smalle numbe o d ugs30. Su oga e measu es o polypha macy like
65
he medica ion bu den index39-41 and he como bidi y-polypha macy sco e42, 43 a e use ul, bu
66
s ill igno e many aspec s o d ug ea men con ex s39.
67
In his s udy, we cha ac e ized he complexi y o polypha macy in an in-hospi al se ing by
68
in es iga ing concomi an d ug pai s subjec o mo e equen dosage adjus men s. We
69
analysed in he o de o 185 million concomi an ea men episodes de i ed om mo e han
70
24 million d ug p esc ip ions. We aimed o p o ide insigh in o polypha macy in a eal-wo ld
71
se ing and decide whe he iden i ica ion o co-medica ion pai s subjec o dosage
72
adjus men s may imp o e he unde s anding o d ug e icacy, oxici y, and known and
73
possibly unknown DDIs. Finally, we a gue ha his s udy can complemen
74
pha macogenomics s udies o add essing esponse a iabili y in pa ien s subjec o
75
polypha macy.
76
,
5
Resul s
77
In-hospi al d ug use and polypha macy
78
In his obse a ional s udy, we used p esc ip ion and admission da a om 1,069,873
79
inpa ien s co e ing app oxima ely 50% o he Danish popula ion in an eigh -yea pe iod
80
(2008-2016). Table 1 p esen s a quan i a i e summa y o he da a. The median numbe o
81
uniquely p esc ibed d ugs pe pa ien was highe among hospi alized pa ien s compa ed o
82
he gene al popula ion38. Polypha macy was p e alen in all age g oups and he numbe o
83
di e en d ugs co ela ed posi i ely wi h age (Pea son ρ:0.40, 95% con idence in e al
84
CI:0.39-0.40, p- alue < 2.2x10-16, Supplemen a y igu e 1). The deg ee o polypha macy
85
a ied ac oss di e en admission easons (p ima y diagnoses) wi h a median o d ugs anging
86
om h ee o eigh (Supplemen a y able 1). D ugs we e classi ied acco ding o he
87
Ana omical The apeu ic Chemical (ATC) Classi ica ion Sys em and we used he i s and
88
second le els o e e o ana omical and he apeu ic d ug classes, espec i ely. Among he
89
24,379,285 inpa ien d ug p esc ip ions, 902 dis inc d ugs we e p esc ibed o a leas 50
90
pa ien s; and hese d ugs we e concomi an ly adminis e ed wi h up o 857 o he d ugs. The
91
he apeu ic classes ha we e p esc ibed o mos pa ien s we e analgesics (N02), an ibio ics
92
(J01), and an i h ombo ic agen s (B01). These we e also he ones wi h he highes numbe o
93
concomi an medica ions (ρ:0.93, 95% CI: 0.91-0.96, p- alue = 1.70x10-37) (Fig. 1).
94
Condensa ion o millions o p esc ip ions in o d ug dosage pa e ns
95
In o de o s udy d ug dosage al e a ions comp ehensi ely and p o ide a amewo k o he
96
disco e y o po en ially unknown DDIs, we cha ac e ized he 24 million d ug p esc ip ions in
97
e ms o ea men episodes, which hen comp ised he da a ounda ion o a Bayesian
98
hie a chical logis ic eg ession. Fo each pa ien , we iden i ied bo h concomi an ea men
99
episodes (whe e wo o mo e d ugs we e adminis e ed simul aneously) and mono he apy
100
episodes (whe e only one d ug was adminis e ed). Concomi an ea men episodes we e
101
s uc u ed as pai s o an index d ug and a co-medica ion con aining in o ma ion ega ding
102
d ug addi ion, p esc ibed dosage, and discon inua ion (Fig. 2). Only d ugs wi h mono he apy
103
episodes, and d ugs ha appea ed in concomi an ea men episodes o a leas 50 pa ien s,
104
we e included in he Bayesian hie a chical logis ic eg ession (Supplemen a y igu e 2).
105
F om he se o 902 unique d ugs, 413 d ugs we e included in he eg ession model ha
106
es ima ed likelihoods o dosage adjus men s du ing concomi an ea men episodes. In o al,
107
,
6
184,026,179 concomi an ea men episodes we e iden i ied among he 413 d ugs ha
108
combined in o 77,494 di e en co-medica ion pai s (i.e. an index d ug and a co-medica ion).
109
Each co-medica ion pai was hen cha ac e ized by he likelihood o dosage adjus men o
110
he index d ug using odd a ios (ORs) and he mono he apy episode o he co esponding
111
index d ug as e e ence. Fo 309 index d ugs. he dosage was mo e likely o be adjus ed wi h
112
co-medica ions han du ing he mono he apy episodes. Among he 77,494 co-medica ion
113
pai s, 3,993 co-medica ion pai s had an OR >1 (Supplemen a y able 2).
114
Among he co-medica ion pai s, 59 ATC he apeu ic subg oups we e ep esen ed and 56 had
115
a leas one d ug wi h ORs >1. All index d ugs belonging o an i h ombo ic agen s (B01, e.g.
116
dal epa in, enoxapa in, aspi in, clopidog el), d ugs ac ing on he enin-angio ensin sys em
117
(C09, e.g. amip il, losa an, enalap il, andolap il), and lipid modi ying agen s (C10, e.g.
118
a o as a in, sim as a in, osu as a in and eze imibe) we e subjec o dosage adjus men s
119
du ing concomi an ea men episodes. Ye , he e was no clea pa e n in classes gene ally
120
widely co-adminis e ed (e.g. N02, A02, J01, C03, N05) and he p opo ion o index d ugs
121
subjec o adjus men s wi h co-medica ions. Th ee d ug classes wi h low co-medica ion
122
bu den had no adjus men s wi h OR >1, i.e. asop o ec i es (C05), an ibio ics and
123
chemo he apeu ics o de ma ological use (D06), and nasal p epa a ions (R01)
124
(Supplemen a y igu e 3).
125
The complexi y o d ug dosage adjus men s in a eal-wo ld se ing
126
Nex , we cha ac e ized he co-medica ion pai s in ela ion o he he apeu ic class o he index
127
d ug and he ana omical class o he co-medica ions. While co-medica ions appea ed
128
homogeneously dis ibu ed ac oss he apeu ic classes (Fig. 3a, le ), co-medica ion pai s wi h
129
ORs >1 we e domina ed by ne ous sys em (N), ca dio ascula sys em (C), and alimen a y
130
ac and me abolism sys em (A) co-medica ions (>60%) (Fig 3a, igh ).
131
Co-medica ion pai s whe e index d ugs we e classi ied in he he apeu ic subg oups
132
psycholep ics (N05); psychoanalep ics (N06); an iepilep ics (N03); o he ne ous sys em
133
d ugs (N07); d ugs o obs uc i e ai way diseases (R03); co icos e oids, de ma ological
134
p epa a ions (D07); and an ihype ensi es (C02) had co-medica ions o he same ana omical
135
main g oup in >40% o he co-medica ion pai s wi h ORs >1. Fo he co-medica ions wi h
136
ORs >1 and an index d ug classi ied as an ieme ics and an i-nausean s (A04), he associa ed
137
co-medica ion classes we e mo e e enly dis ibu ed. Dosages o immunosupp essan (L04)
138
,
!
7!
Supplemen a y igu e 6. Compa ison be ween dosage adjus men s and DDI
e idence
!
Boxplo s indica ing he odds a io (OR) and he DDI e idence among co-medica ion pai s wi h
ORs >1 ac oss he di e en index he apeu ic d ug classes. Two-sample Mann-Whi ney U- es ,
whe e *: p- alue ≤ 0.05, **: p- alue ≤ 0.01, ***: p- alue ≤ 0.001, ****: p- alue ≤ 0.0001.
The apeu ic g oups wi h less han i e obse a ions o each DDI class (DDI/Unknown DDI) a e
no shown.
!

,
!
8!
Supplemen a y igu e 7. Compa ison be ween pa ien olume and DDI e idence
!
Boxplo s indica ing he p e alence (numbe o pa ien s) and he DDI e idence among co-
medica ion pai s wi h ORs >1 ac oss he di e en index he apeu ic d ug classes. Two-sample
Mann-Whi ney U- es , whe e *: p- alue ≤ 0.05, **: p- alue ≤ 0.01, ***: p- alue ≤ 0.001, ****:
p- alue ≤ 0.0001. The apeu ic g oups wi h less han i e obse a ions o each DDI class
(DDI/Unknown DDI) a e no shown. OR=Odds Ra io.
!
!
,
!
9!
Supplemen a y igu e 8. Compa ison be ween li e a u e and DDI e idence
!
Boxplo s indica ing he numbe o publica ions wi h co-men ioning and he DDI e idence among
co-medica ion pai s wi h ORs >1 ac oss he di e en index he apeu ic d ug classes. Two-sample
Mann-Whi ney U- es , whe e *: p- alue ≤ 0.05, **: p- alue ≤ 0.01, ***: p- alue ≤ 0.001, ****:
p- alue ≤ 0.0001. The apeu ic g oups wi h less han i e obse a ions o each DDI class
(DDI/Unknown DDI) a e no shown. OR=Odds Ra io.
!
,
!
10!
Supplemen a y igu e 9. A e age daily dose
!
Calcula ion o p esc ibed dosage as he p esc ibed a e age daily dose (ADD). A p esc ip ion
is a heal h-ca e p og am implemen ed by a physician o o he quali ied heal h ca e p ac i ione in
he o m o ins uc ions ha go e n he plan o ca e o an indi idual pa ien .The e m o en
e e s o a heal h ca e p o ide ’s w i en au ho iza ion o a pa ien o pu chase a p esc ip ion
d ug om a pha macis . P esc ip ions can be classi ied in o ou di e en g oups: (1) one- ime
p esc ip ions; (2) scheduled p esc ip ions; (3) P o Necessi a a (PN) o ‘as needed’
p esc ip ions; and (4) Va iable dosage (VAO) p esc ip ions. One- ime p esc ip ions consis on
he adminis a ion o a single dose in a single day. As his ype o p esc ip ions do no allow he
possibili y o s udy longi udinal dosage adjus men s, we did no conside hem o his s udy. PN
consis o d ugs only adminis e ed i needed and hey a e indica ed wi h a maximum numbe o
adminis a ions/day ha can be aken. An example o a PN p esc ip ion is he painkille
pa ace amol o o he analgesics indica ed o he ea men o pain i he pa ien equi es so a e
e.g. su ge y. VAO d ugs e e o d ugs whose dosage is a iable on a biochemical alue o
ano he physiological cons an o he pa ien a he ime o each adminis a ion. Fo example, his
is he case o insulin, whose dosage depends on he blood glucose le els. Fo PN and VAO d ug
p esc ip ions, he ADD was coded ca ego ically as ‘PN’ o ‘VAO’. Hence o h, quan i a i e
ADD was calcula ed only o scheduled p esc ip ion ypes. Scheduled p esc ip ions consis o
d ug indica ions ha ing an in e al o cycle wi h a de ined ime pa e n o each adminis a ion.
Among his ype, we can ind many a ia ions wi h di e en in e als and pa e ns. The igu es a
and b exempli y wo di e en examples: a Dosing egimen o acyclo i (J05AB01), a d ug used
o ea in ec ions caused by ce ain ypes o i uses (e.g. cold so es, shingles, chickenpox). The
p esc ip ion egimen is indica ed wi h a daily in e al (1 day), a equency o h ee
adminis a ions ollowing he ime pa e n a 8 am, 2 pm and 10 pm. The p esc ibed dose
,
!
11!
indica ed a each adminis a ion is o 200 mg. The inal calcula ed ADD o acyclo i is o 600
mg (see Equa ion 1). b A mo e complex egimen is p esen ed wi h he medica ion
Le o hy oxine sodium (H03AA01), a hy oid ho mone medica ion ha is used o ea
hypo hy oidism and o he ho monal condi ions. The ea men plan o his d ug ollows a
weekly in e al (7 days) and di e en p esc ibed doses o e e y o he day. The p esc ibed dose
is 50 mic og ams (mcg) o days 1,3,5 and 100 mcg o days 0,2,4,6. The inal calcula ed ADD
is 78.57 mcg, which is he sum o he wo p e ious p esc ibed doses in he same weekly in e al.
!
,
!
12!
Supplemen a y igu e 10. T ea men episodes
!
Cons uc ion o concomi an and mono he apy ea men episodes. Pa ien s can ha e a d ug
p esc ibed mul iple imes du ing an admission. We cons uc ed concomi an ea men episodes
as he in e als o ime when wo di e en d ug p esc ip ions we e con empo aneously ac i e.
Mono he apy ea men episodes we e used as he e e ence ea men episodes as he in e als o
ime when a d ug was no concomi an ly gi en wi h any o he d ug.
!

,
!
13!
Supplemen a y igu e 11. Li e a u e co-men ions diag am
!
!
Diag am o he ull- ex a icle and abs ac co pus gene a ion o he ex mining o d ug-d ug
in e ac ions. DTU DTM Co pus (~15 million a icles), PubMed (~27 million a icles) and
MEDLINE we e used o he collec ion o ull- ex a icles and abs ac s.
!
,
!
14!
Tables
Supplemen a y able 1: Admission p ima y diagnosis
ICD-10
chap e
Chap e name
No. unique d ugs, median
(IQR)
1
Ce ain in ec ious and pa asi ic diseases
8 (4-14)
2
Neoplasms
8 (5-12)
3
Diseases o he blood and blood o ming o gans and ce ain
diso de s in ol ing he immune mechanism
7 (4-11)
4
Endoc ine, nu i ional and me abolic diseases
8 (4-12)
5
Men al and beha iou al diso de s
6 (4-9)
6
Diseases o he ne ous sys em
5 (3-9)
7
Diseases o he eye and adnexa
4 (2-7)
8
Diseases o he ea and mas oid p ocess
3 (2-5)
9
Diseases o he ci cula o y sys em
8 (4-12)
10
Diseases o he espi a o y sys em
8 (4-13)
11
Diseases o he diges i e sys em
6 (4-10)
12
Diseases o he skin and subcu aneous issue
4 (2-9)
13
Diseases o he musculoskele al sys em and connec i e
issue
8 (4-11)
14
Diseases o he geni ou ina y sys em
6 (3-10)
15
P egnancy, childbi h and he pue pe ium
3 (2-5)
16
Ce ain condi ions o igina ing in he pe ina al pe iod
2 (1-5)
17
Congeni al mal o ma ions, de o ma ions and ch omosomal
abno mali ies
4 (2-7)
18
Symp oms, signs and abno mal clinical and labo a o y
indings, no elsewhe e classi ied
5 (2-9)
19
Inju y, poisoning and ce ain o he consequences o ex e nal
causes
7 (3-12)
20
Ex e nal causes o mo bidi y and mo ali y
5 (3-9)
21
Fac o s in luencing hea h s a us and con ac wi h heal h
se ices
5 (1-9)
!
,
!
15!
Supplemen a y able 2: Co-medica ion pai s wi h associa ed dosage adjus men
Lis o co-medica ion pai s wi h ORs >1 o dosage adjus men . OR=Odds Ra io.
Supplemen a y able 3: DDI e idence
Lis o co-medica ion pai s wi h ORs >1 o dosage adjus men and hei DDI e idence.
OR=Odds Ra io.
Supplemen a y able 4: Sha ed CYPs and anspo e s
Lis o co-medica ion pai s wi h ORs >1 o dosage adjus men and hei pha macokine ic
ac i i ies: sha ed me abolism and anspo , and associa ed a ian s. OR=Odds Ra io.
Supplemen a y able 5: Li e a u e co-men ions
Lis o co-medica ion pai s wi h ORs >1 o dosage adjus men , DDI e idence and co-
men ioning in he li e a u e. OR=Odds Ra io.
Supplemen a y able 6: One- ime d ugs
Lis o excluded index d ugs o which he p esc ip ion ype was ‘one- ime’ o ‘single-
adminis a ion’ in mo e han 70% o he cases.
D|Appendix
253
,
Supplemen a y igu e 4 | Longi udinal ajec o ies o med by he same ATC g oup. a,
ajec o ies o med by d ugs om he same ana omical ATC g oup we e included and hen
ho izon ally sepa a ed by a e age ime be ween p esc ip ion edemp ions. Each g oup o
ajec o ies con ains edges ha sepa a e he di e en p esc ip ion edemp ions, whose heigh
ep esen he numbe o ajec o ies ha go om leng h 2 o leng h 3, o leng h N. Each g oup
o ajec o ies is o de ed om sho es o longes ( op o bo om). *Re e o Figu e 1 o ATC
g oup legend. b, Numbe o indi iduals edeeming he subg oup indica ed on he le side
(node size) and how many o hese pa ien s edeem ano he d ug used in hype ension (C02,
C03, C07, C08, C09) pos e io ly (o de ed by chemical subg oup in ATC).

,
Supplemen a y igu e 5 | Indi idual g ouping o Poisson modelling change o e ime.
Each g oup o indi iduals, sepa a ed by sex and yea o bi h a e dynamically alloca ed o
di e en g oups a each ime poin , depending on hei p esc ip ion edemp ion a he ime.
Tha p ocess is epea ed o each s a um o he da a (i.e., o each sex, yea o bi h and ime
o p esc ip ion), o each d ug pai . As an example, women bo n in 1991 a e used in he
igu e. They en e he s udy in 1995 ( 1) o p esc ip ion edemp ion pai (P1P2). Some o
hese indi iduals edeemed bo h p esc ip ions a 1, so hey a e di ec ly coun ed in P12 (P12 =
has edeemed P1 and P2; P10 = has edeemed P1 bu no P2; P02 = has edeemed P2 bu no
P1; P00 = has edeemed nei he P1, no P2). This indi idual is hen censo ed om he es o
he yea s in he model. O he indi iduals in he same s a um a 1 will no ha e edeemed any
o he d ugs in he pai s unde s udy, P1 o P2, hence hey will be coun ed as P00. A each
ime, he posi ion o he di e en indi iduals migh change, as hey migh ha e edeemed one
o bo h p esc ip ions, o hey migh ha e le he s udy (emig a ion o dea h).
,
Supplemen a y igu e 6 | Pe iod o inclusion, 1995-2019, o Cox p opo ional haza d
eg ession model coho . Indi iduals who edeemed he p esc ip ion be o e 1995 a e
excluded om he coho (le - unca ed) and pa ien s who lea e he s udy, ei he due o end
o window o s udy (2019), emig a ion o o he a e censo ed ( igh censo ed). The ime ame
is he pe iod o which p esc ip ion egis y has in o ma ion (1995-2019), lea ing hose
pa ien s who edeemed he p esc ip ion be o e he beginning o he egis y unca ed ou o
he model (Supplemen a y Fig. 2). Cha lson Como bidi y Index (CCI)
20
was calcula ed using
he ICD-10 coded disease da a om he Danish Na ional Pa ien Regis y. Age, sex and CCI
we e added as co a ia es in he model and an in e ac ion wi h hem was included i he
haza d p opo ionali y assump ion was iola ed – using a es based on Schoen eld esiduals
wi h a p- alue lowe han 0.0001.
,
Supplemen a y able 1 | Coho cha ac e is ics.
Male Female
Numbe o pa ien s (%) 3 586 032 (49 %) 3 669 887 (51%)
Median age 1995, yea s 30 30
Mean age 1995, yea s 31·2 32·9
Range age 1995, yea s 0-106 0-110
Median age 2019, yea s 44 47
Mean age 2019, yea s 43·9 47·1
Range age 2019, yea s 0-110 0-110
P esc ip ion edemp ion
Median 11 17
Mean 13·9 19·1
SD 10·7 13·4
Range 1-164 1-129
,
Supplemen a y able 2 | P esc ip ion dis ibu ion o e he Ana omical The apeu ic
Chemical Classi ica ion Sys em.
Male Female
Age 0-14 15-44 45-64 65-84 >85 0-14 15-44 45-64 65-84 >85
ATC
mean
(SD)
A 1·3 (0·6) 1·7 (1·1) 2·1 (1·7) 2·5 (1·9) 2·3 (1·6) 1·3 (0·6) 2·0 (1·4) 2·3 (1·8) 2·9 (2·1) 2·7 (1·7)
B 1·0 (0·3) 1·1 (0·4) 1·2 (0·5) 1·4 (0·7) 1·4 (0·6) 1·0 (0·2) 1·2 (0·5) 1·2 (0·5) 1·4 (0·7) 1·4 (0·6)
C 1·3 (0·7) 1·5 (1·1) 2·8 (2·1) 3·2 (2·2) 2·2 (1·5) 1·3 (0·7) 1·5 (1·0) 2·6 (1·9) 3·2 (2·3) 2·4 (1·6)
D 1·9 (1·3) 2·3 (1·7) 2·2 (1·6) 2·2 (1·6) 1·9 (1·3) 2·0 (1·3) 2·7 (2·0) 2·3 (1·6) 2·3 (1·6) 1·9 (1·3)
G 1·0 (0·2) 1·1 (0·3) 1·2 (0·5) 1·4 (0·6) 1·2 (0·5) 1·1 (0·3) 2·1 (1·3) 1·7 (1·0) 1·4 (0·7) 1·2 (0·5)
H 1·0 (0·1) 1·0 (0·2) 1·0 (0·2) 1·1 (0·3) 1·0 (0·2) 1·0 (0·2) 1·1 (0·4) 1·1 (0·4) 1·1 (0·4) 1·0 (0·3)
J 1·0 (1·0) 2·2 (1·3) 2·1 (1·4) 2·5 (1·6) 2·2 (1·4) 1·9 (0·2) 3·3 (1·9) 2·6 (1·7) 2·7 (1·8) 2·4 (1·6)
L 1·1 (0·3) 1·1 (0·3) 1·1 (0·3) 1·1 (0·3) 1·0 (1·8) 1·1 (0·2) 1·0 (0·2) 1·1 (0·3) 1·0 (0·2) 1·0 (0·2)
M 1·1 (0·3) 1·5 (0·7) 1·7 (0·9) 1·7 (0·9) 1·4 (0·7) 1·1 (0·3) 1·6 (0·9) 1·8 (1·1) 1·9 (1·2) 1·5 (0·8)
N 1·3 (0·8) 2·5 (2·3) 2·7 (2·3) 3·0 (2·3) 2·8 (1·9) 1·3 (0·8) 2·7 (2·4) 3·0 (2·5) 3·4 (2·5) 3·1 (2·1)
P 1·0 (0·2) 1·2 (0·5) 1·1 (0·4) 1·1 (0·3) 1·0 (0·2) 1·0 (0·2) 1·3 (0·5) 1·2 (0·5) 1·1 (0·3) 1·0 (0·2)
R 2·0 (1·3) 2·0 (1·3) 1·9 (1·3) 2·2 (1·7) 1·7 (1·1) 1·8 (1·2) 2·4 (1·7) 2·2 (1·7) 2·3 (1·8) 1·7 (1·5)
S 1·5 (0·8) 1·6 (1·0) 1·7 (1·1) 2·1 (1·5) 1·7 (1·1) 1·4 (0·8) 1·7 (1·0) 1·8 (1·3) 2·2 (1·6) 1·9 (1·3)
V 1·0 (0·1) 1·0 (0·1) 1·0 (0·1) 1·0 (0·1) 1·0 (0·2) 1·0 (0·1) 1·0 (0·1) 1·0 (0·1) 1·0 (0·1) 1·0 (0·1)
Range
A 1-13 1-23 1-21 1-21 1-14 1-18 1-23 1-22 1-21 1-18
B 1-6 1-7 1-7 1-7 1-6 1-6 1-8 1-8 1-8 1-6
C 1-9 1-19 1-20 1-21 1-15 1-9 1-18 1-20 1-21 1-15
D 1-16 1-21 1-21 1-19 1-15 1-18 1-21 1-19 1-18 1-15
G 1-4 1-7 1-11 1-7 1-5 1-5 1-14 1-11 1-9 1-7
H 1-4 1-5 1-5 1-5 1-4 1-5 1-7 1-6 1-5 1-4
J 1-17 1-20 1-16 1-14 1-11 1-17 1-21 1-16 1-15 1-13
L 1-4 1-5 1-6 1-5 1-3 1-5 1-5 1-5 1-5 1-4
M 1-6 1-9 1-11 1-10 1-7 1-6 1-10 1-11 1-11 1-9
N 1-19 1-33 1-34 1-27 1-18 1-14 1-36 1-30 1-28 1-19
P 1-5 1-6 1-7 1-5 1-3 1-5 1-7 1-6 1-5 1-4
R 1-14 1-17 1-17 1-17 1-13 1-14 1-19 1-19 1-18 1-13
S 1-13 1-16 1-16 1-17 1-13 1-13 1-17 1-16 1-15 1-14
V 1-2 1-3 1-3 1-3 1-2 1-2 1-3 1-3 1-2 1-2
,
Supplemen a y able 3 | Rela i e isk o pai s in igu e 2
ATC 1 ATC 12
Numbe
pa ien s RR 95% CI P- alue
N02AA N02AB 138 110 2.76 2.74 – 4.13 2.70e-08
R03CC R03BA 226 898 2.72 1.86 - 3.99 2.68e-07
C09AA C09BA 164 680 2.79 1.93 - 4.03 5.22e-08
C07AB B01AA 125 137 3.17 2.17 - 4.63 2.72e-09
C07AB C01AA 103 043 3.55 2.31 - 5.47 9.07e-09
A03FA N02AB 102 783 2.77 1.89 - 4.05 1.42e-07
N06AB N05AF 116 605 2.86 2.02 - 4.0 4.11e-09
B01AC C01AA 135 262 2.88 1.89 - 4.38 8.06e-07
N05BA N02AG 118 223 2.83 2.09 - 3.83 1.83e-11
C09AA C01DA 137 306 3.14 2.1 - 4.62 8.19e-09
J01CE D07BB 195 989 2.81 2.19 - 3.61 5.51e-16
M01AB N02AG 118 272 2.80 2.0 - 3.76 5.61e-12
M01AC M01AB 121 302 2.73 2.08 - 3.58 3.70e-13
G03AA J01EB 391 118 3.05 1.93 - 4.82 1.87e-06
J01EB G01AF 189 887 3.28 2.33 - 4.63 1.19e-11
A02BA N05BA 138 066 2.89 2.18 - 3.83 1.16e-13
M01AX C09AA 105 249 2.88 2.06 - 4.04 8.30e-10
C10AA C03CA 218 656 2.74 1.94 - 3.87 1.17e-08
G03AA G01AF 279 484 3.98 2.42 - 6.55 5.27e-08
A02BA A03FA 149 396 2.74 2.06 - 3.64 4.28e-12
D07BB J01FA 131 233 3.13 2.42 - 4.04 1.93e-18
C10AA A12BA 232 191 2.85 1.98 - 4.10 1.74e-08
S01GA R06AE 113 185 2.74 2.07 - 3.61 1.27e-12
G03AA D06BB 129 086 3.12 2.00 - 4.86 5.17e-07
M01AX B01AC 113 398 2.98 2.11 - 4.21 5.70e-10
S01GA S01GX 125 768 2.75 2.06 - 3.68 1.01e-11
C07AB C03DA 121 672 2.90 1.99 - 4.23 3.44e-08
D07BB D07AC 112 418 3.43 2.59 - 4.54 7.00e-18
A08AA M01AB 185 846 2.97 2.28 - 3.87 7.04e-16
A02BA R05FA 126 611 2.87 2.16 - 3.81 3.85e-13
G03AA J02AC 414 591 2.82 1.76 - 4.53 1.66e-05
D07BB S01AA 123 584 2.78 2.14 - 3.62 2.15e-14
M01AX C08CA 111 180 2.95 2.07 - 4.20 2.48e-09
D07BB D07AB 111 589 3.33 2.50 - 4.45 3.02e-16
D07BB D01AC 120 234 3.26 2.47 - 4.29 4.61e-17
M01AX C03CA 102 801 3.10 2.17 - 4.42 4.24e-10
C03AB R05CB 102 408 2.78 1.92 - 4.03 5.53e-08
D07BB A10BK 3 349 0.002 1.20e-03 - 3.15e-03 4.44e-142
D07CB A10BK 1 538 2.27e-03 1.24e-03 - 4.17e-03 2.46e-86
M03BA A10BK 1 264 2.32e-03 1.24e-03 - 4.35e-03 3.30e-80
S03BA B01AF 1 505 3.24e-03 1.80e-03 - 5.82e-03 8.93e-82
S03BA N05CH 1 789 3.25e-03 2.04e-03 - 5.19e-03 1.03e-127
G03CB B01AF 6 461 3.39e-03 1.74e-03 - 6.62e-03 2.50e-62
S03BA N05CH 1 789 3.25e-03 2.04e-03 - 5.19e-03 1.03e-127
G03CB B01AF 6 461 3.39e-03 1.74e-03 - 6.62e-03 2.50e-62
G03CB N05CH 6 023 3.71e-03 1.91e-03 - 7.18e-03 9.73e-62
C03CB A10BH 449 6.49e-01 2.74e-01 - 1.53 3.24e-01
G03CB R02AX 1 121 4.01e-03 2.02e-03 - 7.94e-03 2.24e-56
B01AF N06DX 1 344 4.15e-01 3.74e-01 - 4.60e-01 1.38e-62
A11EA G03FA 1 009 1.06e+01 8.51 – 13.1 6.12e-102
G01AG S01CA 1 696 1.06e+01 7.10 - 15.7 3.04e-31
B03AE A06AD 1 503 1.06e+01 6.51 – 17.1 1.40e-21

,
S01KA M01AH 1 012 1.08e+01 7.57 – 15.4 1.49e-39
A11EA G04CA 1 052 1.09e+01 8.58 – 14.0 3.49e-82
A11EA G03CB 1 030 1.10e+01 9.47 - 12.8 8.01e-217
S01KA A02AA 1 081 1.12e+01 7.81 – 16.0 7.48e-40
S01KA A06AD 1 745 1.15e+01 7.85 – 16.7 1.40e-36
A10BB N02BE 97 371 1.60e+03 1.52e+03 - 1.68e+03 2.22e-308
A10BF N02BE 3 717 1.66e+03 1.61e+03 - 1.72e+03 2.22e-308
A10BB N02AX 68 483 1.67e+03 1.61e+03 - 1.75e+03 2.22e-308
A10BB N03AF 4 148 1.70e+03 1.65e+03 - 1.75e+03 2.22e-308
A10BB N02AJ 22 433 1.72e+03 1.67e+03 - 1.77e+03 2.22e-308
A10BB N03AA 1 403 1.74e+03 1.68e+03 - 1.80e+03 2.22e-308
A10BB N02AA 52 922 1.74e+03 1.68e+03 - 1.81e+03 2.22e-308
A10BB N03AG 2 891 1.75e+03 1.69e+03 - 1.80e+03 2.22e-308
A10BB N02AG 14 877 1.76e+03 1.67e+03 - 1.85e+03 2.22e-308
A10BB N02AE 12 015 1.79e+03 1.72e+03 - 1.87e+03 2.22e-308
A10BB N02AB 17 553 1.79e+03 1.71e+03 - 1.88e+03 2.22e-308
A10BF N03AX 1 178 1.86e+03 1.82e+03 - 1.90e+03 2.22e-308
A10BH A10BK 13 699 4.05e-01 3.24e-01 - 5.06e-01 2.34e-15
A10BD A10BK 12 727 4.06e-01 3.15e-01 - 5.24e-01 3.95e-12
A10BJ A10BK 13 156 4.19e-01 3.35e-01 - 5.23e-01 1.93e-14
A10BH A10BJ 13 537 4.55e-01 3.73e-01 - 5.56e-01 1.34e-14
N06DX N05AD 2 607 4.72e-01 3.90e-01 - 5.70e-01 8.22e-15
A10BH A10BD 8 676 4.80e-01 3.94e-01 - 5.85e-01 4.16e-13
A10BD A10BJ 11 035 5.03e-01 4.03e-01 - 6.28e-01 1.24e-09
B01AE B01AF 11 838 5.04e-01 4.23e-01 - 6.00e-01 1.41e-14
A10BG A10BH 1 436 5.25e-01 4.01e-01 - 68.6 2.31e-06
N06DX N02AB 4 151 5.30e-01 4.48e-01 - 62.7 1.10e-13
A11EA A01AB 2 882 9.20 6.88 - 12.3 1.93e-50
A11EA A11CA 1 065 9.44 6.95 - 12.8 9.89e-47
G01AG G03DA 1 351 9.57 6.64 – 13.8 8.72e-34
A11EA A10BA 1 331 9.73 6.71 - 14.1 2.88e-33
S01KA S01BC 1 116 9.89 6.56 – 14.91 6.24e-28
S01KA S01XA 1 175 10.28 7.71 - 13.7 9.65e-57
A11EA A12AX 1 857 10.5 7.97 – 13.85 2.62e-62
G01AG G02BA 2 112 10.70 6.62 – 17.5 8.48e-22
S01KA S01CA 2 092 10.8 7.11 – 16.4 8.13e-29
B03AE B03BA 1 103 9.87 6.78 – 14.37 6.95e-33
,
Supplemen a y able 4 | S a i ica ion o pa ien s by p esc ip ion ajec o ies
ACE ea men wi h
no change
ARB ea men
wi h no change
ACE ea men
wi h change o
ARB
ARB ea men
wi h change o
ACE
Numbe o pa ien s 549 436 (49.3) 287 488 (25.8) 229 216 (20.6) 48 054 (4.3)
Age, yea s
Mean (SD) 63.3 (13·9) 61.7 (13·3) 60.7 (12·3) 61.0 (13·2)
Median (IQR) 63.7 (20.0) 62.0 (18.9) 61.0 (17.4) 61.2 (18.9)
Sex
Male 257 099 (46.8) 153 527 (53.4) 122 549 (53.5) 24 484 (51.0)
Female 292 337 (53.2) 133 961 (46.6) 106 667 (46.5) 23 570 (49.0)
Cha lson como bidi ies
Myoca dial in a c ion 74 244 (13.5) 14 962 (5.2) 23 253 (10.1) 5 485 (11.4)
Conges i e hea ailu e 90 247(16.4) 13 696 (4.8) 26 115 (11.4) 6 668 (13.9)
Pe iphe al ascula diseases 59 155 (10.7) 18 643 (6.5) 20 914 (9.1) 5 666 (11.8)
Ce eb o ascula disease 102 441 (18.6) 37 758 (13.1) 35 109 (15.3) 9 894 (20.6)
Demen ia 29 033 (4.9) 8 387 (2.9) 6 181 (2.7) 2 452 (5.1)
Ch onic obs uc i e pulmona y disease 77 099 (14.0) 31 317 (10.9) 28 892 (12.6) 7 029 (14.6)
Rheuma oid disease 21 897 (3.9) 10 172 (3·5) 9 500 (4.1) 2 135 (4.4)
Pep ic ulce disease 36 803 (6.7) 13 572 (4.7) 12 545 (5.4) 3 315 (6.9)
Mild li e disease 12 827 (2.3) 5 527 (1.9) 4 739 (2.0) 1 175 (2.4)
Diabe es wi hou complica ions 90 186 (16.4) 27 143 (9.4) 37 033 (16.1) 8 391 (17·4)
Diabe es wi h complica ions 31 608 (5·7) 7 560 (2.6) 12 894 (5.6) 2 806 (5.8)
Hemiplegia o pa aplegia 3 633 (0.7) 1 433 (0.5) 1 064 (0·4) 298 (0.6)
Renal disease 29 369 (5.3) 8 112 (2.8) 11 624 (5.1) 3 236 (6.7)
Cance 102 249 (18.6) 44 873 (15.6) 39 402 (17.1) 9 362 (19.5)
Mode a e o se e e li e disease 4 272 (0·7) 1 438 (0.5) 1 174 (0·5) 341 (0.7)
Me as a ic solid umou 20 741 (3.7) 8 219 (2.8) 6 841 (2·9) 1 774 (3.7)
AIDS/HIV 295 (0·05) 109 (0.04) 85 (0·03) 23 (0·04)
Da a a e n (%) unless o he wise speci ied
,
Supplemen a y able 5 | Popula ion con inen o o igin
Pe cen age
Eu ope 92.50
Asia 2.93
Middle Eas 1.94
A ica 1.24
No h Ame ica 0.81
Sou h and Cen al Ame ica 0.39
Oceania 0.16
No s a ed 0.03
E|Appendix
269