Living in Stroke Patients and Validating their Performance

In e na ional Jou nal o En i onmen al Resea ch and Public Heal h A icle Measu ing Ac i i ies o Daily Li ing in S oke Pa ien s wi h Mo ion Machine Lea ning Algo i hms: A Pilo S udy Pin-Wei Chen 1,2,†, Na han A. Baune 1,2,† , Igo Zwi 3,4, Jiayu Wang 3, Vic o ia Swamidass 1 and Alex W.K. Wong 2,3,5,6,*   Ci a ion: Chen, P.-W.; Baune, N.A.; Zwi , I.; Wang, J.; Swamidass, V.; Wong, A.W.K. Measu ing Ac i i ies o Daily Li ing in S oke Pa ien s wi h Mo ion Machine Lea ning Algo i hms: A Pilo S udy. In . J. En i on. Res. Public Heal h 2021,18, 1634. h ps://doi.o g/10.3390/ ije ph18041634 Academic Edi o s: Michael L. Jones, F ank De uy e and John Mo is Recei ed: 30 Decembe 2020 Accep ed: 5 Feb ua y 2021 Published: 9 Feb ua y 2021 Publishe ’s No e: MDPI s ays neu al wi h ega d o ju isdic ional claims in published maps and ins i u ional a il- ia ions. Copy igh : © 2021 by he au ho s. Licensee MDPI, Basel, Swi ze land. This a icle is an open access a icle dis ibu ed unde he e ms and condi ions o he C ea i e Commons A ibu ion (CC BY) license (h ps:// c ea i ecommons.o g/licenses/by/ 4.0/). 1Pla o mSTL, S . Louis, MO 63110, USA; benny[email p o ec ed] (P.-W.C.); [email p o ec ed] (N.A.B.); [email p o ec ed] (V.S.) 2P og am in Occupa ional The apy, Washing on Uni e si y School o Medicine, S . Louis, MO 63108, USA 3Depa men o Psychia y, Washing on Uni e si y School o Medicine, S . Louis, MO 63110, USA; [email p o ec ed] (I.Z.); [email p o ec ed] (J.W.) 4Depa men o Compu e Science and A i icial In elligence, Uni e si y o G anada, 18010 G anada, Spain 5Depa men o Neu ology, Washing on Uni e si y School o Medicine, S . Louis, MO 63110, USA 6Cen e o Rehabili a ion Ou comes Resea ch, Shi ley Ryan Abili yLab, Chicago, IL 60611, USA *Co espondence: [email p o ec ed] † These au ho s a e co- i s au ho s. Abs ac : Measu ing ac i i ies o daily li ing (ADLs) using wea able echnologies may o e highe p ecision and g anula i y han he cu en clinical assessmen s o pa ien s a e s oke. This s udy aimed o de elop and de e mine he accu acy o de ec ing di e en ADLs using machine-lea ning (ML) algo i hms and wea able senso s. Ele en pos -s oke pa ien s pa icipa ed in his pilo s udy a an ADL Simula ion Lab ac oss wo s udy isi s. We collec ed blocks o epea ed ac i i y (“a omic” ac i i y) pe o mance da a o ain ou ML algo i hms du ing one isi . We e alua ed ou ML algo i hms using independen semi-na u alis ic ac i i y da a collec ed a a sepa a e session. We es ed Decision T ee, Random Fo es , Suppo Vec o Machine (SVM), and eX eme G adien Boos ing (XGBoos ) o model de elopmen . XGBoos was he bes classi ica ion model. We achie ed 82% accu acy based on en ADL asks. Wi h a model including se en asks, accu acy imp o ed o 90%. ADL asks included chopping ood, acuuming, sweeping, sp eading jam o bu e , olding laund y, ea ing, b ushing ee h, aking o /pu ing on a shi , wiping a cupboa d, and bu oning a shi . Resul s p o ide p elimina y e idence ha ADL unc ioning can be p edic ed wi h adequa e accu acy using wea able senso s and ML. The use o ex e nal alida ion (independen aining and es ing da a se s) and semi-na u alis ic es ing da a is a majo s eng h o he s udy and a s ep close o he long- e m goal o ADL moni o ing in eal-wo ld se ings. Fu he in es iga ion is needed o imp o e he ADL p edic ion accu acy, inc ease he numbe o asks moni o ed, and es he model ou side o a labo a o y se ing. Keywo ds: ac i i ies o daily li ing; s oke; ehabili a ion; elemedicine; emo e sensing echnology; machine lea ning; wea able elec onic de ices 1. In oduc ion S oke is a leading cause o long- e m disabili y, wi h nea ly 800,000 adul s in he U.S. expe iencing a s oke annually [ 1 ]. Impai men s in senso imo o unc ion, as a e common ollowing s oke, nega i ely impac pe o mance in ac i i ies o daily li ing (ADLs) [ 2 ]. Almos hal o he s oke su i o s expe ience limi a ions in ADLs [ 3 ]. These limi a ions a e a key conce n among clinicians and pa ien s. Resea ch has shown s ong ela ionships be ween pe o mance in ADLs and pa ien s’ quali y o li e [ 4 ] and he isk o e-hospi aliza ion [ 5 , 6 ] a e s oke. ADLs a e complex senso imo o ac i i ies in ol ing dynamic spa ial- empo al coo dina ion o ou limbs and unk. De eloping e ec i e ools o moni o ing ADLs and complex bodily mo emen ollowing s oke could p o ide a weal h o clinically ele an in o ma ion use ul o ailo ing he apy pos -s oke. In . J. En i on. Res. Public Heal h 2021,18, 1634. h ps://doi.o g/10.3390/ije ph18041634 h ps://www.mdpi.com/jou nal/ije ph In . J. En i on. Res. Public Heal h 2021,18, 1634 2 o 16 Howe e , clinical assessmen o ADLs in s oke is limi ed o sel - epo ed o clinician- a ed scales, such as he Func ional Independence Measu e o he Ba hel Index. While hese measu es ha e been clinically alida ed, hey a e mos ly e ospec i e and suscep ible o epo ing bias and e o [ 7 – 9 ]. An ea lie s udy in ol ed elde ly pa ien s and hei ami- lies o nu ses comple ing he Law on Pe sonal Sel Main enance scale and he Ins umen al ADL (IADL) scale. They ound ha pa ien s consis en ly a ed hemsel es less disabled on bo h measu es han o he a e s [ 10 ]. Ano he s udy ound simila esul s in which pa ien s epo ed less disabili y in ADL scales han amily a e s. The end was held when com- pa ing pa ien sel - epo s o esea che obse a ions [ 11 ]. These indings indica ed ha pa ien s migh ha e ouble no icing hei disabili y o unde play hei disabili y. Fu he - mo e, because hese measu es a e a ed on an o dinal scale, hey may lack he esolu ion o de ec sub le changes, limi ing he abili y o moni o a pa ien ’s eco e y p ecisely. Due o pa ien bu den, cos s, o con ex ual ac o s, clinicians in equen ly implemen clinical measu es, so his app oach is inadequa e o cha ac e izing he day- o-day a iabili y o an indi idual’s unc ion [12]. Fo una ely, he las decade has seen massi e g ow h in he capaci y o collec an a ay o physiologic da a o moni o ing he heal h and unc ioning o an indi idual ia senso s, such as hose commonly ound in sma phones o wea able de ices [ 13 ]. Al hough wea able echnologies ha e enabled au oma ic and con inuous measu emen o me ics use ul o p edic i e medicine, such as ene gy expendi u e o pedome e s, cu en u iliza ion o wea able da a p o ides an incomple e depic ion o an indi idual’s pe o mance o ADLs, especially o aging and disabled popula ions. Mobili y is o en limi ed in hese popula ions, educing he u ili y o s ep-coun e s, hea a e measu es, and o he c ude me ics o accu a ely measu e an indi idual’s capaci y o pa icipa e in meaning ul ADLs. Mo eo e , me ics de i ed om wea able senso s (e.g., accele a ion o angula eloci y) do no ha e di ec clinical meaning and a e, he e o e, di icul o use o clinical decision making. I is becoming impe a i e o de elop be e echnology ha p o ides an objec i e and clinically- alued measu emen o an indi idual’s ADLs. Human ac i i y ecogni ion has ga ne ed in ense esea ch in e es o e he pas ew decades and can be gene ally sepa a ed in o ine ia-based de ec ion and ideo-based de ec ion app oaches. On he one hand, ideo-based ac i i y de ec ion has been g ow- ing emendously and has shown p omising esul s (please e e o hese e iews o de ails [14–16] ). Howe e , ideo-based ac i i y de ec ion aises p i acy conce ns due o po en ial b each o da a o da a misuse [ 17 ]. On he o he hand, ine ia-based ac i - i y ecogni ion is less in usi e, and has shown signi ican imp o emen s o e ime [ 17 ]. Ine ia-based ac i i y ecogni ion sys ems u ilize IMU’s (ine ial measu emen uni s con- sis ing o an accele ome e and gy oscope) o measu e body kinema ics di ec ly. These di ec measu es may p o ide addi ional in o ma ion o ecognizing pos -s oke ac i i ies. This echnology would allow esea che s and clinicians o moni o ehabili a ion p og ess, alida e ea men e icacy, and de ec unc ional decline ea ly. Such adical change in mea- su emen may u he heigh en ou abili y o ack eal-wo ld ou comes in ehabili a ion and p o ide accu a e measu es o decen alized ( ully emo e) clinical ials in he u u e. Al hough he ine ia-based ac i i y ecogni ion app oach has ecei ed inc easing a en ion [ 18 – 21 ], p io esea ch was based on a la ge eposi o y da abase o able-bodied indi iduals. In addi ion, p e ious esea ch did no di ec ly eco d mobili y-impai ed indi iduals (e.g., indi iduals who expe ienced a s oke) unde a clinical se ing o machine lea ning (ML) pu poses. In addi ion, mos s udies ha e pe o med in e nal alida ion, whe e a po ion o he same da a se is used o es a model c ea ed om he emaining da a. These a e ba ie s o he e en ual applica ion o ADL moni o ing in eal-wo ld se ings. The e o e, his s udy aimed o de elop a no el p edic ion model based on ML algo i hms and o de e mine he accu acy o de ec ing di e en ADLs pe o med by s oke su i o s using wea able senso s. We conduc ed his s udy in a simula ion li ing oom and ki chen. Las ly, we collec ed independen aining and es ing da a o pe o m ex e nal alida ion, which mo e closely imi a es eal-wo ld p edic ion condi ions. In . J. En i on. Res. Public Heal h 2021,18, 1634 3 o 16 2. Ma e ials and Me hods 2.1. Pa icipan s Pa icipan s we e communi y-dwelling adul s wi h s oke. We ec ui ed 11 s oke su i o s om he s oke egis y a Washing on Uni e si y School o Medicine, a da abase o indi iduals who consen ed o u u e esea ch pa icipa ion a he ime o hei s oke hospi aliza ion. Inclusion c i e ia included: (1) age 18+; (2) English-speaking; and (3) mild s oke as de ined by baseline Na ional Ins i u es o Heal h S oke Scale (NIHSS) sco e om 0 o 5. We chose o s udy pa ien s wi h mild s oke because mo o unc ion and abili y o accomplish basic ADLs a e only minimally a ec ed, inc easing he chance ha he p o ocol will be ully comple ed in an app op ia e ime and allowing us o adhe e o he p ojec budge . Exclusion c i e ia: (1) p e ious neu ologic o neu opsychia ic diso de (e.g., demen ia, schizoph enia) ha makes in e p e a ion o he sel - a ed scales di icul ; (2) Sho Blessed Tes sco e >8 (indica ing signi ican cogni i e impai men ); (3) his o y o mode a e disabili y p io o s oke (p e-mo bid Rankin Scale sco e <3); (4) ision ha is poo e han 20/100 (as de e mined by he Ligh house Nea Visual Acui y Tes ); (5) Ap axia sc een o Tulia <9; and (6) e idence o se e e aphasia (NIHSS aphasia i em >2). Eligible pa icipan s pa icipa ed in ou s udy a an ADL simula ion lab in he P og am in Occupa ional The apy ac oss wo s udy isi s. All pa icipan s we e a leas six mon hs pos hei s oke inciden , meaning hei unc ional eco e y had la gely pla eaued. 2.2. P ocedu es Du ing each isi , we i ed pa icipan s wi h i e ine ial measu emen uni s (IMUs) o collec accele ome e and gy oscope da a: one on each w is , one on each o hei uppe a ms, and one on hei hip. The IMU used in his s udy we e Apple Wa ch Se ies 3 (Apple Inc., Cupe ino, CA, USA). We de eloped he sys em using he Apple Wa ches due o hei a o dabili y, comme cial a ailabili y, and abili y o ans e mo ion da a wi elessly o e use s WiFi ne wo k. We u ilized wo s udy isi s o collec independen aining ( isi wo) and es ing ( isi one) da a se s. Du ing isi one, we used a me hod simila o ha in oduced by Bao and In ille [ 18 ] o cap u ing na u alis ic beha io s. Pa icipan s engaged in a se ies o ADL asks ollowing a s anda dized sc ip o he examine o p o ide minimal guidance. The examine in o med pa icipan s wha asks hey would pe o m nex bu did no commen on how o achie e hem. Fo example, he examine would ask pa icipan s o cook s i - y ollowing a ecipe (which would equi e ga he ing ing edien s, chopping, cooking), hen o ea a se ing (which would equi e pla ing he ood, e ie ing ea ing u ensils, and si ing down o ea ). Pa icipan s we e allowed o shi be ween ac i i ies na u ally and pe o m ac ions like chopping and cooking as hey p e e ed. The expe imen e aimed o p o ide as li le guidance as possible, gi ing cues o he nex s eps and answe ing ques ions; o he wise, pa icipan s ope a ed p ima ily independen ly. This ca e was aken o mos closely mimic eal-wo ld scena ios o ADL p edic ion. Pa icipan s pe o med hese semi-na u alis ic ac i i ies o eco d es ing da a o e alua ing he ML algo i hms. We used su eillance came as o eco d and label all ac i i ies om isi one and la e e iew o possible e o s. Du ing isi wo, pa icipan s pe o med ou minu es o each a omic ac i i y (a o al o 19 a omic ac i i ies; Table 1). An a omic ac i i y is a simple mo emen in ol ed in ADL asks (e.g., s i ing a pan o chopping ege ables a e wo a omic ac i i ies in ol ed in cooking). A omic ac i i ies equi e pa icipan s o con inuously pe o m speci ic mo e- men s wi h expe imen e guidance, which con as s wi h semi-na u alis ic ac i i ies whe e we p o ided pa icipan s wi h a high-le el goal (e.g., cook pancakes) bu did no gi e explici pe o mance ins uc ions. Pa icipan s pe o med a omic ac i i ies epea edly o ou minu es; o example, he pa icipan would s i pancake ba e o ou minu es, place/ e ie e spices on/ om a shel o ou minu es, o old laund y o ou minu es. We collec ed a omic ac i i y pe o mance da a o ain ou ML algo i hms. The expe imen e eco ded he iming o a omic ac i i y pe o mance using in-house so wa e (hence o h, ac i i y labeling so wa e) on a se en h-gene a ion iPad. While ou minu es o a omic In . J. En i on. Res. Public Heal h 2021,18, 1634 4 o 16 ac i i y da a we e collec ed pe ac i i y, pa icipan s could pe o m ac i i ies in mul iple blocks o educe bo edom o a igue due o he homogenous epea ed mo ion. Pa icipan s pe o med he semi-na u alis ic ac i i ies i s ( isi one) be o e hey comple ed he a omic ac i i ies ( isi wo) o educe possible bias. Speci ically, we a oided pa icipan s pe o m- ing a omic ac i i ies i s because he homogenous epea ed mo ion migh in luence he way pa icipan s pe o med he semi-na u alis ic ac i i ies. Table 1. A omic ac i i ies wi h ca ego ies label be ween ac i i ies o daily li ing (ADL) o ins umen- al ac i i ies o daily li ing (IADL). A omic Ac i i ies ADL/IADL B ush Tee h ADL Mixing Powde s IADL Sp eading Bu e /Jam IADL Ea ing wi h Hands ADL Ea ing wi h U ensils ADL Bu oning shi /coa ADL Pu and ake o he coa ADL Mo ing i ems ho izon ally ADL Reaching Up ADL Reaching Down ADL Washing Dishes IADL Chopping IADL Pan S i ing IADL Se e on a pla e IADL Sweeping IADL Vacuuming IADL Wiping ho izon al su ace IADL Wiping e ical su ace IADL Folding clo hes IADL P io o his s udy, we es ed on heal hy adul s and achie ed adequa e accu acy when collec ing only wo minu es o a omic ac i i y da a pe ac i i y. Ne e heless, we collec ed ou minu es pe a omic ac i i y in his s udy o inc ease he amoun o aining da a. We an icipa ed his amoun o da a o be bene icial due o po en ial a iabili y be ween pos -s oke pa icipan s. We could ensu e adequa e da a emained o each ac i i y in any case ha segmen s o da a had o be emo ed due o e o s (e.g., echnical di icul ies wi h IMU eco ding). The ou -minu e was chosen as he uppe -limi as a p ecau ion o a oid pa icipan o e -exe ion. We could u he ensu e da a we e collec ed o all ac i i ies wi hin he allo ed h ee-hou s udy sessions. The s udy sessions we e all acili a ed by a licensed occupa ional he apis ained o moni o hese ac i i ies and ensu e pa ien sa e y. We chose he 19 ac i i ies (Table 1) based on se e al c i e ia. We de e mined he inal ADL and IADL lis h ough consul a ion wi h he apis s and physicians specialized in s oke. Ac i i ies chosen we e: (1) commonly used o assess pos -s oke unc ional ac i i y, (2) amenable o pe o mance in ou simula ed li ing en i onmen , and (3) easible o mild s oke pa ien s o pe o m du ing he allo ed h ee-hou es ing sessions. This s udy ecei ed e hics app o al om he ins i u ional e iew boa d a Washing on Uni e si y. All pa icipan s p o ided w i en consen and ecei ed an hono a ium o acknowledge hei esea ch con ibu ion. In . J. En i on. Res. Public Heal h 2021,18, 1634 5 o 16 2.3. Da a Analysis We used desc ip i e s a is ics o cha ac e ize he s udy sample. We combined ini ial da a om i e de ices o each pa icipan and assigned ac i i y labels. These p edic o a iables included he ansla ional and o a ional accele a ion along h ee axes (x, y, z) ac oss each o he i e senso s, esul ing in 30 a iables (Table A1 o Appendix A ). We used h ee-second epochs o summa ize he aw da a in o ea u e space wi h ime-domain (e.g., mean, s anda d de ia ion, au oco ela ion, and slope) and equency-domain ea u es (see Figu e A1 o Appendix A o one a iable as an exam- ple). This echnique p o ided s a iona i y o he ime se ies, allowing each ac i i y o be analyzed as a s a iona y s ochas ic p ocess in model de elopmen [ 22 ]. Da a we e p ep ocessed, including slicing da a wi h an op imal window size o a oid edundancies, inpu a iables we e no malized and hen comp essed in o 6012 da a poin s. We chose h ee-second epochs based on ou p e ious es ing in heal hy adul s, whe e h ee-second was he op imal epoch leng h while es ing i e epoch leng hs (in ege s 1 h ough 5). Mo ion da a om he semi-na u alis ic ac i i y session ( isi 1) was labeled using he da a om ou ac i i y labeling so wa e. We used da a om isi 2 (a omic ac i i ies) o he aining, whe eas da a om isi 1 (semi-na u alis ic ac i i ies) o he es ing. We implemen ed -Dis ibu ed S ochas ic Neighbo Embedding ( -SNE) analysis and clus e analysis o analyze he samples’ dis ibu ion in e ms o he i s and second isi s. The classes/ asks’ sizes o ac i i ies subs an ially a y among he obse ed subjec s, so we ebalanced he aining da a se s o p e en o e - eaching he classi ie only o p edic he majo (nega i e) classes. We accomplished his using bo h he Syn he ic Mino i y O e sampling Technique (SMOTE), a boo s apping algo i hm ha p o ides mo e da a poin s o he smalle class based on a iable dis ibu ions o ha class [ 23 – 25 ] and he Ex ended Nea es Neighbo (ENN) algo i hm ha downs size he la ge class [ 25 ] o he aining da a. I should be no ed ha bo h SMOTE and ENN we e only used on he aining da a a e he da a we e spli in o he aining and alida ion da a se s. Mo eo e , since lea ning he hype pa ame e s in ol ed he 70–30% hold-ou alida ion, SMOTE and ENN was used only in 70% hold-ou . We also used he Shapley Addi i e exPlana ions (SHAP) o iden i y he con ibu ion o he indi idual IMU ea u es in he model p edic ions [ 26 , 27 ]. 2.4. Classi ied Model De elopmen We achie ed he p ima y classi ica ion model wi h he g adien boos ing model (GBM), XGBoos [ 28 ]. This me hod has been shown o ha e imp o ed supe ised classi ie model o maliza ion, compa ed o o he GBMs o andom o es algo i hms, and be e e iciency and con ols o o e i ing [29]. Hype pa ame e s suppo ed by he XGBoos package we e ine- uned using he T ee Pa zen Es ima o (TPE) Bayesian op imiza ion algo i hm [ 30 ] o ind he bes combina ion o pa ame e s based on he 70–30% hold-ou alida ion. The i ness loss unc ion was de ined by he AUC [ 29 ]. Da a we e andomly pa i ioned in o wo se s based on he hold-ou sampling pa i ion used o he in e nal aining alida ion es wi h unbalanced classes o a omic ac i i ies. The e was no need o o he sampling me hods a he han hold-ou because he ex e nal semi-na u alis ic sample was comple ely disjoin , including a iabili y in e ms o imes o asks, o de , comple eness, and he p esence o he aining asks. The ex e nal independen alida ion da ase based on semi-na u alis ic ac i i ies was no seen a any aining session o da a p ocessing. Tasks we e selec ed by ollowing a s anda d s epwise eg ession using XGBoos , whe e models we e i ed based on he choice o p edic i e a iables ca ied ou by an au oma ic p ocedu e. In each s ep, a a iable is conside ed o addi ion o o sub ac ion om he se o explana o y a iables based on some p e-speci ied c i e ion. The s op i e a ion c i e ia was a pe o mance o 90% ±2%. All sc ip s we e w i en in py hon, and he XGBoos package and he o he classi ie s we e used wi h a py hon sklea n API (h ps://sciki -lea n.o g (accessed on 9 Decembe 2020)) [ 28 , 31 ]. Fo pa ame e ine- uning, we used he hype op package (h ps://sciki - In . J. En i on. Res. Public Heal h 2021,18, 1634 6 o 16 lea n.o g/s able/modules/g id_sea ch.h ml (accessed on 9 Decembe 2020)). We used he imbalance package o apply SMOTE and ENN. We used Sklea n o all da a pa i ions and he calcula ion o he me ics. We selec ed he inal se o ea u es, as es ima ed abo e, and hen es ed mul iple ML models: Decision T ee, Random Fo es , SVM, and XGBoos o de e mine he op imal classi- ica ion model. De aul pa ame e s o each o hese me hods we e lea ned as desc ibed abo e (sklea n-sciki ) by op imizing he AUC me ic. We assessed model pe o mance using accu acy, ecall (sensi i i y), p ecision (posi i e p edic i e alue), and ROC ( ecei e ope a ing cha ac e is ic) cu e (AUC) me ics [32]. The loss unc ion was based on AUC. 3. Resul s 3.1. Cha ac e is ics o S udy Pa icipan s Table 2p o ides an o e iew o pa icipan demog aphic cha ac e is ics. The majo i y o pa icipan s we e males (72%), igh -handed (91%), and expe ienced an ischemic s oke (100%). The a e age age was 60 yea s old. 55% (n= 6) pa icipan s expe ienced a igh hemisphe ic s oke, 36% (n= 4) expe ienced a le hemisphe e s oke, and one pa icipan did no ha e in o ma ion on he side o s oke. The a e age ime since he s oke inciden was 2.76 yea s (SD = 1.73). Table 2. Demog aphic cha ac e is ics o he pa icipan s (n= 11). Cha ac e is ics Mean (SD)/Coun Age 59.64 (9.04) Yea s since s oke 2.76 (1.73) NIHSS 1To al Sco e 1.45 (1.29) Sex 8 males; 3 emales Race 6 Caucasian; 4 A ican Ame ican; 1 unknown Educa ion 4 high school g adua es; 5 some college; 2 bachelo ’s deg ee Handedness 10 igh ; 1 le S oke Type 11 ischemic; 0 hemo hagic S oke Side 6 igh -hemisphe ic; 4 le -hemisphe ic; 1 unknown 1NIHSS = Na ional Ins i u es o Heal h S oke Scale. 3.2. Da a Ex ac ion and Model De elopmen The o al du a ion o he obse ed da a was 153,159 s. We di ided da a in o h ee- second epochs; hus, he o al epochs we e 51,053. I should be no ed ha we encoun e ed a ha dwa e mal unc ion o one pa icipan ’s isi , bu we did no iden i y his mal unc ion un il he expe imen was comple ed. As a esul , we excluded a as amoun o da a o his pa icipan o subsequen analyses. Fo ea u e selec ion, we i s u ilized he -SNE g aph o unde s and whe he he e we e dissocia ions be ween aining da a om isi wo and es ing da a om isi one (Figu e 1). We ound no signi ican sepa a ions be ween he wo da a segmen a ion, sugges ing no dissocia ions be ween he wo da ase s. This esul indica es ha we can sa ely use he aining se based on a omic measu emen s in isi 2 o p edic ing semi-na u alis ic ac i i ies in isi 1. In . J. En i on. Res. Public Heal h 2021,18, 1634 7 o 16 In . J. En i on. Res. Public Heal h 2021, 18, x 7 o 16 Figu e 1. -Dis ibu ed S ochas ic Neighbo Embedding ( -SNE) g aph showing segmen a ion o bo h aining and es da a se s. Figu e 2 shows h ee con usion ma ices co esponding o classi ica ions based on 19, 10, and 7 asks. We selec ed asks e ained in he model based on he s epwise eg ession p ocess, in which we emo ed one ADL/IADL ask on each un wi h he lowes accu acy. Using he XGBoos algo i hm, we achie ed an a e age accu acy o 97% on he aining se and an accu acy o 90% in an independen es based on se en asks in a sample composed o all subjec s while pe o ming semi-na u alis ic ac i i ies. We also ob ained 0.91 o e- call (sensi i i y), 0.83 o p ecision, and 0.98 o AUC in he independen es se . The se en asks include cu ing, acuuming, sweeping, sp eading jam o bu e , olding laund y, ea ing, and b ushing ee h. Ten asks p oduced 86% accu acy on he aining se and an 82% accu acy on he independen es se . Ten asks include he p e ious se en asks plus: Taking o /pu ing on a shi , wiping cupboa ds, and bu oning a shi . We u he com- pu ed he -SNE dis ibu ions o he en asks and ound many o hem clus e ed oge he (Figu e 3), sugges ing ha hese a e p edic able s uc u ed classes. SHAP indica ed ha , among he i e IMUs, bo h w is s and he hip cap u ed he mos c i ical in o ma ion o ac i i y ecogni ion. The impo an ea u es con ained gy oscopic s anda d de ia ion and accele ome e s anda d de ia ion and mean. Table 3 shows he pe o mance e alua ion me ics ac oss ML models. We ound ha XGBoos op imizes o equa es wi h me hods ha p oduce he bes esul s. These esul s p o ided ini ial e idence ha ADL/IADL unc- ioning can be p edic ed wi h adequa e accu acy by using wea able senso s (IMUs) and ML algo i hms. We also used he SHAP alues o iden i y he con ibu ion o he indi id- ual IMU ea u es o p edic ac i i ies in he model p edic ion (Figu e 4). Among 5 IMUs, bo h w is s and he hip, cap u ing he mos c i ical in o ma ion o p edic ing ac i i ies. Essen ial IMU ea u es included s anda d de ia ion and mean ea u es om gy oscope and accele ome e . Table 3. Pe o mance me ics o se en ADL asks classi ica ion ac oss machine lea ning (ML) models. Pe o mance Me ic 1 Decision T ee Random Fo es SVM XGBoos T aining Se Accu acy 0.56 0.79 0.97 0.97 AUC 0.74 0.88 0.99 0.98 P ecision 0.50 0.80 0.97 0.97 Recall 0.56 0.79 0.97 0.97 Tes Se Accu acy 0.43 0.80 0.90 0.90 AUC 0.68 0.89 0.95 0.98 P ecision 0.47 0.84 0.92 0.83 Recall 0.43 0.80 0.90 0.91 1 No e: Numbe s in bold ep esen he bes alues ac oss ML models. Figu e 1. -Dis ibu ed S ochas ic Neighbo Embedding ( -SNE) g aph showing segmen a ion o bo h aining and es da a se s. Figu e 2shows h ee con usion ma ices co esponding o classi ica ions based on 19, 10, and 7 asks. We selec ed asks e ained in he model based on he s epwise eg ession p ocess, in which we emo ed one ADL/IADL ask on each un wi h he lowes accu acy. Using he XGBoos algo i hm, we achie ed an a e age accu acy o 97% on he aining se and an accu acy o 90% in an independen es based on se en asks in a sample composed o all subjec s while pe o ming semi-na u alis ic ac i i ies. We also ob ained 0.91 o ecall (sensi i i y), 0.83 o p ecision, and 0.98 o AUC in he independen es se . The se en asks include cu ing, acuuming, sweeping, sp eading jam o bu e , olding laund y, ea ing, and b ushing ee h. Ten asks p oduced 86% accu acy on he aining se and an 82% accu acy on he independen es se . Ten asks include he p e ious se en asks plus: Taking o /pu ing on a shi , wiping cupboa ds, and bu oning a shi . We u he compu ed he -SNE dis ibu ions o he en asks and ound many o hem clus e ed oge he (Figu e 3), sugges ing ha hese a e p edic able s uc u ed classes. SHAP indica ed ha , among he i e IMUs, bo h w is s and he hip cap u ed he mos c i ical in o ma ion o ac i i y ecogni ion. The impo an ea u es con ained gy oscopic s anda d de ia ion and accele ome e s anda d de ia ion and mean. Table 3shows he pe o mance e alua ion me ics ac oss ML models. We ound ha XGBoos op imizes o equa es wi h me hods ha p oduce he bes esul s. These esul s p o ided ini ial e idence ha ADL/IADL unc ioning can be p edic ed wi h adequa e accu acy by using wea able senso s (IMUs) and ML algo i hms. We also used he SHAP alues o iden i y he con ibu ion o he indi idual IMU ea u es o p edic ac i i ies in he model p edic ion (Figu e 4). Among 5 IMUs, bo h w is s and he hip, cap u ing he mos c i ical in o ma ion o p edic ing ac i i ies. Essen ial IMU ea u es included s anda d de ia ion and mean ea u es om gy oscope and accele ome e . In . J. En i on. Res. Public Heal h 2021,18, 1634 8 o 16 In . J. En i on. Res. Public Heal h 2021, 18, x 8 o 16 (a) (b) (c) Figu e 2. Con usion ma ices co esponding o classi ica ions based on (a) 19 asks, (b) 10 asks, and (c) 7 asks. Tasks selec ion was pe o med ollowing a s anda d s epwise eg ession p ocess using XGBoos . Figu e 2. Con usion ma ices co esponding o classi ica ions based on ( a ) 19 asks, ( b ) 10 asks, and ( c ) 7 asks. Tasks selec ion was pe o med ollowing a s anda d s epwise eg ession p ocess using XGBoos . In . J. En i on. Res. Public Heal h 2021,18, 1634 9 o 16 In . J. En i on. Res. Public Heal h 2021, 18, x 9 o 16 Figu e 3. -SNE g aph o he en ADL asks. Do s indica e da a om di e en ADL asks. Figu e 4. In luence o a ious ine ial measu emen uni s (IMU) ea u es o p edic ing ADL asks on he XGBoos model. 4. Discussion Using mo ion-based da a collec ed om IMUs, we ha e de eloped and alida ed ML algo i hms o ecognize a lis o ADLs among a sample o communi y-dwelling s oke su i o s. A signi ican con ibu ion o ou s udy is he use o independen aining and es ing da a se s. We asked pa icipan s o comple e semi-na u alis ic ac i i ies in an ob- s acle cou se design (e.g., unload g oce ies and hen p epa e and ea s i - y using ing e- dien s). We also asked hem o comple e a omic ac i i ies eco ded wi hin a block design (e.g., ou minu es o chopping ege ables). A e ha , we ained he ML algo i hms wi h he a omic ac i i ies and es ed hem wi h semi-na u alis ic ac i i ies. We ound ha ou ML algo i hms could ecognize a se o ac i i ies in a semi-na u alis ic en i onmen wi h adequa e accu acy. Enabling pa icipa ion in daily ac i i ies is he ul ima e ou come in ehabili a ion. These esul s p o ide ini ial e idence ha mo ion-based senso s and ML, especially he XGBoos and SVM algo i hms, can p edic pos -s oke daily ac i i ies. Figu e 3. -SNE g aph o he en ADL asks. Do s indica e da a om di e en ADL asks. Table 3. Pe o mance me ics o se en ADL asks classi ica ion ac oss machine lea ning (ML) models. Pe o mance Me ic 1Decision T ee Random Fo es SVM XGBoos T aining Se Accu acy 0.56 0.79 0.97 0.97 AUC 0.74 0.88 0.99 0.98 P ecision 0.50 0.80 0.97 0.97 Recall 0.56 0.79 0.97 0.97 Tes Se Accu acy 0.43 0.80 0.90 0.90 AUC 0.68 0.89 0.95 0.98 P ecision 0.47 0.84 0.92 0.83 Recall 0.43 0.80 0.90 0.91 1No e: Numbe s in bold ep esen he bes alues ac oss ML models. In . J. En i on. Res. Public Heal h 2021, 18, x 9 o 16 Figu e 3. -SNE g aph o he en ADL asks. Do s indica e da a om di e en ADL asks. Figu e 4. In luence o a ious ine ial measu emen uni s (IMU) ea u es o p edic ing ADL asks on he XGBoos model. 4. Discussion Using mo ion-based da a collec ed om IMUs, we ha e de eloped and alida ed ML algo i hms o ecognize a lis o ADLs among a sample o communi y-dwelling s oke su i o s. A signi ican con ibu ion o ou s udy is he use o independen aining and es ing da a se s. We asked pa icipan s o comple e semi-na u alis ic ac i i ies in an ob- s acle cou se design (e.g., unload g oce ies and hen p epa e and ea s i - y using ing e- dien s). We also asked hem o comple e a omic ac i i ies eco ded wi hin a block design (e.g., ou minu es o chopping ege ables). A e ha , we ained he ML algo i hms wi h he a omic ac i i ies and es ed hem wi h semi-na u alis ic ac i i ies. We ound ha ou ML algo i hms could ecognize a se o ac i i ies in a semi-na u alis ic en i onmen wi h adequa e accu acy. Enabling pa icipa ion in daily ac i i ies is he ul ima e ou come in ehabili a ion. These esul s p o ide ini ial e idence ha mo ion-based senso s and ML, especially he XGBoos and SVM algo i hms, can p edic pos -s oke daily ac i i ies. Figu e 4. In luence o a ious ine ial measu emen uni s (IMU) ea u es o p edic ing ADL asks on he XGBoos model. In . J. En i on. Res. Public Heal h 2021,18, 1634 16 o 16 31. Be gs a, J.; Yamins, D.; Cox, D.D. Making a Science o Model Sea ch: Hype pa ame e Op imiza ion in Hund eds o Dimensions o Vision A chi ec u es. In P oceedings o he 30 h In e na ional Con e ence on Machine Lea ning, A lan a, GA, USA, 17–19 June 2013; Volume 28, pp. 115–123. 32. Fe nández, A.; Ga cía, S.; Gala , M.; P a i, R.C.; K awczyk, B.; He e a, F. Lea ning om Imbalanced Da a Se s; Sp inge : Cham, Swi ze land, 2018. 33. Cheung, V.H.; G ay, L.; Ka unani hi, M. Re iew o accele ome y o de e mining daily ac i i y among elde ly pa ien s. A ch. Phys. Med. Rehabil. 2011,92, 998–1014. [C ossRe ] [PubMed] 34. Capela, N.A.; Lemai e, E.D.; Baddou , N. Fea u e selec ion o wea able sma phone-based human ac i i y ecogni ion wi h able bodied, elde ly, and s oke pa ien s. PLoS ONE 2015,10, e0124414. [C ossRe ] 35. Capela, N.A.; Lemai e, E.D.; Baddou , N.; Rudol , M.; Golja , N.; Bu ge , H. E alua ion o a sma phone human ac i i y ecogni ion applica ion wi h able-bodied and s oke pa icipan s. J. Neu oeng. Rehabil. 2016,13, 5. [C ossRe ] 36. Bailey, R.R.; Bi kenmeie , R.L.; Lang, C.E. Real-wo ld a ec ed uppe limb ac i i y in ch onic s oke: An examina ion o po en ial modi ying ac o s. Top. S oke Rehabil. 2015,22, 26–33. [C ossRe ] 37. Lang, C.E.; Bland, M.D.; Bailey, R.R.; Schae e , S.Y.; Bi kenmeie , R.L. Assessmen o uppe ex emi y impai men , unc ion, and ac i i y a e s oke: Founda ions o clinical decision making. J. Hand The . 2013,26, 104–114; quiz 115. [C ossRe ] 38. Bailey, R.R.; Klaesne , J.W.; Lang, C.E. An accele ome y-based me hodology o assessmen o eal-wo ld bila e al uppe ex emi y ac i i y. PLoS ONE 2014,9, e103135. [C ossRe ]