Machine learning models in predicting health care costs in patients with a recent acute coronary syndrome : A prospective pilot study

This is a sel -a chi ed e sion o an o iginal a icle. This e sion may di e om he o iginal in pagina ion and ypog aphic de ails. Au ho (s): Ti le: Yea : Ve sion: Copy igh : Righ s: Righ s u l: Please ci e he o iginal e sion: CC BY 4.0 h ps://c ea i ecommons.o g/licenses/by/4.0/ Machine lea ning models in p edic ing heal h ca e cos s in pa ien s wi h a ecen acu e co ona y synd ome : A p ospec i e pilo s udy © 2023 Hea Rhy hm Socie y Published e sion Hau ala, A o J.; Sha azipou , Babooshka; A sa , Beki ; Tulppo, Mikko P.; Mie inen, Kaisa Hau ala, A. J., Sha azipou , B., A sa , B., Tulppo, M. P., & Mie inen, K. (2023). Machine lea ning models in p edic ing heal h ca e cos s in pa ien s wi h a ecen acu e co ona y synd ome : A p ospec i e pilo s udy. Ca dio ascula Digi al Heal h Jou nal, 4(4), 137-142. h ps://doi.o g/10.1016/j.c dhj.2023.05.001 2023 Machine lea ning models in p edic ing heal h ca e cos s in pa ien s wi h a ecen acu e co ona y synd ome: A p ospec i e pilo s udy A o J. Hau ala, PhD,*Babooshka Sha azipou , PhD, † Beki A sa , PhD, † Mikko P. Tulppo, PhD, ‡ Kaisa Mie inen, PhD † F om he *Facul y o Spo and Heal h Sciences, Uni e si y o Jy askyla, Jy askyla, Finland, † Facul y o In o ma ion Technology, Uni e si y o Jy askyla, Jy askyla, Finland, and ‡ Resea ch Uni o Biomedicine and In e nal Medicine, Medical Resea ch Cen e Oulu, Oulu Uni e si y Hospi al, Uni e si y o Oulu, Oulu, Finland. BACKGROUND Heal h ca e budge s a e limi ed, equi ing he op imal use o esou ces. Machine lea ning (ML) me hods may ha e an eno mous po en ial o e ec i e use o heal h ca e e- sou ces. OBJECTIVE We assessed he applicabili y o selec ed ML ools o e alua e he con ibu ion o known isk ma ke s o p ognosis o co ona y a e y disease o p edic heal h ca e cos s o all easons in pa ien s wi h a ecen acu e co ona y synd ome (n 565, aged 65 69 yea s) o 1-yea ollow-up. METHODS Risk ma ke s we e assessed a baseline, and heal h ca e cos s we e collec ed om elec onic heal h egis ies. The C oss- decomposi ion algo i hms we e used o ank he conside ed isk ma ke s based on hei impac s on a iances. Then eg ession anal- ysis was pe o med o p edic cos s by en e ing he fi s op- anking isk ma ke and adding he nex -bes ma ke s, one by one, o build up al oge he 13 p edic i e models. RESULTS The a e age annual heal h ca e cos s we e V2601 6 V5378 pe pa ien . The Dep ession Scale showed he highes p e- dic i e alue ( 50.395), accoun ing o 16% o he cos s (P5.001). When he nex 2 anked ma ke s (LDL choles e ol, 5 0.230; and le en icula ejec ion ac ion, 5-0.227, espec- i ely) we e added o he model, he p edic i e alue was 24% o he cos s (P5.001). CONCLUSION Highe dep ession sco e is he p ima y a iable o e- cas ing heal h ca e cos s in 1-yea ollow-up among acu e co ona y synd ome pa ien s. The ML ools may help decision-making when planning op imal u iliza ion o ea men s a egies. KEYWORDS Co ona y a e y disease; Co ona y hea disease; A ifi- cial in elligence; Heal h ca e cos s; Economic e alua ion (Ca dio ascula Digi al Heal h Jou nal 2023;4:137–142) ©2023 Hea Rhy hm Socie y. This is an open access a icle unde he CC BY license (h p://c ea i ecommons.o g/licenses/by/4.0/). In oduc ion Ca dio ascula disease incidence and mo ali y a es a e declining in many coun ies in Eu ope bu s ill emain a majo cause o mo bidi y and mo ali y 1 wi h a significan impac on heal h ca e cos s. The economic bu den o ca dio ascula diseases in he Eu opean Union egion was e alua ed o be V169 billion annually and 62% o hese cos s we e ela ed o heal h ca e. 2 The da a om he Uni ed S a es show ha expendi u e on ca dio ascula disease and ca dio ascula isk ac o s in 2016 was $320 billion. Heal h se ices o ischemic hea disease ($80 billion) and ea men o hype - ension ($71 billion) we e he main causes o he cos s, ol- lowed by ea men o hype lipidemia. 3 Based on he EUROASPIRE su ey, 4 cos s o op imized ailo ed p e en- ion such as smoking cessa ion, die and exe cise, be e man- agemen o ele a ed blood p essu e and/o low-densi y lipop o ein (LDL) choles e ol, and sa ings o a oided e en s we e es ima ed based on coun y-specific da a. The esul s showed ha op imizing seconda y p e en ion is clea ly cos -e ec i e compa ed wi h he cu en gene al guideline– o ien ed p e en ion. 4 Heal h ca e p o ide s wo ldwide a e equi ed o se p io - i ies and alloca e esou ces wi hin he cons ain o limi ed unding. Howe e , decision make s may no be well equip- ped o make explici a ioning decisions and may o en ely on his o ical o poli ical esou ce alloca ion p ocesses. 5 The e o e, economic e alua ion o heal h ca e o ope a ional planning and decision-making is i al o alloca ion o e- sou ces o e ec i e ea men s ha p o ide pa ien s he g ea - es possible heal h benefi s a easonable cos s. Since heal h ca e sys ems, ca e p ac ices, and ela i e p ices o heal h ca e in es men s a y om coun y o coun y, i is impo an o ha e coun y-specific da a o suppo decision-making. 6 ClinicalT ials.go iden ifie : Reco d NCT01916525. Add ess ep in eques s and co espondence: D A o J. Hau ala, Facul y o Spo s and Heal h Sciences, Uni e si y o Jy askyla, PO Box 35, FI-40014 Uni e si y o Jy askyla, Finland. E-mail add ess: a o.j.hau ala@jyu.fi. 2666-6936/© 2023 Hea Rhy hm Socie y. This is an open access a icle unde he CC BY license (h p://c ea i ecommons.o g/licenses/by/4.0/). h ps://doi.o g/10.1016/j.c dhj.2023.05.001 Machine lea ning (ML) and a ificial in elligence me hods may ha e a conside ably high po en ial o bo h e ec i e and cos -e ec i e use o heal h ca e esou ces when implemen ed in clinical p ac ice. 7,8 Fo example, Schwalm and colleagues 9 showed ecen ly ha an ML p edic ion model used as an on- line decision suppo ool by e e ing physicians could imp o e he diagnos ic yield o in asi e co ona y angiog aphy in s able co ona y a e y disease (CAD) pa ien s. The decision- make planning o op imal use o heal h ca e esou ces may benefi om he p edic ion model o he mos impo an isk ac o o combina ion o hose con ibu ing mos o heal h ca e cos s. The ea u e impo ance analysis is widely used in p edic i e modeling, ep esen ing he significance o he inpu ea u es a he a ge a iables p edic ion by calcula ing p edic- i e sco es. 10 We conduc ed an analysis using selec ed ea u e impo ance analysis ools o assess he con ibu ion o well- add essed causal and modifiable isk ma ke s o p ognosis o CAD a baseline o p edic heal h ca e cos s in pa ien s wi h a ecen acu e co ona y synd ome (ACS) o 1-yea ollow-up in he Finnish heal h ca e sys em. Me hods S udy popula ion This s udy is pa o he EFEX-CARE (E ec i eness o Ex- e cise Ca diac Rehabili a ion) s udy ha has been egis e ed a ClinicalT ials.go (Iden ifie Reco d NCT01916525). The pa ien s in he EFEX-CARE s udy ha e been ec ui ed om a consecu i e se ies o ACS pa ien s in he Di ision o Ca di- ology o he Oulu Uni e si y Hospi al. They all unde wen co ona y angiog aphy o confi m he CAD. The s udy popu- la ion o he EFEX-CARE s udy has been p e iously desc ibed in de ail, 11 bu o pu i b iefly, exclusion c i e ia included NYHA class III, scheduled o eme gency p oced- u e o bypass su ge y, uns able angina pec o is, se e e pe- iphe al a he oscle osis, diabe ic e inopa hy o neu opa hy, o inabili y o independen daily physical ac i i y, eg, owing o musculoskele al p oblems. In his s udy, we epo heal h ca e cos s o 1 yea ollow-up and isk ma ke da a a base- line measu ed abou 2–3 weeks a e hei hospi al discha ge o he pa ien s ea ed acco ding o usual ca e. Al oge he , all da a needed o analysis we e a ailable o 65 pa ien s. The s udy epo ed in his pape adhe ed o he CONSORT guide- lines and was ca ied ou acco ding o he Decla a ion o Hel- sinki; he local commi ee o esea ch e hics o he No he n Os obo hnia Hospi al Dis ic app o ed he p o ocol. All he subjec s ga e w i en in o med consen . Assessmen o pa ien cha ac e is ics, isk ma ke s, and heal h ca e cos s Body weigh and heigh we e measu ed o assess body composi ion. Blood p essu e was measu ed in a supine posi- ion a e a 10-minu e es ing pe iod acco ding o he cu en guideline. Sel - a ed dep ession was assessed by using he Dep ession Scale (DEPS) ques ionnai e. 12 The hospi al eg- is y and s anda d ques ionnai es we e used o ga he he da a ega ding smoking s a us, alcohol use diso de s iden ifi- ca ion (AUDIT-C), 13 medica ion, his o y o acu e myoca - dial in a c ion, and e ascula iza ion. Assessmen o le en icula sys olic unc ion was pe o med using 2-D echo- ca diog aphy (Vi id 7; GE Heal hca e, Wauwa osa, WI). Blood samples om as ing s age we e ob ained o analysis o plasma glucose and glyca ed hemoglobin (HbA1c), blood lipids, insulin, and high-sensi i i y C- eac i e p o ein a e a 12-hou o e nigh as using consis en me hods (Oulu Uni e si y Hospi al, Oulu, Finland). An inc emen al symp om-limi ed maximal exe cise es was pe o med a he Oulu Uni e si y Hospi al on a bicycle e gome e (Mona k E gomedic 839 E; Mona k Exe cise AB, Vansb o, Sweden) o assessmen o maximal physical exe cise capaci y (me a- bolic equi alen s). The 15D ques ionnai e was used o eco d heal h- ela ed quali y o li e 14 and i was comple ed by he pa- ien s a he hospi al be o e hospi al discha ge. In he es ima- ion o heal h ca e cos s, bo h specialized and p ima y heal h ca e se ices, as well as he cos s o occupa ional heal h ca e se ices, we e conside ed. Social secu i y ID numbe s we e used o de e mine isi s o ambula o y ca e, numbe o ea - men days, and use o ex e nal se ices o calcula e heal h ca e cos s a ising om he use o heal h se ices on he pa o specialized heal h ca e. The exac cos s we e measu ed based on in oicing (using Diagnosis Rela ed G oups classifica ion). In o ma ion on he use o p ima y heal h ca e and cos s ela ed o i was ob ained om elec- onic heal h egis ies by using unique social secu i y ID numbe s o de e mine isi s o he doc o , o he significan examina ions such as la ge adiog aphs, and in-wa d ea - men days. Fu he mo e, he use o home ca e and possible ins i u ional ca e (eg, assis ed ca e home, e c) was de e mined om egis ies. The epo o he Social Insu ance Ins i u e o Finland (KELA) 15 was used o es ima e occupa ional heal h ca e se ice cos s. Finally, all cos s we e managed as 2015 alues. Because o he 1-yea ime ho izon o he analysis, no discoun ing was applied. KEY FINDINGS Ad anced da a analy ics and machine lea ning ools can po en ially be used o p edic heal h ca e cos s in eal- wo ld clinical se ings. Ou pilo s udy showed o he fi s ime, using machine lea ning ools, ha dep ession, exp essed as highe dep ession sco es, is he p ima y heal h measu emen o ecas ing heal h ca e cos s o all easons, ollowed by low-densi y lipop o ein choles e ol and le en ic- ula ejec ion ac ion, in 1-yea ollow-up among acu e co ona y synd ome pa ien s. Applica ions o machine lea ning and a ificial in elli- gence me hods o heal h ca e cos s can p o ide in o - ma ion ha may be help ul o decision-making when planning op imal u iliza ion o ea men s a egies and esou ces in heal h ca e se ings. 138 Ca dio ascula Digi al Heal h Jou nal, Vol 4, No 4, Augus 2023 De elopmen o p edic i e models In p edic i e modeling, he significance o he inpu ea u es a he a ge a iable p edic ions is ep esen ed by some sco es calcula ed h ough a so-called ea u e impo ance analysis. 10 In o he wo ds, hese sco es demons a e he impo ance o a ea u e/ a iable o a p edic ion. Indeed, ea u e impo ance analysis is o en used, as ea u e selec ion, o educe he num- be o inpu a iables, bo h o educe he compu a ional cos o modeling and, in some cases, o imp o e he model’spe o - mance. These ea u e impo ance sco es p o ide insigh in o he da a and models and play a c ucial ole in imp o ing he e ficiency o he p edic i e models h ough ea u e selec- ion 16,17 and dimensionali y educ ion. 18,19 Va ious ea u e impo ance me hods ha e been de eloped in he li e a u e, eg, based on s a is ical co ela ions and a iances. Howe e , he choice o me hods depends on a iables and he ype o da a. The e o e, i is ecommended o e alua e a ious ech- niques o find sui able ones. We conduc ed a ea u e impo ance analysis on ou da a- se o check he significance o selec ed isk ac o s in p edic - ing all heal h ca e cos s. A e we es ed a ious ea u e impo ance me hods, he ollowing ones p o ed hei s abili y in se e al es s on andomly selec ed subg oups o samples om he da ase : C oss decomposi ion 20 ; pa ial leas squa es (PLS) canonical analysis (PLSC), PLS based on singula alue decomposi ion (PLSSVD), PLS eg ession (PLSRe- g ession), and canonical co ela ion analysis (CCA) algo- i hms ank he conside ed isk ac o s based on hei impac s on a iances (ie, show which isk ac o leads o he highes a iance in cos s). PLSReg ession anks consid- e ed isk ac o s based on absolu e alues o he co ela ions be ween he isk ac o and he cos s. Analysis o a iance (ANOVA) es has also been used o ea u e selec ion o ank conside ed isk ac o s based on hei P alues. The used me hods eflec he in insic p edic i e alue o he isk ac o s and a e no dependen on a pa icula p edic i e model ha makes hem mo e sui able in ou case. PLS es i- ma o s a e pa icula ly sui ed when he e is mul icollinea i y among he isk ac o s. 21 A e anking o isk ac o s o heal h ca e cos s, a linea eg ession analysis was pe o med o p edic cos s by en e ing he fi s op- anking isk ma ke and adding he nex -bes ma ke s, one by one, o build up al oge he 13 p e- dic i e models. Desc ip i e s a is ical analyses we e conduc - ed using means, s anda d de ia ions (SDs), and p opo ions, as app op ia e. SPSS so wa e (SPSS 26; SPSS Inc, Chicago, IL) was used o p edic i e da a analyses. S a is ical signifi- cance was defined as a P alue ,.05 o all es s. Resul s Baseline demog aphics, clinical cha ac e is ics, and medica- ion use o he s udy pa icipan s a e illus a ed in Table 1. The o al a e age cos pe ACS pa ien o all easons was V2601 65378 o a 1-yea ollow-up. The anking o isk ac o s o p edic ion o he heal h ca e cos is p esen ed in Figu e 1. The colo code on he igh side ep esen s he isk ac o s anking (1–13) in each ea u e selec- ion me hod. The lowe ank alue (da ke colo in he hea - map) deno es he highe impo ance o he isk ac o . The numbe s in pa en heses (1–13) show he agg ega ed ank o each isk ac o o e hei anking ound in a ious me hods. Table 2 shows he final p edic i e models and hei con i- bu ions o he cos s. Fu he mo e, he di ec ion o each isk ma ke con ibu ion (nega i e o posi i e) is shown as co ela- ion alues. The DEPS showed he highes p edic i e alue Table 1 Baseline demog aphics, clinical cha ac e is ics, heal h ca e cos s, and medica ion use o he s udy g oup (n 565) Va iable ACS pa ien s Men 46 (71%) Pa ien s wi h T2D 11 (17%) Age, yea s 65 69 Weigh , kg 83 614 BMI, kg/m 2 28.0 64.3 Sys olic BP, mm Hg 137 622 Dias olic BP, mm Hg 78 611 Maximal exe cise capaci y, MET 5.6 61.7 Quali y o li e, 15D scale 0.90 60.08 AUDIT-C o alcohol use 2.9 62.4 Dep ession scale 4.6 65.3 Cu en smoke 8 (12%) To al a e age heal h ca e cos pe pa ien Cos o all easons (V) 2601 65378 His o y o AMI NSTEMI 45 (51%) STEMI 22 (34%) Re ascula iza ion PCI 55 (85%) Ea lie CABG 8 (12%) Ca diac unc ion LVEF, % 62 67 CCS class 1.6 60.6 Labo a o y analyses HbA1c, % 6.0 60.8 Fas ing plasma glucose, mmol/L 6.0 61.0 To al choles e ol, mmol/L 3.8 60.7 HDL choles e ol, mmol/L 1.2 60.3 LDL choles e ol, mmol/L 2.1 60.7 T iglyce ides, mmol/L 1.3 60.6 hs-CRP, mg/L 2.7 66.0 Medica ion Be a-blocke s 56 (86%) ACEI o ARB 54 (83%) Lipids 64 (98%) An icoagulan s 64 (98%) Calcium an agonis s 17 (26%) Ni a es 18 (28%) Diu e ics 15 (23%) Values a e means 6SD o numbe (pe cen age) o subjec s. 15D 5heal h- ela ed quali y o li e ques ionnai e; ACEI 5angio ensin- con e ing enzyme inhibi o ; ACS 5acu e co ona y synd ome; AMI 5acu e myoca dial in a c ion; ARB 5angio ensin ecep o blocke ; AUDIT-C 5 Alcohol Use Diso de s Iden ifica ion Tes ; BMI 5body mass index; BP 5 blood p essu e; CABG 5co ona y a e y bypass g a ; CCS 5Canadian Ca - dio ascula Socie y g ading o angina pec o is; HbA1c 5glyca ed hemoglo- bin; HDL 5high-densi y lipop o ein; hs-CRP 5high-sensi i i y C- eac i e p o ein; LDL 5low-densi y lipop o ein; LVEF 5le en icula ejec ion ac- ion; MET 5me abolic equi alen ; NSTEMI 5non-ST-segmen ele a ion myoca dial in a c ion; PCI 5pe cu aneous co ona y in e en ion; STEMI 5ST-segmen ele a ion myoca dial in a c ion; T2D 5 ype 2 diabe es. Hau ala e al P edic ion Models o Heal h Ca e Cos s 139 ( 50.395), accoun ing o 16% o he cos s (P5.001). Those pa ien s who showed highe sco es o dep ession had highe heal h ca e cos s. When he nex 2 anked ma ke s (LDL choles e ol, 50.230; and le en icula ejec ion ac ion [LVEF], 5-0.227, espec i ely) we e added o he model, he p edic i e alue was 24% o he cos s (P5.001). Finally, ha ing all 13 isk ma ke s (including, eg, smoking, sys olic blood p essu e, and diabe es) in he model, hey p edic ed 30% o he cos s (P5.094). Discussion The p esen s udy demons a ed ha selec ed ML ools a e applicable o p edic heal h ca e cos s o all easons in 1- yea ollow-up when assessing he con ibu ion o well- add essed causal and modifiable isk ma ke s o p ognosis o CAD collec ed a baseline in pa ien s wi h a ecen ACS. We ound ha dep ession exp essed as he highe DEPS sco e is he p ima y con ibu ing ac o o heal h ca e cos s in 1-yea ollow-up, ollowed by a highe LDL choles e ol le el and lowe alues o LVEF. These esul s may be use ul o decision-making when planning and ocusing on op imal u iliza ion o heal h ca e esou ces. Addi ionally, ou findings may highligh he po en ial use o sophis ica ed ML da a an- aly ics ools in eal-wo ld clinical se ings when making eco- nomic analyses o suppo decision-making. A baseline, he mos dominan p edic o s o all heal h ca e se ice cos s in 1-yea ollow-up we e ela ed o a highe le el o dep ession, a highe le el o LDL choles e ol, a lowe le el o ejec ion ac ion, a lowe le el o exe cise capaci y, and a lowe le el o heal h- ela ed quali y o li e, accoun ing o abou 25% o he cos s in s able ACS pa ien s ea ed ac- co ding o he cu en guidelines. All hose isk ac o s a e well add essed as impo an causal and modifiable ac o s o he p ognosis o disease. 22 In e es ingly, psychosocial isk ac o s, such as dep ession, ha e shown hei impo ance in a ec ing ca dio ascula p ognosis, ea men adhe ence, quali y o li e, and sudden ca diac dea h. 23,24 I is no able ha in he p esen s udy, he highes le el o co ela ion be- ween heal h ca e cos s and he dep ession sco e migh be conside ed as a mode a e associa ion ( 50.395), since a high le el o co ela ion usually exceeds alues o .0.5 and could be in e p e ed as a s ong associa ion. Howe e , Figu e 1 Rank agg ega ion o he isk ac o s calcula ed by di e en me hods (each column ep esen s a ea u e selec ion me hod). The lowe ank alue (da ke colo in he hea map) deno es he highe impo ance o he isk ac o . CCA 5canonical co ela ion analysis; LDL 5low-densi y lipop o ein; PLSC 5pa ial leas squa es canonical analysis; PLSR 5pa ial leas squa es eg ession; PLSSVD 5pa ial leas squa es based on singula alue decomposi ion; F alue 5 alue om he analysis o a iance. Table 2 Linea eg ession analysis models o p edic ion o heal h ca e cos s Risk ma ke s R co ela ion Model R 2 P alue Dep ession Scale 0.395 1 0.156 .001 LDL choles e ol 0.230 2 0.190 .001 Ejec ion ac ion -0.227 3 0.240 .001 Maximal exe cise capaci y -0.142 4 0.245 .002 Quali y o li e, 15D -0.106 5 0.251 .004 Age 0.092 6 0.273 .004 Sex 0.090 7 0.273 .008 Smoking -0.089 8 0.274 .016 Sys olic blood p essu e -0.070 9 0.276 .026 AUDIT-C o alcohol use -0.050 10 0.285 .035 Diabe es -0.048 11 0.298 .042 HbA1c -0.027 12 0.299 .064 Body mass index -0.022 13 0.300 .094 The models we e defined acco ding o anking analysis o well-add essed causal and modifiable isk ma ke s o p ognosis o co ona y a e y disease a baseline. Model 1 includes op- anking isk ma ke Dep ession Scale. Models om 2 o 13 we e defined by en e ing he second-highes pa ame e (LDL choles e ol) o he model, hen he nex -highes isk ma ke s we e added one by one o he defined models (3, ejec ion ac ion; 4, maximal exe cise capaci y; 5, quali y o li e; 6, age; 7, sex; 8, smoking; 9, sys olic blood p es- su e; 10, AUDIT-C o alcohol use; 11, diabe es; 12, HbA1c; and 13, body mass index). Abb e ia ions as in Table 1. 140 Ca dio ascula Digi al Heal h Jou nal, Vol 4, No 4, Augus 2023 dep essi e symp oms ha e been shown o be s ongly associ- a ed wi h highe le els o s ess, low social suppo , unem- ploymen , low amily income, and unheal hy li es yle such as low physical ac i i y, low ui and ege able in ake, and excessi e sal consump ion in CAD pa ien s. 25 Symp oms o dep ession a e highly p e alen in s able CAD pa ien s, and hei long- e m ajec o ies a e sugges ed o be he single bigges d i e o heal h ca e cos s. 26 The e o e, managemen o dep ession symp oms migh be one o he p ima y ocuses o policymake s and decision make s in planning ea men and esou ces o s able CAD pa ien s. 23 As men ioned abo e, dep ession is a common como bidi y in CAD pa ien s and nume ous po en ial mechanisms ha e been pos ula ed o he ela ionship be ween dep ession and CAD. I has been documen ed ha se e al clinical ac o s can be d i ing dep ession concu en ly and hus may con ound esul s when aiming o in e p e and define a causal isk ac o o dep ession. This kind o analysis may equi e e idence ha educ ion o he isk ac o educes isk. 27,28 The DEPS scale we used in his s udy is a 10-i em sel - epo scale ha assesses he se e i y o dep essi e symp oms. Rega ding he clinically meaning ul alue o he DEPS scale, i has been sugges ed ha a sco e o 10 o highe on he DEPS scale is a use ul cu o o iden i ying clinically significan dep essi e symp oms. 29 Fu he mo e, i is impo an o no e ha a DEPS sco e o 10 o highe should no be used as he sole basis o diagnosing dep ession. A comp ehensi e clinical e alua ion, including a ho ough his o y and physical examina ion, is necessa y o make an accu a e diagnosis and de elop an app op ia e ea - men plan o dep ession. E en hough no in he scope o he p esen s udy, we pe - o med u he analysis using he DEPS sco e o 10 o find ou i he measu es we ha e assessed, including medica ion (p esen ed in Table 1), a e associa ed wi h he DEPS scale. Se en pa ien s had a alue o 10 o highe . The only pa am- e e associa ed wi h he DEPS scale was quali y o li e, as- sessed wi h 15D ques ionnai e ( 5-0.396, P5.001). In he p esen s udy, he selec ed ea u e impo ance me hods showed hei applicabili y o ank well-known isk ma ke s o find he mos p e e ed fi s -o de a ge s o isk ma ke s o con ibu e o heal h ca e cos s. We also used and e alua ed some o he ele an isk ma ke s o he ACS popula ion in he de elopmen p ocess. Fo example, since high-sensi i i y C- eac i e p o ein has been shown o be an independen p edic o o ad e se ca dio ascula e en s among CAD pa ien s, 30 we es ed i i con ibu es o he o de o leading p edic i e isk ma ke s. We ound ha including high-sensi i i y C- eac i e p o ein in he analysis as an ex a isk ma ke will no change he esul s. Fu he - mo e, we assessed i he o de o he leading p edic i e isk ma ke will change i we emo e, one by one, he isk ma ke s anked om 6 o 13. Despi e exclusion o he isk ma ke s om he ea u e impo ance analysis, he o de o he 5 leading ma ke s emained he same. The e o e, we belie e ha in addi ion o selec ed ea u e impo ance anal- ysis ools, he selec ed isk ma ke s included we e ele an and alid o he pe o med analysis. The use o heal h ca e se ices in he p esen s udy was de i ed om hospi al eco ds ins ead o , o example, om pa- ien sel - epo s, he eby elimina ing ecall bias. Secondly, he cha ac e is ics o pa ien s a he baseline we e widely assessed, including clinical s a us, medica ion, and comp ehensi e lab- o a o y analysis. We eel ha hese a e he s eng hs o his s udy. A limi a ion o his pilo s udy is ha he pa ien sample in he EFEX-CARE s udy is small and may be pa ly selec ed, which could limi he gene alizabili y o a b oade popula ion o ACS pa ien s wi h significan como bidi ies. We showed ha a e combining all he 13 s udied ma ke s, only 30% o he cos s could be p edic ed by his model. This could be in e - p e ed as a ela i ely low a e. Howe e , he p op ie a y na u e o economic da a, and he ac ha elemen s o heal h ca e cos s a e coming om di e en en i ies, may a leas pa ly explain ou esul s. Fo example, we we e able o analyze di ec heal h ca e cos s, bu no indi ec cos s such as he ex- penses incu ed om he cessa ion o educ ion o wo k p o- duc i i y. The o he ques ion ha emains open is whe he i would be possible o aise he p edic i e alue by adding mo e a iables; his could be a a ge o u u e esea ch. Addi- ionally, al hough we ca e ully assessed a ious ea u e impo - ance me hods o p o e hei s abili y, he ela i ely low numbe o samples and mul icollinea i y among he isk ac- o s may aise some cau ion in o e all in e p e a ion and gene alizabili y in o e all in e p e a ion o he esul s. Howe - e , he p oposed me hodology is gene ic enough o be applied in any field o se ing o medical and heal h ca e in which isk p ofiles o pa ien s exis and heal h ca e cos s o ce ain pe- iods a e assessed. Because heal h ca e budge s a e limi ed wo ldwide, he e is a c ucial need o s a egies o heal h ca e sys ems ha p o e o be cos -e ec i e. The need o ca e s a egies should a he same ime be low cos and gi e he bes e ec o ca e. How- e e , di ec assessmen o cos s is no easonable in di e en coun ies because o di e ences in social and heal h ca e se - ices na ionally. The e o e, he esul s o he p esen s udy may be use ul o policymake s especially in he Finnish heal h ca e sys em when planning and deciding how limi ed heal h ca e esou ces should be used in he op imal way. Conclusion Ou s udy showed ha dep ession, exp essed as highe dep ession sco es, is he p ima y ac o o ecas ing heal h ca e cos s o all easons, ollowed by LDL choles e ol and LVEF, in 1-yea ollow-up among ACS pa ien s. These e- sul s a e help ul o decision-making when planning op imal u iliza ion o ea men s a egies and esou ces in di e en heal h ca e se ings. Fu he mo e, ou findings confi m he po en ial use o ad anced da a analy ics and ML ools in eal-wo ld clinical se ings. Acknowledgmen s The au ho s would like o hank he EFEX-CARE s udy g oup o hei excellen wo k and assis ance h oughou he EFEX-CARE s udy. This esea ch is ela ed o he Hau ala e al P edic ion Models o Heal h Ca e Cos s 141 hema ic esea ch a ea o Decision Analy ics u ilizing Causal Models and Mul iobjec i e Op imiza ion (DEMO, jyu.fi/ demo) o he Uni e si y o Jy askyla. Funding Sou ces This s udy was pa ly unded by he Academy o Finland, Finland (g an no. 322221). Disclosu es The au ho s ha e no conflic s o disclose. Au ho ship All au ho s a es hey mee he cu en ICMJE c i e ia o au ho ship. Pa ien Consen All pa ien s p o ided w i en in o med consen . E hics S a emen The au ho s designed he s udy and ga he ed and analyzed he da a acco ding o he Helsinki Decla a ion guidelines on human esea ch. The esea ch p o ocol used in his s udy was e iewed and app o ed by he ins i u ional e iew boa d. Re e ences 1. Visse en FLJ, Mach F, Smulde s YM, e al. 2021 ESC Guidelines on ca dio as- cula disease p e en ion in clinical p ac ice. Eu Hea J 2021;42:3227–3337. 2. Leal J, Luengo-Fe nandez R, G ay A, Pe e sen S, Rayne M. Economic bu den o ca dio ascula diseases in he enla ged Eu opean Union. Eu Hea J 2006; 27:1610–1619. 3. Bi ge M, Kaldjian AS, Ro h GA, Mo an AE, Dieleman JL, Bellows BK. Spending on ca dio ascula disease and ca dio ascula isk ac o s in he Uni ed S a es: 1996 o 2016. Ci cula ion 2021;144:271–282. 4. De Smed D, Ko se a K, De Bacque D, e al. Cos -e ec i eness o op imizing p e en ion in pa ien s wi h co ona y hea disease: he EUROASPIRE III heal h economics p ojec . Eu Hea J 2012;33:2865–2872. 5. Mi on C, Donaldson C. Heal h ca e p io i y se ing: p inciples, p ac ice and chal- lenges. Cos E Resou Alloc 2004;2:3. 6. Salo H, Sin onen H. [Economic e alua ion o an immuniza ion p og am]. Duo- decim 2002;118:93–97. 7. Liu Y, Ren H, Fanous H, e al. A machine lea ning model in p edic ing hemody- namically significan co ona y a e y disease: a p ospec i e coho s udy. Ca di- o asc Digi Heal h J 2022;3:112–117. 8. Shimizu M, Suzuki M, Fujii H, Kimu a S, Nishizaki M, Sasano T. Machine lea ning o mic o ol -le el 12-lead elec oca diog am can help dis inguish ako - subo synd ome and acu e an e io myoca dial in a c ion. Ca dio asc Digi Heal h J 2022;3:179–188. 9. Schwalm JD, Di S, She h T, e al. A machine lea ning-based clinical decision sup- po algo i hm o educing unnecessa y co ona y angiog ams. Ca dio asc Digi Heal h J 2022;3:21–30. 10. Kuhn M, Johnson K. Fea u e enginee ing and selec ion: a p ac ical app oach o p edic i e models. CRC P ess; 2019. 11. Hau ala AJ, Ki iniemi AM, Makikallio T, e al. Economic e alua ion o exe cise- based ca diac ehabili a ion in pa ien s wi h a ecen acu e co ona y synd ome. Scand J Med Sci Spo s 2017;27:1395–1403. 12. Salokangas RK, Pou anen O, S enga d E. Sc eening o dep ession in p ima y ca e. De elopmen and alida ion o he Dep ession Scale, a sc eening ins umen o dep ession. Ac a Psychia Scand 1995;92:10–16. 13. Saunde s JB, Aasland OG, Babo TF, de la Fuen e JR, G an M. De elopmen o he Alcohol Use Diso de s Iden ifica ion Tes (AUDIT): WHO collabo a i e p oj- ec on ea ly de ec ion o pe sons wi h ha m ul alcohol consump ion–II. Addic ion 1993;88:791–804. 14. Sin onen H. The 15D ins umen o heal h- ela ed quali y o li e: p ope ies and applica ions. Ann Med 2001;33:328–336. 15. Hujanen T, Mikkola H. Ty€ o e eyshuollon pal elujen kus annus en alueellise e o . Kelan u kimusosas on julkaisuja; 2013. Ne i y€ opape ei a 42. 16. Chand asheka G, Sahin F. A su ey on ea u e selec ion me hods. Compu Elec Eng 2014;40:16–28. 17. Li J, Cheng K, Wang S, e al. Fea u e selec ion: a da a pe spec i e. ACM compu ing su eys (CSUR) 2017;50:1–45. 18. Fodo IK. A su ey o dimension educ ion echniques (No. UCRL-ID-148494). CA (US): Law ence Li e mo e Na ional Lab; 2002. 19. Van De Maa en L, Pos ma E, Van den He ik J. Dimensionali y educ ion: a compa a i e e iew. J Mach Lea n Res 2009;10:66–71. 20. Wegelin JA. A Su ey o Pa ial Leas Squa es (PLS) Me hods, wi h Emphasis on he Two-Block Case; 371. Depa men o S a is ics, Uni e si y o Washing on. 2000. h ps://www.s a .washing on.edu/ esea ch/ epo s/2000/ 371.pd . 21. Vinzi VE, Chin WW, Hensele J, Wang H. Handbook o pa ial leas squa es, Vol 201. Be lin: Sp inge ; 2010. 22. Amb ose i M, Ab eu A, Co a U, e al. Seconda y p e en ion h ough comp e- hensi e ca dio ascula ehabili a ion: om knowledge o implemen a ion. 2020 upda e. A posi ion pape om he Seconda y P e en ion and Rehabili a ion Sec- ion o he Eu opean Associa ion o P e en i e Ca diology. Eu J P e Ca diol 2021;28:460–495. 23. Lah inen M, Ki iniemi AM, Jun ila MJ, Kaa iainen M, Huiku i HV, Tulppo MP. Dep essi e symp oms and isk o sudden ca diac dea h in s able co ona y a e y disease. Am J Ca diol 2018;122:749–755. 24. Pogoso a N, Sane H, Pede sen SS, e al; Ca diac Rehabili a ion Sec ion o he Eu opean Associa ion o Ca dio ascula P e en ion, Rehabili a ion o he Eu- opean Socie y o Ca diology. Psychosocial aspec s in ca diac ehabili a ion: om heo y o p ac ice. A posi ion pape om he Ca diac Rehabili a ion Sec- ion o he Eu opean Associa ion o Ca dio ascula P e en ion and Rehabili a- ion o he Eu opean Socie y o Ca diology. Eu J P e Ca diol 2015; 22:1290–1306. 25. Pogoso a N, Boy so S, De Bacque D, e al. Fac o s associa ed wi h anxie y and dep essi e symp oms in 2775 pa ien s wi h a e ial hype ension and co ona y hea disease: esul s om he COMETA Mul icen e S udy. Glob Hea 2021; 16:73. 26. Palacios J, Khondoke M, Mann A, Tylee A, Ho op M. Dep ession and anxie y symp om ajec o ies in co ona y hea disease: associa ions wi h measu es o disabili y and impac on 3-yea heal h ca e cos s. J Psychosom Res 2018; 104:1–8. 27. Lich man JH, F oeliche ES, Blumen hal JA, e al. Dep ession as a isk ac o o poo p ognosis among pa ien s wi h acu e co ona y synd ome: sys ema ic e iew and ecommenda ions: a scien ific s a emen om he Ame ican Hea Associa- ion. Ci cula ion 2014;129:1350–1369. 28. Ca ney RM, F eedland KE, Mille GE, Ja e AS. Dep ession as a isk ac o o ca diac mo ali y and mo bidi y: a e iew o po en ial mechanisms. J Psychosom Res 2002;53:897–902. 29. Almeida OP, Almeida SA. Sho e sions o he ge ia ic dep ession scale: a s udy o hei alidi y o he diagnosis o a majo dep essi e episode acco ding o ICD- 10 and DSM-IV. In J Ge ia Psychia y 1999;14:858–865. 30. Ka jalainen JJ, Ki iniemi AM, Hau ala AJ, e al. De e minan s and p ognos ic alue o ca dio ascula au onomic unc ion in co ona y a e y disease pa ien s wi h and wi hou ype 2 diabe es. Diabe es Ca e 2014;37:286–294. 142 Ca dio ascula Digi al Heal h Jou nal, Vol 4, No 4, Augus 2023