Using machine learning to predict deterioration of symptoms in COPD patients within a telemonitoring program

Using machine lea ning o p edic de e io a ion o symp oms in COPD pa ien s wi hin a elemoni o ing p og am Ja ie Mo aza1,2, C is óbal Es eban-Aizpi i8, Amaia A ambu u1,2, Ped o Ga cía8, Fe nando Sancho6, Se gio Resino5, Ley e Chasco1,2, F ancisco José Conde7, José An onio Gu ié ez7, Dabi San ano9 & C is óbal Es eban1,2,3,4 COPD exace ba ions ha e a p o ound clinical impac on pa ien s. Accu a ely p edic ing hese e en s could help heal hca e p o essionals ake p oac i e measu es o mi iga e hei impac . Fo o e a decade, elEPOC, a eleheal hca e p og am, has collec ed da a ha can be u ilized o ain machine lea ning models o an icipa e COPD exace ba ions. The objec i e o his s udy is o de elop a machine lea ning model ha , based on a pa ien ’s his o y, p edic s he p obabili y o an exace ba ion e en wi hin he nex 3 days. A e cleaning and ha monizing he di e en subse s o da a, we spli he da a along he empo al axis: one subse o model aining, ano he o model selec ion, and ano he o model e alua ion. We hen ained a g adien ee boos ing app oach as well as neu al ne wo k-based app oaches. A e conduc ing ou analysis, we ound ha he Ca Boos algo i hm yielded he bes esul s, wi h an a ea unde he p ecision- ecall cu e o 0.53 and an a ea unde he ROC cu e o 0.91. Addi ionally, we assessed he signi icance o he inpu a iables and disco e ed ha b ea hing a e, hea a e, and SpO2 we e he mos in o ma i e. The esul ing model can ope a e in a 50% ecall and 50% p ecision egime, which we conside has he po en ial o be use ul in daily p ac ice. Ch onic obs uc i e pulmona y disease (COPD) is a ch onic espi a o y disease ha se es as a pa adigm o ch onic diseases. Wi h a high global p e alence (12.16%)1,2 i ep esen s a signi ican heal h bu den3. Du ing he cou se o he disease pa ien s may expe ience de e io a ion o hei baseline clinical s a us (exace ba ion), occasionally equi ing hospi aliza ion o con ol such wo sening. In COPD, he baseline disease se e i y and hospi aliza ions due o COPD exace ba ions (eCOPD) a e he wo ac o s ha mos signi ican ly impac on di ec cos s4. Fu he mo e, hospi aliza ions also ha e a p o ound clinical impac on pa ien s, leading o a loss o pulmona y unc ion5, de e io a ion o heal h- ela ed quali y o li e in he sho and long e m6,7, heigh ened isk o ca dio ascula e en s8, inc eased mo ali y9, and g ea e p obabili y o sho - e m eadmission10. This cycle o hospi aliza ion and eadmission no only escala es cos s bu also pe pe ua es ad e se ou comes o pa ien s. The cu en si ua ion has spu ed a ious in e en ions aimed a modi ying he sequence o nega i e e en s associa ed wi h eCOPD. These in e en ions in ol e p edic ing he isk o se e e eCOPD (hospi aliza ion)11 o eadmissions10,12, as well as ea ly de ec ion o eCOPD13, pa icula ly in pa ien s wi h a highe likelihood o admission14. This no ion o an icipa ion and he ools ha suppo i ep esen a new ca e pa adigm o ch onic diseases, pa icula ly COPD, which would necessi a e a edesign o he ch onic disease ca e model. In his new scena io, elemoni o ing and Machine Lea ning a e expec ed o play c ucial oles in changing clinical p ac ice. Al hough elemoni o ing has been a con o e sial ool in managing eCOPD13, i may be necessa y o es ablish an adequa e pa ien p o ile o achie e success14. Addi ionally, elemoni o ing can p o ide a con inuous and well-s uc u ed s eam o quali y da a ha can be le e aged wi h Machine Lea ning echniques. These echniques can lea n and unco e ela ionships and pa e ns ha a e no isible o adi ional me hods cu en ly in use. 1Respi a o y Depa men , Hospi al Galdakao-Usansolo, Galdakao, Vizcaya, Spain. 2BioC uces-Bizkaia Heal h Resea ch Ins i u e, Ba acaldo, Spain. 3Heal h Se ices Resea ch on Ch onic Pa ien s Ne wo k (REDISSEC), Mad id, Spain. 4Ch onici y, P ima y Ca e, and Heal h P omo ion Resea ch Ne wo k (RICAPPS), Mad id, Spain. 5Subdi ec o a e o In o ma ion Technology, Osakide za, Bilbao, Spain. 6Depa men o Compu e Science and A i icial In elligence, Uni e si y o Se ille, Se illa, Spain. 7In o ma ion Technology Depa men , Hospi al Galdakao- Usansolo, Galdakao, Vizcaya, Spain. 8Camb ian In elligence SLU, Mad id, Spain. 9Subdi ec o a e o Quali y and In o ma ion Sys ems, Osakide za, Bilbao, Spain. email: [email p o ec ed] OPEN Scien i ic Repo s | (2025) 15:7064 1 | h ps://doi.o g/10.1038/s41598-025-91762-x www.na u e.com/scien i ic epo s A he hospi al o Galdakao, we de eloped a elemoni o ing p og am ( elEPOC)14–16 aimed o moni o COPD pa ien s ha ha e been equen ly admi ed o COPD exace ba ion o he hospi al. The main goal o his p og am is o educe he numbe o admissions o he hospi al, and i s esul s so a ha e been e y sa is ac o y ( he p og am has been sp ead o o he hospi als o he Basque Go e nmen Heal h Depa men ). I has also been shown ha his p og am imp o es se e al aspec s o heal h ca e o pa ien s wi h espec o hose in he con ol g oup14. In his s udy, ou objec i e is o de elop an ea ly wa ning sys em based on Machine Lea ning ( elEPOCML) ha can p edic when a pa ien in he elEPOC p og am is likely o expe ience a ed ala m (i.e., he highes le el o ala m, see14 o mo e de ails). In o he wo ds, we aim o an icipa e eCOPD in ou coho o elemoni o ed pa ien s. Me hods Telemoni o ing da a se The elEPOC da ase is composed o he daily ques ionnai es submi ed by pa ien s on a daily basis. Mo e speci ically, he da ase is composed o he ollowing a iables: • SpO2: Measu ed by a pulse oxime e . • Hea Ra e: Measu ed by a pulse oxime e . • B ea hing Ra e: Manually measu ed espi a o y a e. • Numbe o S eps (p e ious day): Measu ed wi h a pedome e . • Tempe a u e. • Do you ha e a igue? (Yes/No). • Do you ha e mo e a igue han usual? (Yes/No). • Do you ha e a cough? (Yes/No). • Do you ha e mo e coughing han usual? (Yes/No). • Do you ha e spu um? (Yes/No). • Is you spu um amoun inc eased, he same, o dec eased compa ed o usual? (Inc ease/Same/Dec ease). • Wha is he colo o you spu um? (Whi e / G eenish yellow / Wi h blood). • How a e you eeling in gene al? (Be e /Equal/Wo se). • A e you eeling be e , equal, o wo se han usual? (Be e /Equal/Wo se). Based on hese epo s and a se o ules, an ala m sys em wi h h ee le els is es ablished14. The highes le el, which indica es a se e e exace ba ion, is labeled as a ed ala m. The da ase is subdi ided in o h ee subse s as a esul o he di e en so wa e pla o ms used o collec he da a. Table1 shows he cha ac e is ics o each da a subse . Due o majo changes in he da a acquisi ion p ocess and ala m de ini ions, he subse 1 was le ou o his p ojec . In o al, 166 COPD pa ien s om he TelEPOC p og am we e included in his s udy (74.7% men, 15% cu en smoke s). Table2 p esen s baseline cha ac e is ics ob ained om he medical eco ds a ailable in he hospi al o hese pa ien s. These cha ac e is ics a e no pa o he TelEPOC da ase and we e no used o ain he machine lea ning models. Ins ead, hey a e included o p o ide addi ional con ex ega ding he demog aphic, clinical, and physiological p o iles o he s udy popula ion, enhancing he in e p e abili y o ou esul s. Da a cleaning and ha moniza ion The da a cleaning and ha moniza ion s eps consis ed o ixing w ongly labeled ala ms, ha monizing ca ego ical a iable alues ha changed o e ime, emo ing un eliable ime pe iods when ha happened du ing back end mig a ions, emo ing duplica e submissions and emo ing un ealis ic alues. Table3 shows he ange o allowed alues o he di e en a iables. Some imes, pa ien s make mis akes while submi ing hei daily da a. In such cases, hey send a second submission o co ec hei e o s. Du ing he da a cleaning p ocess, we only e ain one submission pe day. We p io i ize he alues submi ed in he mos ecen submission. Addi ionally, a iables and ca ego ical alues ha e been modi ied h oughou he p og am. The e o e, we pe o med necessa y ope a ions o s anda dize hem wi hou comp omising he in o ma ion. In addi ion, we excluded samples whe e he pa ien was al eady in a ed ala m s a e. Howe e , we no iced ha he model achie ed signi ican ly be e sco es when we e ained hese samples in he da ase . This was likely due o he ac ha pa ien s in a ed ala m s a e a e much mo e likely o expe ience ano he ed ala m he ollowing day han hose who a e no . Ne e heless, we belie e ha such p edic ions a e o limi ed alue in daily p ac ice. This is because doc o s a e al eady awa e o he pa ien s’ condi ion, and consecu i e days wi h a ed ala m can be Da a subse 1 Da a subse 2 Da a subse 3 S a ing da e 2010-05-31 2014-09-29 2017-12-15 End da e 2014-11-13 2018-04-17 2021-05-19 Numbe o submissions 72,870 80,303 82,675 % o missing da a 13.44% 12.90% 14.34% Numbe o pa ien s 74 111 117 Table 1. The TelEPOC da ase is di ided in o h ee subse s. Scien i ic Repo s | (2025) 15:7064 2 | h ps://doi.o g/10.1038/s41598-025-91762-x www.na u e.com/scien i ic epo s/ conside ed as pa o he same exace ba ion pe iod. The e o e, we concluded ha i would be mo e app op ia e and ealis ic o exclude hese samples om he model’s aining and e alua ion. A e hese p e-p ocessing s eps, we combined i o c ea e a single da ase . The inal da ase consis ed o 149 pa ien s and 159,719 samples. I co e s a ime ame om 09-29-2014 o 05-19-2021. Model inpu and a ge a iables The a ge a iable ha we chose o p edic is whe he a pa ien will expe ience a ed ala m wi hin he nex h ee days. As pa o he da a p ep ocessing, we compu ed his a ge a iable o each pa ien submission. The model u ilizes he pa ien ’s submi ed da a om he p e ious days o make his p edic ion. We expe imen ed wi h di e en numbe s o p e ious days, which we e e ed o as he inpu window size. A e es ing, we selec ed an inpu window size o 11 days (cu en day plus a his o ical window o 10 days) o ou inal model. This window size yielded he bes esul s among he ones we es ed. Addi ionally, i co esponds o ou clinical obse a ions and he indings epo ed in o he s udies17. Besides, ou model will use he medium e m s a is ics o each pa ien a he ime o each submission. Speci ically, in o de o allow he model o ind bo h pa e ns in he aw da a as well as de ia ion om indi idual baseline alues, we add new a iables wi h mean, median and s anda d de ia ion om he las 4 mon hs p io o each submission o empe a u e, hea a e, numbe o s eps and b ea hing a e a iables. Expe imen al se up We spli he da ase in o h ee subse s along he empo al axis. The i s subse , which con ains 75% o he samples, is he ain se used o ain he model. The second subse is he alida ion se , comp ising 15% o he samples, used o model selec ion. The hi d subse is he holdou es se , also consis ing o 15% o he samples, used o e alua e he selec ed model’s pe o mance on unseen da a. Spli ing he da a along he empo al axis, a he han spli ing by pa ien s, ensu es ha he model is ained on da a p eceding he alida ion and es se s, mimicking eal-wo ld scena ios. I also enables us o assess he model’s abili y o gene alize o new da a by es ing i on he holdou es se , which co e s a ime pe iod he Va iable Min alue Max alue SpO2 70 100 Hea a e 40 160 Daily s eps 0 30,000 B ea h a e 9 50 Tempe a u e 30 42 Table 3. Range o allowed a iables o each a iable. To al 166 Age* 64.7 (5.82) Men 124 (74.7) BMI (kg/m2) 27.7 (5.8) Dyspnea (mMRC) scale* 2 [1–3] Six minu es walking dis ance (m) 395.6 (114.9) Physical ac i i y (s eps/day) 5458.9 (183.2) FVC (%) 78.0 (21.0) FEV1 (%) 45.4 (13.8) FEV1/FVC 45.0 (9.4) DLCO (%) 48.8 (19.4) KCO (%) 65.7 (23.7) Cha lson Index* 3 [2–4] COPD assessmen es (CAT) 15.8 (8.0) Hospi al anxie y-dep ession (HAD) Anxie y¥ 6.2 (4.1) Dep ession¥ 4.9 (4.1) BODE-index¥ 2.67 (1.9) Table 2. Baseline cha ac e is ics o he s udy pa icipan s. Resul s show as mean (s anda d de ia ion). * Resul s show as median [in e qua ile ange]. BMI: body mass index; mMRC: modi ied medical esea ch; FVC: o ced i al capaci y; FEV1: o ced expi a o y olume; DLCO: di usion lung capaci y; KCO: CO ans e coe icien ; BODE index (Body mass index, ai low obs uc ion, dyspnea, exe cise capaci y). Scien i ic Repo s | (2025) 15:7064 3 | h ps://doi.o g/10.1038/s41598-025-91762-x www.na u e.com/scien i ic epo s/ model has no p e iously encoun e ed, hus educing he isk o in o ma ion leakage om he ain se o he es se . Al e na i ely, we could di ide he da ase based on he pa ien axis, esul ing in each spli consis ing o a g oup o pa ien s and hei comple e empo al sequence. This app oach o e s he ad an age o p e en ing he model om lea ning pa ien -speci ic pa e ns, which can enhance i s abili y o gene alize o new pa ien s. Howe e , his se up could po en ially leak pa e ns speci ic o a pa icula momen in ime o he es se . Fu he mo e, i is less ep esen a i e o he eal-wo ld use case han he empo al spli . Addi ionally, a p ospec i e da ase con aining 14,253 samples was collec ed. This ex a da ase was used o conduc a eal-wo ld alida ion o he inal model, assessing i s pe o mance beyond he o iginal ain– alida ion– es spli . Models We ied mul iple machine lea ning models o ind he mos sui able one o he ed ala m p edic ion ask, speci ically g adien ee boos ing (Ca Boos ), eed o wa d neu al ne wo ks, and con olu ional neu al ne wo ks. The selec ion o hese h ee app oaches was mo i a ed by hei complemen a y s eng hs. G adien ee boos ing excels wi h abula da a and p o ides ea u e impo ance ankings, making i pa icula ly aluable o clinical applica ions whe e model in e p e abili y is c ucial. Feed o wa d neu al ne wo ks can cap u e complex non- linea ela ionships and a e adep a p ocessing high-dimensional heal hca e da a. Con olu ional Neu al Ne wo ks, pa icula ly 1D CNNs, a e e ec i e o ime se ies da a, au oma ically ex ac ing ele an ea u es om empo al s uc u es. All h ee me hods ha e p o en ack eco ds in heal hca e applica ions, can handle longi udinal da a e ec i ely, and ha e well-es ablished implemen a ions, making hem sui able choices o his p edic ion p oblem. By employing hese di e se algo i hms, we aimed o explo e di e en app oaches o modeling he complex pa e ns in COPD exace ba ion da a, ul ima ely selec ing he bes -pe o ming model o ou ask. T aining and model selec ion We conduc ed a andom sea ch o hype pa ame e s o all models and chose he one ha pe o med he bes on he alida ion se . We e alua ed he pe o mance o all models based on bo h he A ea unde he ROC cu e (AUROC) and he A ea unde he P ecision-Recall cu e (AUPRC). We selec ed he AUPRC as ou p ima y me ic o model selec ion because i be e e lec s model pe o mance in ou hea ily imbalanced da ase —one in which non-exace ba ions g ea ly ou numbe exace ba ions—and i di ec ly cap u es he ade-o be ween ecall and p ecision o he posi i e (exace ba ion) class. While AUROC is a use ul measu e o o e all disc imina ion, i does no e lec he alse posi i e a e’s impac on model applicabili y in eal-wo ld scena ios. A model wi h high AUROC can s ill gene a e an imp ac ical numbe o alse posi i es in an imbalanced da ase , which can o e whelm clinical eams, causing ala m a igue and unde mining he sys em’s u ili y. In con as , AUPRC ocuses on he balance be ween iden i ying ue posi i es ( ecall) and a oiding alse posi i es (p ecision), which is essen ial o a model o be ac ionable in p ac ice. Resul s Table 4 shows he pe o mance o each model when using an inpu window size o 11 days and selec ing he hype pa ame e s ha p o ided he highes AUPRC in he alida ion se . As can be seen, he ca boos model ou pe o ms he neu al ne wo k app oaches. We also Ran expe imen s wi h o he g adien ee boos ing app oaches, such as Ligh GBM and XGBoos , which p o ided almos iden ical bu sligh ly wo se esul s han ca boos . We conduc ed expe imen s wi h a ying inpu window sizes o obse e how sco es change as we al e he numbe o ecen days ha he model can use o make p edic ions. Table5 displays he ca boos model sco es Inpu window size AUROC AUPRC Cu en day 0.89 0.47 Cu en day + 10 0.91 0.53 Cu en day + 20 0.90 0.52 Table 5. AUROC and AUPRC in he es se o he ca boos model when using di e en inpu window sizes. Highes alues a e in bold. Model AUROC AUPRC Ca Boos 0.91 0.53 Feed o wa d NN 0.89 0.47 Con olu ional NN 0.87 0.43 Table 4. AUROC and AUPRC in he es se p o ided by he bes hype pa ame e con igu a ion ha we ound o each model ype. Highes alues a e in bold. Scien i ic Repo s | (2025) 15:7064 4 | h ps://doi.o g/10.1038/s41598-025-91762-x www.na u e.com/scien i ic epo s/ when u ilizing inpu windows o only he cu en day, he cu en day + 10 days, and he cu en day + 20 days. We obse ed ha his pa ame e appea ed o be mo e c i ical o he AUPRC me ic and we chose o employ he cu en day + 10 days e sion o he emainde o he s udy, since i p o ided he bes esul o bo h AUROC and AUPRC. The de ailed esul s o he ca boos model a e shown in Table 6, whe e “AUPRC andom” con ains he a ea unde he P ecision-Recall cu e ob ained by andom p edic ions ( he AUROC o andom p edic ions is always 0.5).Figu es1, 2 and 3 depic he AUROC and AUPRC cu es o he ain, alida ion, and es se s, espec i ely. Model in e p e abili y To achie e a mo e in e p e able model ha would allow us o examine he impo ance o each a iable, we ained a model on a simpli ied da ase wi hou empo al windows. Speci ically, we excluded he 10-day his o y and he 4-mon h summa y, and ocused solely on a single submission o p edic he p obabili y o a ed ala m wi hin he nex h ee days. The expe imen was conduc ed using he same da a spli s as he p e ious expe imen . Table7 p esen s he sco es ob ained om his expe imen . As we can see, he esul s a e signi ican ly wo se han hose o he p e ious expe imen . This demons a es he impo ance o he 10-day his o y and he 4-mon h summa y, which p o ide aluable in o ma ion o he ed ala m p edic ion ask. Figu es3, 4, 5 and 6 depic he AUROC and AUPRC cu es o his expe imen . To ensu e ha ou model had lea ned meaning ul ela ionships and o iden i y he mos in o ma i e a iables o he ask a hand, we conduc ed a SHAP analysis on he simpli ied da ase . SHAP is a game- heo e ic app oach ha a ibu es he con ibu ion o each ea u e o he ou pu . We can see he esul s o his analysis in Fig.7, which shows how b ea h a e, hea a e and SpO2 a e he h ee mos in o ma i e a iables o p edic ed ala ms wi h he Ca Boos model ained on he simpli ied da ase . P ospec i e analysis and applicabili y A e aining ou bes Ca Boos model, an addi ional da ase was collec ed in o de o pe o m a p ospec i e analysis. Speci ically, we used da a submi ed om 2021-05-20 un il 2022-01-30, wi h a o al numbe o 14,253 samples. The sco es ob ained in his subse we e simila o he p e ious ones. Table8 shows he sco es achie ed and Fig.8 hei co esponding cu es. Fig. 1. AUROC and AUPRC cu es o he bes Ca Boos model on he ain se . AUROC AUPRC AUPRC andom T ain 0.95 0.79 0.09 Valida ion 0.89 0.59 0.09 Tes 0.91 0.53 0.08 Table 6. AUROC and AUPRC sco es o he bes Ca Boos model o each da a subse . AUPRC andom con ains he sco e o andom p edic ions. Scien i ic Repo s | (2025) 15:7064 5 | h ps://doi.o g/10.1038/s41598-025-91762-x www.na u e.com/scien i ic epo s/ AUROC AUPRC AUPRC andom T ain 0.91 0.61 0.08 Valida ion 0.88 0.46 0.10 Tes 0.88 0.41 0.07 Table 7. AUROC and AUPRC sco es o he ca boos model o each simpli ied da a subse . AUPRC andom con ains he sco e o andom p edic ions. . Fig. 3. AUROC and AUPRC cu es o he bes Ca Boos model on he es se . Fig. 2. AUROC and AUPRC cu es o he bes Ca Boos model on he alida ion se . Scien i ic Repo s | (2025) 15:7064 6 | h ps://doi.o g/10.1038/s41598-025-91762-x www.na u e.com/scien i ic epo s/ One po en ial applica ion o his sys em in he eal wo ld is o pe o m a ho ough assessmen on pa ien s who, while no expe iencing a ed ala m, ha e he highes isk sco es acco ding o ou model. To e alua e his app oach using ou p ospec i e da ase , we iden i ied 63 days du ing he p ospec i e pe iod when a leas 30 pa ien s had comple e in o ma ion o model e alua ion. Speci ically, hese we e days when no pa ien cu en ly had a ed ala m, he a ge label could be compu ed o he nex h ee days, and he pa ien s had no missing da a in he p e ious 10 days. We hen examined how many ed ala ms could ha e been an icipa ed by selec ing he “n” pa ien s wi h he highes sco e p edic ed by ou model on hese 63 days. To compa e he e ec i eness o his app oach, we also examined he numbe o ed ala ms ha would ha e been an icipa ed by andomly selec ing “n” pa ien s ins ead. The esul s o his expe imen a e p esen ed in Table9. We obse ed ha i we had selec ed he i e pa ien s wi h he highes p edic ed sco es acco ding o ou Ca Boos model, we would ha e iden i ied 107 pa ien s who la e expe ienced a ed ala m. In con as , i we had andomly selec ed i e pa ien s each day, Fig. 5. AUROC and AUPRC cu es o he Ca Boos model on he simpli ied alida ion se . Fig. 4. AUROC and AUPRC cu es o he Ca Boos model ained on he simpli ied ain se . Scien i ic Repo s | (2025) 15:7064 7 | h ps://doi.o g/10.1038/s41598-025-91762-x www.na u e.com/scien i ic epo s/ we would ha e iden i ied only 23.5 ± 1.30 pa ien s who la e expe ienced a ed ala m. These indings sugges ha ou model can e ec i ely iden i y pa ien s a highe isk o expe iencing a ed ala m, and ha his app oach could be a aluable ool o heal hca e p o ide s seeking o p io i ize pa ien ca e. This s udy was conduc ed in acco dance wi h he Decla a ion o Helsinki and ele an guidelines and egula ions. E hical app o al was ob ained om he Comi é de É ica de la In es igación de Euskadi (Euskadi Resea ch E hics Commi ee), wi h p ojec ID PI2019038. In o med consen was ob ained om all pa icipan s p io o hei inclusion in he s udy. Discussion The e is no de ini i e, uni e sally accep ed de ini ion o eCOPD (GOLD23). The de ini ion is inhe en ly subjec i e, p ima ily due o pa ien inpu da a bu also in luenced by heal hca e p o ide s, esul ing in a lack o objec i e da a o suppo he diagnosis o eCOPD. This is why some expe s e e o he diagnosis o eCOPD as an exclusion diagnosis. In ou elEPOC p og am, we we e able o educe he hospi aliza ion a e o eCOPD by 40% compa ed o a hospi al con ol (14). Howe e , despi e his imp o emen , some pa ien s s ill equi ed hospi aliza ion. Consequen ly, ou nex s ep was o a emp o an icipa e he de e io a ion o he pa ien s’ condi ion. In his pape , we aimed o iden i y he clinical de e io a ion o symp oms (a “ ed ala m”), which we used as a su oga e ma ke o exace ba ion. I is impo an o no e ha a “ ed ala m” does no necessa ily indica e an exace ba ion; a he , i se es as an indica o o po en ial de e io a ion and p omp s heal hca e p o ide s o ake app op ia e ac ions wi hin he elEPOC p og am. We will cen e he discussion on a icles ocused on p edic ing COPD pa ien de e io a ion using solely elemedicine da a. In18, he au ho s used a elemoni o ing sys em o collec pulse a e, oxygen sa u a ion, and b ea hing a e as well as symp oms such as ches igh ness, b ea hlessness and spu um o 110 COPD pa ien s moni o ed o 1 yea . They iden i ied s able and p od omal s a es and buil a model o classi y pa ien s’ s a us in o one o he wo s a es. Thei exace ba ion e en s a e de ined as changes in medica ion, and he e o e hey pu sue a di e en a ge ha he one p esen ed in his wo k. They achie e an AUROC o 0.682. In19, he au ho s use 68,139 sel epo s om 2374 pa ien s o de ine COPD exace ba ion e en s based on he epo ed symp oms, and hen hey build a model o an icipa e e en s ha would happen wi hin he nex h ee days. They didn’ ha e access o measu emen s such as SpO2 o empe a u e. They achie ed an AUROC o 0.727. In20, he au ho s a gue ha using classic algo i hms o es ablish eCOPD e en s leads o oo many alse- posi i e ala ms, ende ing hem imp ac ical o daily use. Ins ead, hey ocus on o ecas ing wo a ge s wi hin he nex 24h: hospi aliza ion and ini ia ion o o al s e oids. To achie e his, he au ho s u ilized 363 days o elemoni o ing da a om 135 pa ien s. Thei app oach esul ed in an AUROC o 0.74 o hospi aliza ion p edic ion, co esponding o a 40% alse-posi i e a e o an 80% ecall a e. Fo an icipa ing he ini ia ion o s e oids, hei model esul ed in an AUROC o 0.765. The au ho s demons a e how hei Machine Lea ning app oach p oduces be e esul s o hese asks han he s anda d algo i hms used o de ine exace ba ions. As an example o wha can be achie ed by including clinical ma ke s, in21, he au ho s de eloped a model o classi y pa ien s as being in a mild s a e o a se e e condi ion. Thei inpu da a included 24 ea u es, such as Neu ophil coun and Eosinophil coun om blood es analysis. They expe imen ed wi h se e al Machine Lea ning models and, simila o ou s udy, ound ha he G adien T ee Boos ing app oach pe o med he bes , Fig. 6. AUROC and AUPRC cu es o he Ca Boos model on he simpli ied es se . Scien i ic Repo s | (2025) 15:7064 8 | h ps://doi.o g/10.1038/s41598-025-91762-x www.na u e.com/scien i ic epo s/ wi h an AUROC o 0.963. In e es ingly, hey also disco e ed ha a nai e ensemble o hei models imp o ed he bes model’s sco e, esul ing in an AUROC o 0.968. Addi ionally, among hei 24 a iables, he au ho s epo ed ha he second mos in o ma i e ea u e o hei ensemble model was he weigh o he pa ien s, which is a a iable ha could po en ially be added o ou model as an inpu . A e e iewing hese s udies, we disco e ed ha ou da ase con ains a la ge numbe o da a poin s compa ed o he o he s. This indica es ha he quali y and size o ou da ase played a signi ican ole in achie ing he bes AUROC sco e (0.91) among he p esen ed a icles. Howe e , as poin ed ou in20, a high alse-posi i e a e is he main ac o ha limi s he p ac icali y o hese models in daily use. The p ecision o posi i e p edic i e alue is he me ic ha should be epo ed o assess he numbe o alse ala ms gene a ed in compa ison o he ue ala ms. The e o e, we belie e ha his ype o wo k should epo he AUPRC, which e lec s he model’s p ecision and ecall pe o mance. I is wo h no ing ha a andom model will always achie e an AUROC sco e o 0.5, and he sco e emains he same ega dless o he posi i e o nega i e a io. Howe e , a andom model’s AUPRC sco e is equal o he p e alence o posi i e labels. Consequen ly, he lowe he p e alence, he ha de i is o achie e a high AUPRC sco e, making i di icul o use his me ic o compa e models ained on da ase s wi h di e en a ios o posi i e e en s. None o he ela ed wo ks we e iewed epo his me ic. In ou case, AUROC AUPRC AUPRC andom P ospec i e da ase 0.89 0.56 0.08 Table 8. AUROC and AUPRC sco es o he bes ca boos model on he p ospec i e se . Fig. 7. SHAP sco es o he Ca Boos model on he simpli ied es se . Each poin ep esen s an ins ance’s SHAP alue o a ea u e. The x-axis shows he SHAP alue (impac on p edic ion), while he colo g adien (blue o ed) ep esen s he ea u e’s ac ual alue (low o high). Scien i ic Repo s | (2025) 15:7064 9 | h ps://doi.o g/10.1038/s41598-025-91762-x www.na u e.com/scien i ic epo s/