1
Scien i ic RepoR s | 7:44702 | DOI: 10.1038/s ep44702
www.na u e.com/scien i ic epo s
Ex e nal alida ion o a COPD
p edic ion model using popula ion-
based p ima y ca e da a: a nes ed
case-con ol s udy
B igh I Nwa u1,2, Colin R Simpson1, Aziz Sheikh1,3 & Daniel Ko z1,3,4
Eme ging models o p edic ing isk o ch onic obs uc i e pulmona y disease (COPD) equi e ex e nal
alida ion in o de o assess hei clinical alue. We alida ed a p e ious model o p edic ing new
onse COPD in a di e en da abase. We andomly d ew 38,597 case-con ol pai s ( o al N = 77,194) o
indi iduals aged ≥35 yea s and ma ched o sex, age, and gene al p ac ice om he Uni ed Kingdom
Clinical P ac ice Resea ch Da alink da abase. We assessed accu acy o he model o disc imina e
be ween COPD cases and non-cases by calcula ing a ea unde he ecei e ope a o cha ac e is ic
(ROCAUC) o he p edic ion sco es. Analogous o he de elopmen model, e e smoking (OR 6.70; 95%CI
6.41–6.99), p io as hma (OR 6.43; 95%CI 5.85–7.07), and highe socioeconomic dep i a ion (OR 2.90;
95%CI 2.72–3.09 o highes s. lowes quin ile) inc eased he isk o COPD. The alida ed p edic ion
sco es anged om 0–5.71 (ROCAUC 0.66; 95%CI 0.65–0.66) o males and 0–5.95 (ROCAUC 0.71; 95%CI
0.70–0.71) o emales. We ha e con i med ha smoking, p io as hma, and socioeconomic dep i a ion
a e key isk ac o s o new onse COPD. Ou model seems ex e nally alid a iden i ying pa ien s a
isk o de eloping COPD. An impac assessmen now needs o be unde aken o assess whe he his
p edic ion model can be applied in clinical ca e se ings.
The bu den o ch onic obs uc i e pulmona y disease (COPD) has been ising, now ep esen ing one o he lead-
ing causes o mo bidi y and mo ali y wo ldwide1. Whils es ima es o disease bu den ha e p ima ily come om
de eloped coun ies, he p e alence appea s o be ising in de eloping coun ies as well, and he esul an mo -
ali y is p ojec ed o ise by 30% in he nex decade2. In he Uni ed S a es (US), o e 12 million adul s ha e COPD,
ep esen ing he hi d leading cause o dea h3; i is he second mos common cause o eme gency hospi alisa ion
in he Uni ed Kingdom (UK)4–6.
Whils some s udies ha e e alua ed algo i hms o iden i y indi iduals wi h es ablished COPD7–9, he a e o
undiagnosed disease emains high10, occu ing in one ou o eigh people o e he age o 35 yea s6. The e is a
pauci y o s udies ha ha e de eloped hands-on ools ha enable ea ly iden i ica ion o indi iduals9 a - isk o
u u e COPD. Ou abili y o cons uc models ha will enhance iden i ying indi iduals a isk well be o e dis-
ease onse will p o ide he oppo uni y o de eloping key s a egies o p e en ion11. Using he P ima y Ca e
Clinical In o ma ics Uni (PCCIU) gene al p ac ice (GP) da abase, we de eloped and in e nally alida ed he i s
isk p edic ion model o ea ly de ec ion o inciden COPD, which simul aneously ook in o accoun a ange o
known isk ac o s, including smoking, age, sex, p io as hma, and socio-economic s a us11. A mo e ecen s udy
u ilised he UK Clinical P ac ice Resea ch Da alink (CPRD) da abase o simila ly de elop and alida e a COPD
p edic ion model, de i ing compa able p edic i e alues as ou p e ious s udy12. Howe e , i is impo an , p io
o using hese isk sco ing sys ems in clinical p ac ice ha hey a e ex e nally alida ed in en i ely di e en da a-
se s o compa able popula ions.
1As hma UK Cen e o Applied Resea ch, Cen e o Medical In o ma ics, Ushe Ins i u e o Popula ion Heal h
Sciences, The Uni e si y o Edinbu gh, UK. 2School o Heal h Sciences, Uni e si y o Tampe e, Finland. 3Depa men
o Family Medicine, CAPHRI School o Public Heal h and P ima y Ca e, Maas ich Uni e si y Medical Cen e,
Maas ich , The Ne he lands. 4Ins i u e o Gene al P ac ice, Addic ion Resea ch and Clinical Epidemiology Uni ,
Medical Facul y o he Hein ich-Heine-Uni e si y Düsseldo , Düsseldo , Ge many. Co espondence and eques s
o ma e ials should be add essed o D.K. (email: daniel.k[email p o ec ed].de)
Recei ed: 03 Oc obe 2016
Accep ed: 13 Feb ua y 2017
Published: 17 Ma ch 2017
OPEN
www.na u e.com/scien i ic epo s/
2
Scien i ic RepoR s | 7:44702 | DOI: 10.1038/s ep44702
In he cu en s udy, we he e o e aimed o ex e nally alida e ou p e iously de eloped p edic ion model
(de eloped using he PCCUI GP da abase) using a di e en da abase (i.e. he CPRD GP da abase). The cu en
wo k is he i s o ex e nally alida e a COPD p edic ion algo i hm o ea ly de ec ion o indi iduals a - isk o
u u e COPD in an en i ely di e en da abase bu compa able popula ion.
Resul s
Backg ound cha ac e is ics o he s udy popula ion. O e all, majo i y o he pa ien s had smoked
a some poin , bu he p opo ion o smoke s was highe in COPD cases (n = 33,269; 86%) han in con ols
(n = 19,908; 52%), ega dless o whe he he CPRD smoking codes o he PCCIU codes we e used (Table1). The
p opo ion o hose wi h p io as hma was up o n = 13,161, 17% (cases n = 10,210, 27%; con ols n = 2,951, 7%).
A highe p opo ion o cases was mo e dep i ed han con ols as measu ed by he Index o Mul iple Dep i a ion
(IMD) quin ile (Table1).
Associa ions be ween isk ac o s and COPD. In unadjus ed and adjus ed (i.e. simul aneous adjus -
men o all ac o s) models, e e smoking was associa ed wi h a se en- old inc eased odds o COPD (Table2).
P io as hma was associa ed wi h an inc eased isk o COPD: adjus ed OR o p io as hma was 5.04 (95% CI
4.77–5.33). Compa ed o he leas dep i ed IMD quin ile, hose in he mo e dep i ed IMD quin iles we e inc eas-
ingly a highe isk o COPD (Table2). The es ima es o smoking and p io as hma we e simila in magni ude
and di ec ion when ei he o he codes om CRPD and PCCIU da a we e used in he analyses (da a no shown).
Valida ion o COPD p ognos ic index. Table3 p esen s he alida ed decile p ognos ic sco es when he
sco es de i ed om he de elopmen models based on he PCCIU da a we e applied o he CPRD da a. The sco es
we e de i ed o males and emales sepa a ely and anged om 0 (lowes ) o 5.71 (highes ) o males and 0 o
5.95 o emales. Some deciles did no ha e co esponding p ognos ic sco es, an indica ion ha he sco es we e
no no mally dis ibu ed. The accu acy o he alida ed p edic ion model in disc imina ing be ween COPD and
non-COPD pa ien s was ROCAUC = 0.66 (95% CI 0.65–0.66) o males and ROCAUC = 0.71 (95% CI 0.70–0.71) o
Cha ac e is ic
All N = 77,002
n (%)
COPD Cases
n = 38,511 n (%)
Con ols
n = 38,491 n (%)
Smoking s a us
Ne e smoke 23,825 (30.9) 5,242 (13.6) 18,583 (48.3)
E e smoke 53,177 (69.1) 33,269 (86.4) 19,908 (51.7)
P e ious as hma p io o diagnosis o COPD
No 63,841 (82.9) 28,301 (73.5) 35,540 (92.3)
Yes 13,161 (17.1) 10,210 (26.5) 2,951 (7.7)
IMD Quin iles
1s Quin ile (leas dep i ed) 13,470 (17.5) 5,801 (15.1) 7,669 (19.9)
2nd Quin ile 16,788 (21.8) 7,814 (20.3) 8,974 (23.3)
3 d Quin ile 15,223 (19.8) 7,446 (19.3) 7,777 (20.2)
4 h Quin ile 16,291 (21.1) 8,794 (22.8) 7,497 (19.5)
5 h Quin ile (mos dep i ed) 15,230 (19.8) 8,656 (22.5) 6,574 (17.1)
Table 1. Pa icipan s’ baseline cha ac e is ics by COPD cases and con ols. COPD = ch onic obs uc i e
pulmona y disease. CPRD = Clinical P ac ice Resea ch Da alink gene al p ac ice da abase.
Unadjus ed OR (95% CI)
N = 77,002
Adjus ed1 OR (95% CI)
N = 77,002
Smoking s a us
Ne e smoke 1 1
E e smoke 7.01 (6.70–7.33) 7.21 (5.88–7.57)
P e ious as hma p io o diagnosis o COPD
No 1 1
Yes 4.52 (4.31–4.74) 5.04 (4.77–5.33)
Index o Mul iple Dep i a ion Quin iles
1s Quin ile (leas dep i ed) 1 1
2nd Quin ile 1.34 (1.27–1.41) 1.23 (1.16–1.31)
3 d Quin ile 1.64 (1.55–1.73) 1.47 (1.37–1.57)
4 h Quin ile 2.20 (2.07–2.32) 1.81 (1.68–1.94)
5 h Quin ile (mos dep i ed) 2.90 (2.72–3.09) 2.17 (2.00–2.34)
Table 2. Unadjus ed and adjus ed associa ions be ween isk ac o s and diagnosis o COPD: Odds a io
(OR), 95% con idence in e al (95% CI). COPD = ch onic obs uc i e pulmona y disease. CPRD = Clinical
P ac ice Resea ch Da alink gene al p ac ice da abase. IMD = Index o Mul iple Dep i a ion.
www.na u e.com/scien i ic epo s/
3
Scien i ic RepoR s | 7:44702 | DOI: 10.1038/s ep44702
emales (Table3). The ROC cu es o he alida ed sco es a e shown in Fig.1 and he sensi i i y and speci ici y
alues o he a ious cu poin s on he p ognos ic sco es a e shown in Supplemen a y File 4.
COPD p ognos ic sco es de i ed om he CPRD da a. Table4 p esen s he p ognos ic sco es de i ed
solely om he CPRD da a. The sco es anged om 0 o 4.37. The co esponding ROCAUC o assessing he accu-
acy o he model was 0.74 (95% CI 0.73–0.74). The ROC cu es a e shown in Figu eS1 o Supplemen a y File 5.
The p ognos ic sco es we e simila when ei he o he smoking and p io as hma codes om CRPD and PCCIU
da a we e used in he analyses (da a no shown).
Discussion
In his i s e e ex e nal alida ion exe cise o a COPD p edic ion model using he CPRD da abase, we ha e
ound simila p edic ion es ima es as hose de i ed om he PCCIU da abase de i ed based model, bo h in he
magni ude and di ec ion o e ec . Whils he p ognos ic sco es based on he de elopmen model anged om
0 o 7.50, simila ly o males and emales, he sco es om he cu en alida ion s udy anged om 0 o 5.71 o
males and 0 o 5.95 o emales. Simila ly, whils we obse ed some small di e ences in he associa ions be ween
smoking, p io as hma, and dep i a ion quin iles and he isk o COPD, hese we e la gely in he same di ec ion
o impac as expec ed. As expec ed, he accu acy o he models as measu ed by he ROCAUC we e somewha lowe
Range o alues o deciles o PI sco es A ea unde ROC
cu e (95% CI)1 (lowes ) 2345678910 (highes )
P ognos ic sco es
Males 0.00–0.00 0.61–1.41 1.42–1.91 — 1.92–2.52 — 2.53–3.31 — 3.32–4.49 4.50–5.71 0.66 (0.65– 0.66)
Females 0.00–0.00 0.01–0.45 0.46–1.50 1.51–2.26 — 2.27–2.71 2.72–3.73 3.74–3.76 3.77–4.93 4.94–5.95 0.71 (0.70–0.71)
Table 3. Valida ion o COPD p ognos ic sco es de i ed om he PCCIU da a using he CPRD da a.
COPD = ch onic obs uc i e pulmona y disease. PCCIU = P ima y Ca e Clinical In o ma ics Uni gene al
p ac ice da abase. CPRD = Clinical P ac ice Resea ch Da alink gene al p ac ice da abase. ROC = ecei e
ope a ing cha ac e is ics. Some deciles did no ha e co esponding p ognos ic sco es, an indica ion ha he
sco es we e no no mally dis ibu ed.
Figu e 1. ROC cu es o he alida ed p ognos ic sco es, o males ( op) and emales (bo om): he
p edic ion model de eloped using he PCCIU da a was applied o he CPRD da a.
www.na u e.com/scien i ic epo s/
4
Scien i ic RepoR s | 7:44702 | DOI: 10.1038/s ep44702
in his alida ion s udy (males 0.66, emales 0.71) compa ed o he de elopmen s udy (males 0.83, emales 0.85);
none heless, hese es ima es s ill ha e he po en ial o be e y use ul in clinical p ac ice.
The CPRD da abase is a well-cha ac e ised popula ion-based p ima y ca e da abase and is one o he bes
alida ed la ge p ima y ca e esea ch da abases in he wo ld13,14. Wi h su icien sample size and powe , we ha e
– o he i s ime – alida ed a COPD p edic ion sco e in a di e en da ase , ob aining simila isk ac o s and
analogous accu acy es ima es o disc imina e COPD cases om non-cases compa ed o he measu es de i ed
om he p edic ion de elopmen s udy11. In compa ison o he p edic ion de elopmen s udy using he PCCIU
da abase11, he coding o smoking and as hma was mo e comple e in he cu en s udy using he CPRD da a. No
di e ences we e howe e seen when ei he o he CPRD- and PCCIU-de i ed codes we e used.
A limi a ion o his s udy is ha i was based on a ma ched case-con ol design bu nes ed in a longi udi-
nal popula ion coho (wi h es ima ion o he condi ional odds a ios o he in luence o he isk ac o s on he
de elopmen o COPD), whe eas he p edic ion de elopmen s udy was based on a ollow-up coho s udy (wi h
es ima ion o he haza d a ios o he in luence o he isk ac o s on a 10-yea isk o COPD). Ne e heless, he
es ima es o associa ions be ween he a ious isk ac o s and he de elopmen o COPD we e bo h in he di ec-
ion and magni ude compa able be ween he alida ion and p edic ion de elopmen s udies. Al hough he meas-
u es o accu acy (using he ROCAUC) o he p edic ion model in disc imina ing be ween COPD cases and con ols
we e sligh ly lowe in he cu en s udy compa ed o he p edic ion de elopmen s udy, he di e ences we e as
expec ed, gi en p e ious e idence in his espec , i.e. accu acy measu es o p edic ion models mo e o en a e seen
o be lowe in an ex e nally alida ed da ase compa ed o es ima es de i ed om he p edic ion de elopmen
da ase 15. This wo k has he e o e se ed o con i m he impo ance o unde aking ex e nal alida ion s udies.
Fu he limi a ion o ou wo k is ha , spi ome y measu es, which a e he gold s anda d o diagnosing COPD,
a e no ou inely eco ded in he GP da abases, hence we could no u ilise hem in his s udy. Simila ly, pack-yea s
o smoking, a desi ed exposu e indica o o assess he causal impac o smoking, is no ou inely collec ed in GP
da abases, hence we did no conside i in his s udy.
As he i s alida ion o a isk sco e o p edic ing he de elopmen o COPD in a di e en ex e nal da abase,
he e is no applicable p e ious s udy o compa e he esul s. O e all, he isk es ima es o he s udied isk ac o s
(smoking, p io as hma, and socioeconomic s a us) we e compa able o he es ima es om he p edic ion model
de elopmen s udy. Whils a ecen s udy using CPRD unde ook an ex e nal alida ion o he p edic ion sco es,
he alida ion wo k was done wi hin he same da abase: he au ho s spli he o iginal da a in o wo (de elop-
men and alida ion samples) da ase s, using one da ase o de elop he p edic ion sco es and he second da ase
comp ising o 20 CPRD gene al p ac ices o alida e he sco es12. Ex e nal alida ion o p edic ion models aims
o assess he gene alisabili y o he de i ed model in an app op ia e simila pa ien popula ion, bu in a di e en
con ex ; his wo k he e o e needs o be unde aken in a new da ase 15–17.
In compa ison o he ecen s udy using CPRD da abase o de elop and alida e a p edic ion model12, he
isk es ima es wi h ega ds o smoking and as hma, al hough in he same di ec ion as obse ed in he cu en
s udy, somewha di e ed in magni ude, possibly as di e en de ini ions we e used be ween he wo s udies. In he
cu en s udy we de ined smoking as “e e smoking” while he p e ious s udy di e en ia ed be ween o me and
cu en smoking12. Ano he di e ence be ween he cu en s udy and he p e ious CPRD s udy was ha a iables
included in he inal model di e ed, which may explain he di e ences obse ed: whils he inal model in he
cu en s udy included smoking s a us, p io as hma, and IMD, he p e ious s udy included smoking s a us, p io
as hma, salbu amol p esc ip ion, and lowe espi a o y ac in ec ions. Fu he di e ence ela es o he ime co -
e age o pa icipan s’ en olmen in o he wo s udies: he cu en s udy included pa icipan s be ween 1992 and
2012, whe eas he p e ious s udy co e ed be ween 2000 and 2006. The UK Quali y and Ou comes F amewo k
(QOF) o COPD s a ed in 2004, a which poin he coding o COPD imp o ed5,6. Howe e , since bo h s udies
con ained da a bo h p io o and a e he s a o he QOF, we belie e he ime co e age o s udies may no ha e
subs an ially in luenced he obse ed a ia ions in esul s.
O he p e ious s udies om he US7,8 and Denma k9 de eloped p edic ion algo i hms aimed a iden i ying
COPD pa ien s wi h al eady es ablished disease. They we e also based on seconda y ca e da a, such as admin-
is a i e claims da a, ou pa ien pha macy da a, and hospi al admissions da a, hence con as he cu en s udy,
which used popula ion-based p ima y ca e o alida e p edic ion sco es aimed a iden i ying a - isk indi iduals
be o e he onse o disease.
As he incidence o COPD and mo ali y con inues o ise globally, ou abili y o de ec cases be o e hey man-
i es is c ucial6,11,12. The cu en COPD p edic ion model, now ex e nally alida ed in a di e en da ase , bu com-
pa able popula ion, p o ides a con incing oppo uni y o e alua ing i s use ulness and applicabili y in clinical
ca e se ings o iden i ying indi iduals a high isk o de eloping he disease18. The cu en alida ion con i ms
he impo ance o smoking his o y, p io as hma, and socioeconomic s a us19, which in combina ion p o ide a
composi e p edic ion sco e o accu a ely iden i y indi iduals a di e en isk ca ego ies (based on he ou p e-
dic ion model, low ca ego y had isk sco es ≤ 6, medium isk sco e o 7, and high isk sco e 8–10) o de eloping
COPD11. The cu en s udy is he i s o ex e nally alida e a COPD p edic ion sco e. Whils eme ging COPD
Range o alues o deciles o p ognos ic sco es A ea unde ROC
cu e (95% CI)1 (lowes ) 2345678910 (highes )
P ognos ic sco es 0.00–.21 0.22–0.59 0.60–1.98 1.99–2.19 — 2.20–2.36 2.37–2.57 2.58–2.75 2.76–3.80 3.81–4.37 0.74 (0.73–0.74)
Table 4. P ognos ic sco es de i ed om he CPRD da a. CPRD = Clinical P ac ice Resea ch Da alink
gene al p ac ice da abase. ROC = ecei e ope a ing cha ac e is ic.
www.na u e.com/scien i ic epo s/
5
Scien i ic RepoR s | 7:44702 | DOI: 10.1038/s ep44702
p edic ion models om o he con ex s also need o be ex e nally alida ed in di e en da ase s, we belie e ha
he isk sco es de i ed om he cu en ex e nal alida ion exe cise can now be assessed whe he i can applied
in clinical se ings wi h a concu en impac assessmen o i s pe o mance.
In conclusion, using a la ge in e na ionally espec ed da abase, we ha e – o he i s ime – alida ed a COPD
p edic ion model o iden i ying a - isk indi iduals p io o he onse o disease in a di e en da abase bu com-
pa able popula ion, wi h accep able accu acy a disc imina ing be ween COPD cases and non-COPD cases. Key
p edic o s o onse o COPD include smoking, p io as hma, and socioeconomic s a us. Conside a ion now needs
o be made, wi h a concu en impac assessmen on pe o mance, on whe he he alida ed isk sco e is use ul
and can be eadily applied in clinical p ac ice o iden i ying hose a isk o de eloping COPD.
Me hods
S udy popula ion. The CPRD da abase is a alida ed compu e ised, anonymised and longi udinal p ima y
ca e da abase, conside ed by many as he gold s anda d o ou ine clinical esea ch da a (www.cp d.com)13,14. I is
join ly unded by he UK Na ional Heal h Se ice (NHS) and he Medicines and Heal hca e p oduc s Regula o y
Agency (MHRA). Da a a e linkable o o he heal hca e and social ca e da a sou ces, and a e egula ly used o
conduc bo h obse a ional and in e en ional esea ch in he UK. P esen ly he da abase comp ises a ound 14
million pa ien s de i ed om 660 p ima y ca e p ac ices ac oss he UK13,14. We ecei ed access o CPRD unde
licence om he Medical Resea ch Council (MRC). The CPRD G oup has ob ained e hical app o al om a Mul i-
cen e Resea ch E hics Commi ee o all pu ely obse a ional esea ch using CPRD da abase. The p o ocol o
he cu en s udy was app o ed by he Independen Scien i ic Ad iso y Commi ee o CPRD (p o ocol numbe
10_084 R). All he s udy me hods we e pe o med in acco dance wi h he ele an guidelines and egula ions and
in acco dance wi h bes scien i ic p ac ices.
We ex ac ed a andom sample om he CPRD o COPD cases and con ols aged ≥ 35 yea s: index cases and
hei co esponding con ols we e ma ched a a a io o 1:1 on GP p ac ice, sex, and yea o bi h (wi hin wo
yea s). Cases comp ised o indi iduals wi h i s eco ded COPD diagnosis and who we e ollowed o a leas
i e yea s p io o he index da e (i.e., da e o d awing he sample). Iden i ica ion o cases and con ols and de i-
ni ion o o he s udy a iables we e based on he Read Clinical Classi ica ion Sys em, a s anda d coding sys em
p oduced o clinicians in p ima y ca e and which is used o mos p ima y ca e elec onic pa ien eco ds in he
UK (a comple e lis o Read codes used o his s udy is gi en in Supplemen a y File 1). A con ol mus ha e been
egis e ed a he same p ac ice a he ime o he index da e o he co esponding case and should ha e had a leas
i e yea s o ollow-up in he same p ac ice p io o he index da e o he case. The index da e o he con ols was
he da e o he eco ding o COPD o he co esponding ma ched case.
In o al, we sampled 38,597 case-con ol pai s ( o al N = 77,194). This was he maximum numbe o allowable
pa ien s ex ac ed unde he Medical Resea ch Council license. O hese, 188 (84 [0.22%] cases and 104 con-
ols [0.27%]) had missing da a o IMD, hence we e excluded om analyses. The IMD is he UK go e nmen ’s
measu e o assess household’s socio-economic s a us based on he le el o dep i a ion o an a ea (h p://census.
ukda ase ice.ac.uk/ge -da a/ ela ed/dep i a ion). Two con ols wi hou ID numbe s and hei co esponding
cases we e u he excluded, esul ing in a o al o 77,002 (38,511 cases and 38,491 con ols) as he inal sample
o analyses.
Assessmen and de ini ion o isk ac o s. F om CPRD, we ex ac ed he same a iables used in he
PCCIU da a o de elop he p edic ion model (i.e., smoking, p io as hma, and IMD); he cu en s udy sample
was ma ched o age and sex. Smoking s a us was ca ego ised in o “ne e smoke ” (i.e., pa ien s eco ded as
“non-smoke ” a any ime and no coding as “smoke ” o “ex-smoke ” a any o he ime) o “e e smoke ” (i.e.,
pa ien s eco ded as “smoke ” o “ex-smoke ” a any ime); “ne e smoke ” was he e e ence ca ego y. P io
as hma was ca ego ised as “no” o “yes”; “no” was he e e ence ca ego y. As he s udy was nes ed wi hin a longi-
udinal coho , he iming o occu ence o as hma was de e mined and we ensu ed ha only as hma cases occu -
ing p io o COPD we e de ined as posi i e. IMD was ca ego ised in o quin iles (quin ile 1 as leas dep i ed and
quin ile 5 as mos dep i ed); quin ile 1 was he e e ence ca ego y.
We obse ed some di e ences in coding o smoking and as hma be ween he CPRD and PCCIU da a: in
compa ison o he codes in PCCIU, CPRD comp ised o ou ex a codes o smoking s a us and 15 ex a codes
o as hma (see supplemen a y File 1). This esul ed in abou 0.5% ex a e e smoke s and abou 12% ex a p io
as hma cases by using CPRD codes compa ed wi h PCCIU codes. The Read code, H33.00 (As hma), con ibu ed
o mos (85%) o he ex a as hma cases in CPRD. Whils we used CPRD-de i ed codes h oughou he alida ion
exe cise, we also sepa a ely analysed PCCIU-de i ed codes o assess whe he he e we e any di e ences be ween
he wo codes in e ms o he di ec ion and magni ude o he p edic ion sco es.
S a is ical analysis. The S a a codes o da a p epa a ion and analyses a e p esen ed in Supplemen a y Files
2 and 3, espec i ely. Fo desc ip i e analysis we calcula ed he equencies o pa icipan s’ backg ound cha -
ac e is ics by cases and con ols. Assessmen o he associa ions be ween he isk ac o s and he isk o COPD
was pe o med by calcula ing he unadjus ed and adjus ed isk es ima es using condi ional logis ic eg ession.
These es ima es a e epo ed as odds a io (OR) and accompanied by hei 95% con idence in e als (95% CI). I
should be no ed ha while he de elopmen model was based on a longi udinal da a (hence modelling using Cox
eg ession), he cu en alida ion wo k is based on a ma ched case-con ol s udy (hence he calcula ion using
condi ional logis ic eg ession).
A wo-pa app oach was used o unde ake he ex e nal alida ion o he p edic ion model. Fi s , he isk
sco es de i ed om he p edic ion model (based on coe icien s ac oss ca ego ies o he a iables included in he
ully adjus ed Cox eg ession model)11 we e applied o he co esponding ca ego ies o each p edic o in he al-
ida ion model, hus de i ing he p ognos ic sco es o each indi idual in he alida ion model o he CPRD da a;

www.na u e.com/scien i ic epo s/
6
Scien i ic RepoR s | 7:44702 | DOI: 10.1038/s ep44702
his was done sepa a ely o males and emales. The esul ing isk sco es we e hen di ided in o 10 isk ca ego ies
(deciles). The app oach is desc ibed in Table5.
Fo example, a 60-yea -old male e e smoke , wi h IMD le el 3 and no p e ious his o y o as hma would yield
a p ognos ic index (PI) o 5.0 ( isk ca ego y 9 = high isk o u u e COPD); a 40-yea -old emale ne e smoke ,
wi h IMD 3 and a p e ious his o y o as hma would yield a PI o 2.2 ( isk ca ego y 3 = low isk o u u e COPD).
In he second pa o he alida ion model, he p ognos ic sco es o each indi idual was de i ed solely based
on he CPRD da a using he me hod applied in he model de elopmen : i.e., calcula ion o he isk sco es as he
sum o he eg ession coe icien s ac oss he di e en ca ego ies o he a iables in he ully adjus ed condi ional
logis ic eg ession model. Due o he CPRD alida ion da ase being ma ched o sex and age, he sco es we e
calcula ed o males and emales combined and no sepa a ed by sex. The esul an isk sco es we e hen di ided
in o deciles. Illus a i ely, gi en he be a eg ession coe icien s o he ca ego ies o he di e en ac o s as β,
calcula ion o he isk sco es was gi en as ollows:
=β +β +β +β
+β +β
⁎⁎⁎
⁎⁎ ⁎
P ognos icsco eSmoking IMD2 IMD3
IMD4 IMD5 As hma
(smoking)(IMD2)(IMD) (IMD)
(IMD)(as hma)
Whe e β
(smoking)*Smoking is he coe icien o a smoke s ne e smoke as he e e ence; β
(IMD2)*IMD2 is he
coe icien o an indi idual in he second quin ile o IMD o β
(IMD)*IMD5 o an indi idual in he highes quin ile
o IMD s i s quin ile as he e e ence; and β
(as hma)*As hma o an indi idual wi h as hma s non-as hma as he
e e ence.
Fo each alida ion model, he accu acy o he p ognos ic sco es in disc imina ing be ween COPD and
non-COPD pa ien s was es ima ed by calcula ing he a ea unde he ecei e ope a o cha ac e is ic cu e
(ROCAUC) o all alues o he sco es.
Re e ences
1. Lozano, R. e al. Global and egional mo ali y om 235 causes o dea h o 20 age g oups in 1990 and 2010: a sys ema ic analysis o
he Global Bu den o Disease S udy 2010. Lance . 380, 2095–2128 (2012).
2. Wo ld Heal h O ganiza ion. Gene a: Wo ld heal h epo : ch onic espi a o y diseases. A ailable om: h p://www.who.in /
espi a o y/copd/bu den/en/ (Accessed July 10, 2015).
3. Ame ican Lung Associa ion. T ends in COPD (Ch onic B onchi is and Emphysema): mo bidi y and mo ali y. Ma ch 2013. A ailable
om: h p://www.lung.o g/ inding-cu es/ou - esea ch/ end- epo s/copd- end- epo .pd (Accessed July 10, 2015).
4. P ide, N. B. & So iano, J. B. Ch onic obs uc i e pulmona y disease in he Uni ed Kingdom: ends in mo ali y, mo bidi y, and
smoking. Cu Opin Pulm Med. 8, 95–101 (2002).
5. Calde ón-La aňaga, A. e al. Associa ion o popula ion and p ima y heal hca e ac o s wi h hospi al admission a es o ch onic
obs uc i e pulmona y disease in England: na ional c oss-sec ional s udy. Tho ax 66, 191–196 (2011).
6. Quin , J. F. A e clinical isk sco es o COPD use ul? BMJ Open Resp Res. 2, e000072 (2015).
7. Mapel, D. W. e al. An algo i hm o he iden i ica ion o undiagnosed COPD cases using adminis a i e claims da a. J Manag Ca e
Pha m. 12, 458–465 (2006).
8. Mapel, W. E., Pe e sen H., Robe s M. H. e al. Can ou pa ien pha macy da a iden i y pe sons wi h undiagnosed COPD? Am
J Manag Ca e. 16, 505–512 (2010).
9. Smid h, M., Sokolowski, I., Kae s ang, L. & Veds ed, P. De eloping an algo i hm o iden i y people wi h ch onic obs uc i e
pulmona y disease (COPD) using adminis a i e da a. BMC Med In o m Decis. 12, 38 (2012).
10. Hill, K. e al. P e alence and unde diagnosis o ch onic obs uc i e pulmona y disease among pa ien s a isk in p ima y ca e.
CMAJ. 182, 673–678 (2010).
11. Ko z, D., Simpson, C. R., Viech baue , W., an Schayck, O. C. & Sheikh, A. De elopmen and alida ion o model o p edic he 10-
yea isk o gene al p ac i ione - eco ded COPD. npj P im Ca e Respi Med. 24, 14011 (2014).
12. Ha oon, S. e al. P edic ing he isk o COPD in p ima y ca e: de elopmen and alida ion o a clinical isk sco e. BMJ Open Resp
Res. 1, e000060 (2014).
P ognos ic sco es o Males
0.722590422761375000*AgeG oup2 + 1.354003015630730000*AgeG oup3 +
1.794461805033540000*AgeG oup3 + 2.268067060819270000*AgeG oup5 +
2.640068571991060000*AgeG oup6 + 3.462262041329010000* AgeG oup7 +
1.905731297921800000*Smoking + 0.307312890106045000*IMD2 +
0.468566344550995000*IMD3 + 0.647039365745038000*IMD4 +
0.926189176930742000*IMD5 + 1.214816131312250000*As hma
P ognos ic sco es o Females
0.719505024838863000*AgeG oup2 + 1.311267194962520000*AgeG oup3 +
1.702959068049090000*AgeG oup4 + 2.098189831042430000*AgeG oup5
2.352907331580150000*AgeG oup6 + 3.248496488525190000*AgeG oup7
2.262261867746550000*Smoking + 0.223346323577682000*IMD2 +
0.498918486471617000*IMD3 + 0.666610932992558000*IMD4 +
0.948500789111445000*IMD5 + 1.025043519409850000*As hma
Table 5. Calcula ion o p edic ion isk sco es o males and emales de i ed om he PCCIU da ase . Age
ca ego ies: AgeG oup1 (35–39 y s: e e ence); AgeG oup2 (40–44 y s); AgeG oup3 (45–49 y s); AgeG oup4
(50–54 y s); AgeG oup5 (55–59 y s); AgeG oup6 (60–64 y s); AgeG oup7 (65 + y s). IMD ca ego ies: IMD1
(1s Quin ile, leas dep i ed: e e ence); IMD2 (2nd Quin ile); IMD3 (3 d Quin ile); IMD4 (4 h Quin ile); IMD5
(5 h Quin ile, mos dep i ed).
www.na u e.com/scien i ic epo s/
7
Scien i ic RepoR s | 7:44702 | DOI: 10.1038/s ep44702
13. Williams, T., an S aa, T., Pu i, S. & Ea on, S. Recen ad ances and use o he Gene al P ac ice Resea ch Da abase as an example o a
UK P ima y Ca e Da a esou ce. The Ad D ug Sa . 3, 89–99 (2012).
14. Quin , J. K. e al. Valida ion o ch onic obs uc i e pulmona y disease eco ding in he Clinical P ac ice Resea ch Da alink (CPRD-
GOLD). BMJ Open. 4, e005540 (2014).
15. Sion is, G. C. M., Tzoulaki, I., Cas aldi, P. J. & ioannidis, J. P. Ex e nal alida ion o new isk p edic ion models is in equen and
e eals wo se p ognos ic disc imina ion. J Clin Epidemiol. 68, 25–34 (2015).
16. Al man, D. G. & Roys on. P. Wha do we mean by alida ing a p ognos ic model? S a is Med. 19, 453–473 (2000).
17. Al man, D. G., Ve gouwe, Y., Roys on, P. & Moons, K. G. P ognosis and p ognos ic esea ch: alida ing a p ognos ic model. BMJ.
338, b605 (2009).
18. Robson, J. e al. The NHS Heal h Check p og amme: implemen a ion in eas London 2009–2011. BMJ Open. 5, e007578 (2015).
19. Ge shon, A. S., Wa ne , L., Cascagne e, P., Vic o , J. C. & To, T. Li e ime isk o de eloping ch onic obs uc i e pulmona y disease:
a longi udinal popula ion s udy. Lance 378, 991–996 (2011).
Acknowledgemen s
Access o he CPRD da abase was unded h ough he Medical Resea ch Council’s licence ag eemen wi h he
Medicines and Heal hca e p oduc s Regula o y Agency. BN, CS and AS we e suppo ed by he Fa Ins i u e and
As hma UK Cen e o Applied Resea ch. DK ecei ed addi ional suppo om he Minis y o Inno a ion,
Science and Resea ch o he Ge man Fede al S a e o No h Rhine-Wes phalia (“NRW-Rückkeh p og amm”).
Au ho Con ibu ions
B.N. unde ook da a analysis and d a ing o he manusc ip wi h se e al ounds o c i ical commen s om D.K.
A.S. and C.S. con ibu ed o he de elopmen o he plans o his wo k, we e in ol ed in d a ing he p o ocol and
commen ed c i ically on ea lie d a s o he manusc ip .
Addi ional In o ma ion
Supplemen a y in o ma ion accompanies his pape a h p://www.na u e.com/s ep
Compe ing In e es s: The au ho s decla e no compe ing inancial in e es s.
How o ci e his a icle: Nwa u, B. I. e al. Ex e nal alida ion o a COPD p edic ion model using popula ion-
based p ima y ca e da a: a nes ed case-con ol s udy. Sci. Rep. 7, 44702; doi: 10.1038/s ep44702 (2017).
Publishe 's no e: Sp inge Na u e emains neu al wi h ega d o ju isdic ional claims in published maps and
ins i u ional a ilia ions.
This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 In e na ional License. The images
o o he hi d pa y ma e ial in his a icle a e included in he a icle’s C ea i e Commons license,
unless indica ed o he wise in he c edi line; i he ma e ial is no included unde he C ea i e Commons license,
use s will need o ob ain pe mission om he license holde o ep oduce he ma e ial. To iew a copy o his
license, isi h p://c ea i ecommons.o g/licenses/by/4.0/
© The Au ho (s) 2017