Machine Learning Solutions Applied to Amyotrophic Lateral Sclerosis Prognosis: A Review

SYSTEMATIC REVIEW published: 28 Ap il 2022 doi: 10.3389/ comp.2022.869140 F on ie s in Compu e Science | www. on ie sin.o g 1Ap il 2022 | Volume 4 | A icle 869140 Edi ed by: Pe e Kokol, Uni e si y o Ma ibo , Slo enia Re iewed by: Ricca do Rosa i, Ma che Poly echnic Uni e si y, I aly Qian Du, GNS Heal hca e, Uni ed S a es *Co espondence: Fabiano Papaiz [email p o ec ed] Special y sec ion: This a icle was submi ed o Digi al Public Heal h, a sec ion o he jou nal F on ie s in Compu e Science Recei ed: 03 Feb ua y 2022 Accep ed: 24 Ma ch 2022 Published: 28 Ap il 2022 Ci a ion: Papaiz F, Dou ado MET J , Valen im RAM, Mo ais AHF and A ais JP (2022) Machine Lea ning Solu ions Applied o Amyo ophic La e al Scle osis P ognosis: A Re iew. F on . Compu . Sci. 4:869140. doi: 10.3389/ comp.2022.869140 Machine Lea ning Solu ions Applied o Amyo ophic La e al Scle osis P ognosis: A Re iew Fabiano Papaiz1,2,3*, Ma io Emílio Teixei a Dou ado J .1,4, Rica do Alexsand o de Medei os Valen im1, An onio Higo F ei e de Mo ais 1,3 and Joel Pe diz A ais2 1Labo a o y o Technological Inno a ion in Heal h (LAIS), Fede al Uni e si y o Rio G ande do No e, Na al, B azil, 2Cen e o In o ma ics and Sys ems o he Uni e si y o Coimb a, Depa men o In o ma ics Enginee ing, Uni e si y o Coimb a, Coimb a, Po ugal, 3Ad anced Nucleus o Technological Inno a ion, Fede al Ins i u e o Rio G ande do No e, Na al, B azil, 4Depa men o In e nal Medicine, Fede al Uni e si y o Rio G ande do No e, Na al, B azil The p ognosis o Amyo ophic La e al Scle osis (ALS), a complex and a e disease, ep esen s a challenging and essen ial ask o be e comp ehend i s p og ession and imp o e pa ien s’ quali y o li e. The use o Machine Lea ning (ML) echniques in heal hca e has p oduced aluable con ibu ions o he p ognosis ield. This a icle p esen s a sys ema ic and c i ical e iew o p ima y s udies ha used ML applied o he ALS p ognosis, sea ching o da abases, ele an p edic o bioma ke s, he ML algo i hms and echniques, and hei ou comes. We ocused on s udies ha analyzed bioma ke s commonly p esen in he ALS disease clinical p ac ice, such as demog aphic, clinical, labo a o y, and imaging da a. Hence, we in es iga e s udies o p o ide an o e iew o solu ions ha can be applied o de elop decision suppo sys ems and be used by a highe numbe o ALS clinical se ings. The s udies we e e ie ed om PubMed, Science Di ec , IEEEXplo e, and Web o Science da abases. A e comple ing he sea ching and sc eening p ocess, 10 a icles we e selec ed o be analyzed and summa ized. The s udies e alua ed and used di e en ML algo i hms, echniques, da ase s, sample sizes, bioma ke s, and pe o mance me ics. Based on he esul s, h ee dis inc ypes o p edic ion we e iden i ied: Disease P og ession, Su i al Time, and Need o Suppo . The bioma ke s iden i ied as ele an in mo e han one s udy we e he ALSFRS/ALSFRS-R, disease du a ion, Fo ced Vi al Capaci y, Body Mass Index, age a onse , and C ea inine. In gene al, he s udies p esen ed p omisso y esul s ha can be applied in de eloping decision suppo sys ems. Besides, we discussed he open challenges, he limi a ions iden i ied, and u u e esea ch oppo uni ies. Keywo ds: Amyo ophic La e al Scle osis, p ognosis, Machine Lea ning, heal h in o ma ics, li e a u e e iew 1. INTRODUCTION Amyo ophic La e al Scle osis (ALS) is a a e, incu able, and p og essi e disease ha a ec s he neu ons o he human mo o sys em. The communica ion be ween he b ain and muscles is g adually in e up ed, leading pa ien s o pa alysis and dea h. I s causes a e unknown, ypically commi s men and women be ween he ages o 40 and 70. The a e age li e expec ancy is 3–5 yea s a e symp oms onse , and he wo ldwide incidence is abou 1.9 cases pe 100,000 indi iduals Papaiz e al. Machine Lea ning Applied o ALS P ognosis pe yea . ALS is clinically he e ogeneous, p esen ing di e en si es o disease onse , ex a-mo o in ol emen s, p og ession a es, and su i al imes among hei pa ien s (Ande sen e al., 2012; Chiò e al., 2014; Swinnen and Robbe ech , 2014; Ha diman e al., 2017). Being ALS a complex disease, p o iding an accu a e p ognosis becomes a challenge o he physicians (e.g., su i al ime, disease p og ession, momen o in oducing speci ic ea men s). Thus, i is essen ial o iden i y ele an biological ma ke s (bioma ke s) and unde s and how hey a e ela ed o ALS disease p og ession. Bioma ke s a e pa ame e s collec ed om he pa ien s ha can be used o con i m a disease p esence (diagnosis), ollow up a disease p og ession (p ognosis) o ea men esponse (moni o ing), and calcula e he p obabili y o de eloping a disease ( isk) (G oup, 2001). They can comp ise di e en da a ypes, such as clinical, biome ic, imaging, bio luid, and gene ic. P e ious s udies iden i ied help ul bioma ke s ha can assis in ALS p ognosis, such as age a symp om onse , diagnosis delay, weigh loss, bulba si e o onse , a e o unc ional and espi a o y impai men o e ime, mic oRNAs, neu o ilamen s, and labo a o y es s (ALS, 1996; Ceda baum e al., 1999; Kollewe e al., 2008; Chiò e al., 2009; Va ghese e al., 2013; Ha diman e al., 2017; Walle e al., 2017). Resea ches using A i icial In elligence echniques, like Machine Lea ning (ML) algo i hms, ha e been success ully applied o imp o e he diagnosis and p ognosis o diseases, such as he ecen ad ances in he oncology ield (Kou ou e al., 2015; O’Shea e al., 2016). The ML ield aims o de elop compu e p og ams capable o lea ning using p e ious expe ience ( aining da a) wi hou being explici ly p og ammed o his. ML algo i hms could ex ac in o ma ion om he aining da a, ans o m i in o knowledge, and use i o sol e di e en ca ego ies o p oblems (e.g., classi ica ion, eg ession, clus e ing, Samuel, 1988). In heo y, he g ea e he amoun o aining da a a ailable, he g ea e he algo i hm’s lea ning and pe o mance (Mi chell, 1997; Kuba , 2017). In his sense, ha ing access o ALS pa ien da a is c ucial o pe o m ele an s udies in he p ognos ic a ea and c ea e ML solu ions o help physicians in hei daily wo k. The analysis o medical da a usually in ol es dealing wi h high-dimensional da a, co e ing a la ge numbe o bioma ke s. Thus, some ML echniques (e.g., Fea u e Selec ion, Dimensionali y Reduc ion) can be applied o ans o m a complex da ase in o a simple one by iden i ying he mo e ele an bioma ke s, which imp o e he lea ning pe o mance, da a collec ing e iciency, and algo i hm unde s anding (Lee and Ve leysen, 2007; B ank e al., 2011). ML algo i hms can be used o de elop Clinical Decision Suppo Sys ems (CDSS). The CDSS a e compu e p og ams designed o help physicians make mo e app op ia e and imely decisions abou hei pa ien s (Be ne e al., 2007; Beele e al., 2014; Gul epe e al., 2014; CDS, 2015; Rosa i e al., 2020; Romeo and F on oni, 2022). These sys ems usually p o ide p ognos ic p edic ions o imp o e he decision-making p ocess and, hus, imp o e he pa ien ’s quali y o li e. Some bene i s include imp o ing pa ien s’ quali y o ca e, ea men e iciency, esou ce planning, and educing cos s. CDSS also ep esen s a aluable ool o p omo e knowledge dissemina ion among all in e es ed heal h wo ke s. ML-based CDSS can imp o e clinical decisions TABLE 1 | Resea ch ques ions. RQ Ques ion 01 Wha a e he ALS da abases used in he s udy? 02 How many pa ien s comp ise he coho o he s udy? 03 Wha a e he ypes o p edic ion add essed by he s udy? 04 Wha a e he ML algo i hms and echniques used in he s udy? 05 Wha a e he bioma ke s e alua ed and he mos ele an iden i ied by he s udy? 06 Wha a e he pe o mances o he used ML algo i hms? by helping physicians analyze and make in e ences on a la ge amoun o pa ien da a. Howe e , some ML app oaches p esen esul s ha can no be easily unde s ood, dec easing hei in e p e abili y (e.g., A i icial Neu al Ne wo ks o Suppo Vec o Machines). In e p e abili y e e s o how well a pe son can unde s and he decisions made by he ML algo i hm (Mille , 2019). This issue can di icul he p ocess o accep ance and in eg a ion o a CDSS in he clinical en i onmen ou ine. Consequen ly, he de elopmen o a CDSS mus ha e conce ned abou in e p e abili y issues, being anspa en enough so ha heal h wo ke s can unde s and how any suppo was o e ed. Many coun ies p esen inancial limi a ions on hei heal h sys em. This ac makes i un easible o collec complex and cos ly bioma ke s (e.g., gene ic) in p ima y ca e. In his manne , i is essen ial o ca y ou s udies conside ing hese limi a ions o de elop compu a ional solu ions (e.g., CDSS) ha can assis a highe numbe o p ima y ca e uni s. The main objec i e o his s udy is o in es iga e ML app oaches on ALS p ognosis ha analyzed less complex bioma ke s, which can be po en ially applied o de elop clinical decision suppo sys ems o assis physicians in he eal-wo ld ALS clinical se ing. We ocused on s udies ha analyzed bioma ke s commonly p esen in he ALS disease clinical p ac ice, such as demog aphic, clinical (including unc ional, espi a o y, and nu i ional), labo a o y, and imaging da a. Hence, we in es iga e s udies using bioma ke s ob ained h ough a less complex p ocess, aiming o p o ide an o e iew o solu ions ha can be applied o de elop decision suppo sys ems and be used on a la ge scale in p ima y ca e, conside ing inancial limi a ions. In his sense, we did no include s udies using omics da a (i.e., genomic, ansc ip omic, p o eomic, and me abolomic). We desc ibed he ecen ad ances in his a ea, he cu en ly a ailable da ase s, he bioma ke s analyzed, he ML algo i hms and echniques used, he mos ele an bioma ke s iden i ied, and hei ou comes. Besides, we discussed he open challenges, he limi a ions iden i ied, and u u e esea ch oppo uni ies. 2. METHODS This sys ema ic e iew aims o in es iga e ML solu ions applied o ALS p ognosis. In his sense, we elabo a ed esea ch ques ions (RQ) o guide he conduc o his a icle, which a e p esen ed F on ie s in Compu e Science | www. on ie sin.o g 2Ap il 2022 | Volume 4 | A icle 869140 Papaiz e al. Machine Lea ning Applied o ALS P ognosis TABLE 2 | Inclusion c i e ia. IC Desc ip ion 01 A icles published in Jou nals 02 A icles w i en in English 03 A icles published be ween Janua y 2011 and Ap il 2021 04 A icles in he In o ma ion Technology, Compu e Enginee , o Compu e Science ela ed a eas TABLE 3 | Exclusion c i e ia. EC Desc ip ion 01 Re iew a icles 02 Duplica e a icles 03 A icles no ela ed o Machine Lea ning applied o ALS p ognosis 04 A icles using omics da a (i.e., genomic, ansc ip omic, p o eomic, and me abolomic) in Table 1. Nex , we pe o med he ollowing s ages: (i) sea ch a icles ela ed o ALS p ognosis using ML in scien i ic da abases, (ii) apply he inclusion c i e ia, (iii) apply he exclusion c i e ia, and (i ) analyze and summa ize he selec ed a icles. In he i s s age, he ele an li e a u e was ob ained om he PubMed,Science Di ec ,IEEEXplo e, and Web o Science da abases. The sea ch was pe o med in Ap il 2021 using he ollowing sea ch que y: (“a i icial in elligence” OR “machine lea ning” OR “deep lea ning”) AND (“amyo ophic la e al scle osis” OR “mo o neu one disease”) AND (“p edic ” OR “p ognosis” OR “p og ession”). We used he Rayyan Web Applica ion (Ouzzani e al., 2016) o o ganize he esul ing a icles and also o pe o m he emaining s ages. In he second and hi d s ages, we applied he Inclusion (IC) and Exclusion (EC) C i e ia o il e he a icles acco ding o he scope o his a icle (see Tables 2,3). We conside ed only a icles published in Jou nals, w i en in English, and published be ween Janua y 2011 and Ap il 2021 (IC-01, IC-02, and IC-03). A icles ha did no belong o he In o ma ion Technology, Compu e Enginee , o Compu e Science ela ed a eas we e no included (IC-04). Nex , we ca ied ou he emo al o he e iew a icles (EC-01), he duplica e en ies (EC-02), and a icles no ela ed o ML applied o ALS p ognosis (EC-03). Then, he a icles using omics da a we e emo ed (EC-04). Finally, in he ou h s age, he selec a icles we e ho oughly ead, which allowed he inal analysis and accomplishmen o he objec i es o his esea ch. 3. RESULTS Figu e 1 illus a es he sea ch and sc eening p ocess o his sys ema ic e iew. The sea ch que y and all inclusion c i e ia we e used o pe o m he da abase sea ches. A o al o 52 a icles we e e ie ed, whe e wo e iew a icles we e immedia ely excluded. A e he emo al o 15 duplica es, 35 a icles we e chosen o abs ac e iew. A o al o 25 s udies we e excluded due o he use o omic da a (n= 6) and no being ela ed o ML applied o ALS p ognosis (n= 19). A e comple ing he sea ching and sc eening p ocess, 10 a icles we e selec ed o be analyzed and summa ized. The ollowing sec ions p esen he esul s ha add ess he esea ch ques ions de ined in his s udy (Table 1). 3.1. ALS Da ase s and Sample Sizes Di e en da ase s we e analyzed and hei sample sizes anged om 41 up o o e 10,000 samples. Table 4 desc ibes all he da ase s analyzed. Mos o he s udies (60%) analyzed da a om he PRO-ACT (A assi e al., 2014) da ase , p obably because i was he only publicly a ailable. The o he da ase s used we e local o p op ie a y. The da a o ma s analyzed included abula (all s udies) and image ( an de Bu gh e al., 2017). Mo e de ail abou he sample size used by each s udy a e desc ibed in Tables 6–8. 3.2. Types o P edic ion Add essed Based on he included s udies, h ee dis inc ypes o p edic ion we e iden i ied: Disease P og ession,Su i al Time, and Need o Suppo (mo e de ail in Table 5). Kue ne e al. (2019) add essed he Disease P og ession and Su i al Time ypes simul aneously. The Disease P og ession p edic ion aimed o es ima e he pa ien ’s s a e a a gi en momen in he u u e and was he ype mos add essed by he s udies included (70%). The Su i al Time p edic ion aimed o es ima e he occu ence o dea h om a baseline da e o a poin - ime in he u u e, such as he p obabili y o dea h a e 12 mon hs om symp oms onse . The Need o Suppo p edic ion aimed o es ima e he momen when pa ien s will need mo e specialized suppo . 3.3. P edic i e Machine Lea ning App oaches Fo Disease P og ession p edic ion, mos s udies aimed o es ima e changes in he ALS Func ional Ra ing Scale (ALSFRS) o he Re ised ALS Func ional Ra ing Scale (ALSFRS-R) o e ime. Two o he s udies aimed o classi y pa ien s conce ning hei disease p og ession a es (Slow/Fas Kue ne e al., 2019, Low/High G eco e al., 2021). Table 6 de ails he a ge p edic ions, bes ML algo i hm, pe o mance, da ase s, samples size, echniques, alida ion s a egies, and bioma ke s e alua ed o each s udy. The s udies ha add essed he Su i al Time p edic ion aimed o classi y he pa ien s in o su i al g oups and es ima e he p obabili y o dea h a e a speci ic ime in e al. an de Bu gh e al. (2017) aimed o classi y pa ien s in o Sho (<25 mon hs), Medium (25−50 mon hs), o Long (>50 mon hs) su i al g oups. Kue ne e al. (2019) aimed o es ima e he p obabili y o su i al a e 12, 18, and 24 mon hs. G ollemund e al. (2020) aimed o es ima e he p obabili y o pa ien s being ali e a e 12 mon hs. All h ee s udies used he da e o symp oms onse as he baseline da e. The cha ac e is ics o each s udy a e de ailed in Table 7. Pi es e al. (2018) was he unique s udy ha add essed he Need o Suppo p edic ion, aiming o es ima e he need o Non- In asi e Ven ila ion (NIV) suppo a e 3, 6, and 12 mon hs. The cha ac e is ics o his s udy a e de ailed in Table 8. F on ie s in Compu e Science | www. on ie sin.o g 3Ap il 2022 | Volume 4 | A icle 869140 Papaiz e al. Machine Lea ning Applied o ALS P ognosis FIGURE 1 | PRISMA low cha o his s udy. 3.4. Bioma ke s E alua ed and he Mos Rele an Iden i ied As p e iously men ioned, we ocused on he bioma ke s commonly p esen in he ALS disease clinical p ac ice, being ob ained in a less cos ly and complex way. The bioma ke s e alua ed comp ise clinical, demog aphic, i al signs, espi a o y, unc ional, labo a o y, imaging, neu ophysiological, and medica ion da a. Fo mo e de ail, please see column Bioma ke s E alua ed in Tables 6–8. All he selec ed s udies e alua ed he ALS Func ional Ra ing Scale (ALSFRS) o he Re ised ALS Func ional Ra ing Scale (ALSFRS-R) bioma ke s. This ac highligh s he impo ance o hese bioma ke s in moni o ing ALS pa ien s. Table 9 depic s he mos ele an bioma ke s iden i ied in he s udies, wi h he in o ma ion abou hei associa ed ypes o p edic ion. They comp ised clinical, imaging, unc ional, espi a o y, and labo a o y da a. The bioma ke s iden i ied as ele an in mo e han one s udy we e he ALSFRS/ALSFRS-R (n = 7), disease du a ion (n= 5), Fo ced i al capaci y (n= 4), Body mass index (n= 2), age a onse (n= 2), and C ea inine (n= 2). 3.5. Desc ip ion o he S udies an de Bu gh e al. (2017) demons a ed he posi i e impac o using Magne ic Resonance Images (MRI) along wi h clinical in o ma ion o classi y ALS pa ien s in o h ee su i al g oups: Sho (<25 mon hs), Medium (25−50 mon hs), and Long (>50 mon hs). The bioma ke s e alua ed we e clinical in o ma ion (e.g., si e o onse , age a onse , ALSFRS slope, FVC) and MRI images (S uc u al Connec i i y and B ain Mo phology da a) om 135 ALS pa ien s. They de eloped Deep Neu al Ne wo ks models and e alua e hem in ou scena ios using di e en bioma ke s se s: (i) only Clinical Da a, (ii) only S uc u al Connec i i y MRI Da a, (iii) only B ain Mo phology MRI Da a, and (i ) combining Clinical and MRI Da a. The g ea e accu acy was ob ained using he Clinical-MRI combined da a (84%) compa ed o he o he h ee s a egies (Clinical: 69%; S uc u al Connec i i y MRI: 63%; B ain Mo phology MRI: 63%). They poin ed ou he powe o Deep Neu al Ne wo ks in making p edic ions using complex da a. Howe e , he ela ionships be ween inpu and ou pu a iables could no be easily F on ie s in Compu e Science | www. on ie sin.o g 4Ap il 2022 | Volume 4 | A icle 869140 Papaiz e al. Machine Lea ning Applied o ALS P ognosis TABLE 4 | Lis o da ase s used in he s udies, in e sely o de ed by he sample size. Da ase Samples Re e ences PRO-ACT +10,000 Go don and Le ne , 2019; Halbe sbe g and Le ne , 2019; Kue ne e al., 2019; Tang e al., 2019; G ollemund e al., 2020; Hadad and Le ne , 2020 I eland-I alia 1,479 Kue ne e al., 2019 Tel A i —Sou asky Medical Cen e 1,328 Hadad and Le ne , 2020 Lisbon—Sain Ma y’s Hospi al 1,214 Pi es e al., 2018; Leão e al., 2021 Pa is—Te ia y e e al Cen e o ALS 646 G ollemund e al., 2020 T ophos Company 431 G ollemund e al., 2020 Exonhi Pha ma 172 G ollemund e al., 2020 U ech —Uni e si y Medical Cen e 135 an de Bu gh e al., 2017 I alia 41 G eco e al., 2021 TABLE 5 | Types o p edic ion add essed by he s udies. Type Numbe o s udies Re e ences Disease p og ession 7Go don and Le ne , 2019; Halbe sbe g and Le ne , 2019; Kue ne e al., 2019; Tang e al., 2019; Hadad and Le ne , 2020; G eco e al., 2021; Leão e al., 2021 Su i al ime 3 an de Bu gh e al., 2017; Kue ne e al., 2019; G ollemund e al., 2020 Need o suppo 1 Pi es e al., 2018 ecognized, needing mo e in es iga ion o unde s and ALS p og ession be e . Pi es e al. (2018) de eloped a model o p edic when a pa ien will need NIV suppo acco ding o a gi en ime window (3, 6, and 12 mon hs). They used he Po uguese ALS Da ase (n =1,070), combining he s a ic and empo al da a in o a da a s uc u e called snapsho , which con ains all in o ma ion abou a pa ien a a speci ic da e. The pa ien s we e di ided in o h ee disease p og ession g oups (Slow, Neu al, and Fas ) and, o each g oup, hei espec i e snapsho s we e used as lea ning ins ances o e alua e se e al ML models. A Fea u e Selec ion Ensemble app oach was used o selec he ele an bioma ke s o each g oup. The Random Fo es model ob ained he bes pe o mance o 3, 6, and 12 mon hs ime window alues. The ele an bioma ke s p esen in all g oups we e BMI, FVC, and VC. O he ele an bioma ke s (p esen in 75% o he ime) we e age a onse , disease du a ion, and ALSFRS sco e. The au ho s epo ed he ad an age o using specialized ML models o di e en pa ien g oups (e.g., disease p og ession g oups) a he han c ea e gene alized models ea ing all he pa ien s simila ly. Halbe sbe g and Le ne (2019) demons a ed he bene i o using empo al modeling, sequence clus e ing, and sequen ial pa e n mining o p edic he las pa ien s a e eco ded (ALSFRS sco e) based on his pas in o ma ion. To ind ele an de e io a ion pa e ns in empo al pa ien s da a hey de eloped a amewo k consis ing o h ee s ages: (i) g oup pa ien s wi h simila p og ession using hie a chical clus e ing based on Dynamic Time Wa ping, (ii) pe o m pa e n mining o ound ou common unc ional de e io a ion pa e ns among pa ien s based on he SPADE sequence mining algo i hm, and (iii) de elop a Random Fo es model o classi y pa ien s in o hei mos simila clus e o p edic hei nex disease s a e. The pe o mance ob ained by he p oposed amewo k (Accu acy: 73, F1 sco e: 0.68, Mean Absolu e E o : 0.3) was supe io ela ed o wo o he benchma k models (Random Fo es and Long Sho -Te m Memo y, bo h using no empo al modeling). They used s a ic (e.g., age a onse , ime om onse , gende ) and longi udinal (ALSFRS sco es and subsco es) da a o 2,590 subjec s om he PRO-ACT da ase . The mos impo an p edic o s epo ed we e he p e ious ALSFRS sco e, he p e ious ALSFRS D essing subsco e, he p e ious Climbing S ai s subsco e, he p e ious Tu ning in Bed subsco e, he ime om disease onse , and he de e io a ion pa e n e med <E,G,I> (i.e., a sequen ial declining in he W i ing,D essing, and Walking ALSFRS subsco es). Go don and Le ne (2019) e alua ed he capaci y o o dinal classi ie s o p edic he unc ional decline o he pa ien s. They used da a abou he i s and las pa ien isi s om he PRO- ACT da ase (n=3,772), analyzing he ollowing bioma ke s: clinical, demog aphic, ALSFRS, FVC, medica ion, i al signs, and labo a o y es s. The a ge a iables we e all en ALSFRS i ems (ques ions) sepa a ely. The pa ien s a es we e mapped o he ALSFRS i ems, hus co ela ing pa ien s a e o disease p og ession o each poin in ime. Add essing he o dinal na u e o he ALSFRS, hey e alua ed he ollowing o dinal classi ie s: Cumula i e Link Models (CLM), O dinal Decision T ees (ODT), and Cumula i e P obabili y T ee (CPT). To e alua e hei pe o mances, hey de ined a penalizing sys em ha accoun s o a ious e o se e i ies di e en ly. Thus, a classi ie was less penalized when i p edic ed he alue o 2 ins ead o 1 when he eal alue was 3. These h ee classi ie s we e compa ed wi h he Random Fo es (RF), a non-o dinal classi ie . The esul s showed ha he CLM and ODT o dinal classi ie s p esen ed a simila pe o mance and ou pe o med he RF classi ie ega ding he Mean Absolu e E o measu ed in he bes expe imen scena io (CLM: 0.62−1.06; ODT: 0.63−1.01; RF: 1.01−1.61). Fo ea u e selec ion, he au ho s implemen ed an algo i hm based on he J3 sca e ing ma ix c i e ion o each ALSFRS i em indi idually. The mos ele an p edic o s we e he FVC, he si e o onse , he ime om onse , and he labo a o y es s C ea inine, CK, Chlo ide, Phospho us, and Alkaline Phospha ase. A c owdsou cing s a egy was p esen ed in Kue ne e al. (2019), whe e we e selec ed 30 eams a ound he wo ld o pa icipa e in an ALS s a i ica ion challenge. They asked he pa icipan s o c ea e ML models o pe o m p edic ion asks using he PRO-ACT and he I ish-I alian Regis ies da ase s. The eams used pa ien da a om he i s h ee mon hs and we e limi ed o e alua e only six o all bioma ke s a ailable. The a ge p edic ions we e he Disease P og ession a 12 mon hs (decline o he Func ional Ra e Scale) and he P obabili y o Su i al a 12, F on ie s in Compu e Science | www. on ie sin.o g 5Ap il 2022 | Volume 4 | A icle 869140 Papaiz e al. Machine Lea ning Applied o ALS P ognosis TABLE 6 | O e iew o ML app oaches on disease p og ession. Re e ences Ta ge p edic ion Bes algo i hm Pe o mance Bioma ke s e alua ed Da ase (Samples) Techniques Valida ion Halbe sbe g and Le ne (2019) Las pa ien s a e (ALSFRS sco e) eco ded based on his pas in o ma ion SPADE + DTW + Clus e ing Me hod Accu acy: 73 F1 Sco e: 0.68 MAE: 0.30 Tabula : Clinical, demog aphic, labo a o y, ALSFRS PRO-ACT (2,590) Hold-ou Go don and Le ne (2019) Las pa ien s a e (ALSFRS subsco es) eco ded based on his pas in o ma ion CLM, and ODT MAE (min-max): -CLM: 0.62−1.06 -ODT: 0.63−1.01 Tabula : Clinical, demog aphic, i al signs, labo a o y, ALSFRS PRO-ACT (3,772) FS 10-Fold CV Kue ne e al. (2019) -ALSFRS sco e a 12 mon hs, using da a om he i s 3 mon hs and only 6 bioma ke s. -Pa ien s classi ica ion in o slow/ as p og ession g oups. GBM, and RF GBM (PRO-ACT): -Z-sco e: ≈12 RF (I eland-I alia) -Z-sco e: ≈6 Tabula : Clinical, demog aphic, i al signs, labo a o y, FVC, SVC, ALSFRS PRO-ACT (10,723) I eland-I alia (1,479) Hold-ou Tang e al. (2019) ALSFRS sco e and FVC a 12 mon hs, using 1s isi and 3-mon h da a BART (ALSFRS), and RF (FVC) BART: -R2: 0.22 - RMSE: 0.55 - Co : 0.47 RF: -R2: 0.68 - RMSE: 14.27 - Co : 0.83 Tabula : Clinical, demog aphic, pulse, BMI, FVC, labo a o y, Riluzole medica ion, ALSFRS PRO-ACT (2,424) FS MI 5-Fold CV Hadad and Le ne (2020) ALSFRS sco e a se e al ime in e als, a ying om 6 up o 24 mon hs XGBoos RMSE: 2.65−5.57 MAE: 1.98−4.42 Tabula : Clinical, demog aphic, i al signs, FVC, labo a o y, ALSFRS PRO-ACT (3,171) Tel A i (1,328) FIA Hold-ou G eco e al. (2021) Pa ien s classi ica ion in o low/high p og ession a es g oups SVM Accu acy: 87.25 Tabula : Clinical, demog aphic, labo a o y, ALSFRS-R I alia (41) FS LOO CV Leão e al. (2021) Changes in he ALSFRS-R sco e and subsco es (be o e and a e NIV) Ex ension o DBN Accu acy: 74−88 Sensi i i y: 57−95 AUC: 75−98 Tabula : Clinical, demog aphic, El Esco ial, BMI, C9o 72, FVC, MIP, MEP, PNRA, ALSFRS, ALSFRS-R Lisbon (1,214) MI 5-Fold CV ALSFRS, ALS unc ional a ing scale; ALSFRS-R, e ised ALS unc ional a ing scale; NIV, non-in asi e en ila ion; BMI, body mass index; FVC, o ced i al capaci y; SVC, slow i al capaci y; MIP, maximum inspi a o y p essu e; MEP, maximum expi a o y p essu e; PNRA, ph enic ne e esponse ampli ude; SPADE, sequen ial pa e n disco e y using equi alence class; DTW, dynamic ime wa ping; CLM, cumula i e link models; ODT, o dinal decision ees; GBM, gene alized boos ing model; RF, andom o es ; BART, Bayesian addi i e eg ession ee; SVM, suppo ec o machine; DBN, dynamic Bayesian ne wo k; AUC, a ea unde he ROC cu e; MAE, mean absolu e e o ; MSPE, mean squa ed p edic ion e o ; R2, coe icien o de e mina ion; RMSE, oo mean squa e e o ; Co , Pea son’s co ela ion coe icien ; FS, ea u e selec ion; FIA, ea u e impo ance analysis; MI, missing da a impu a ion; CV, c oss- alida ion; LOO, lea e-one-ou . 18, and 24 mon hs. Rega ding he su i al p edic ion, one eam ou pe o med he o he s signi ican ly using a Gaussian P ocess Reg ession model, p esen ing a be e app oach in leading wi h he igh -censo ed pa ien ou come (dead o ial d opou ). The bes models ela ed o he disease p og ession p edic ion used he Gene alized Boos ing Model and he Random Fo es algo i hms. The mo e ele an bioma ke s we e disease du a ion, age a onse , si e o onse , gende , weigh , BMI, espi a o y exams (FVC and SVC), labo a o y es s (C ea inine and Segmen ed Neu ophils), and ALSFRS sco es and subsco es. Based on he ele an bioma ke s chosen by he eams, he au ho s ha e iden i ied ou dis inc pa ien g oups: Slow P og essing, Fas P og essing, Ea ly S age, and La e S age. The main bioma ke s ela ed o each g oup we e also de ailed in his s udy, whe e he au ho s highligh ed he impo ance o he ALSFRS Bulba subsco e (ques ions 1−3) in disc imina ing be ween g oups. Tang e al. (2019) add essed p edic ions in changing o he ALSFRS sco e and in he FVC pe cen age. They used s a ic and longi udinal bioma ke s om he PROC-ACT da ase (n=2,424), including only hose pa ien s wi h in o ma ion abou ALSFRS sco es o e ime. The longi udinal da a we e ans o med in o signa u e ec o s agg ega ing s a is ics alues (minimum, median, maximum, and slope). Using da a om he i s isi and a he 3-mon h, he au ho s c ea e models o p edic he changes in he ALSFRS slope a 12-mon h. The e alua ed models (Random Fo es and Bayesian Addi i e Reg ession T ee) achie ed modes esul s (Co ela ion: 0.47; RMSE: 0.55; R2: 0.22), hus, indica ing he di icul y in p edic ing F on ie s in Compu e Science | www. on ie sin.o g 6Ap il 2022 | Volume 4 | A icle 869140 Papaiz e al. Machine Lea ning Applied o ALS P ognosis TABLE 7 | O e iew o ML app oaches on su i al ime. Re e ences Ta ge p edic ion Bes algo i hm Pe o mance Bioma ke s e alua ed Da ase (Samples) Techniques Valida ion an de Bu gh e al. (2017) Pa ien s classi ica ion in o Sho (<25 mon hs), Medium (25−50), and Long (>50) su i al g oups Deep neu al ne wo ks Accu acy: 84 Tabula : Clinical, demog aphic, C9o 72, FTD, El Esco ial, ALSFRS. Image: MRI. U e ch (135) FS MI Hold-ou Kue ne e al. (2019) P obabili y o dea h wi hin 12, 18, and 24 mon hs Gaussian Reg ession PRO-ACT: -Z-sco e: ≈14.5 I eland-I alia: -Z-sco e: ≈13 Tabula : Clinical, demog aphic, i al signs, labo a o y, FVC, SVC, ALSFRS PRO-ACT (10,723) I eland-I alia (1,479) Hold-ou G ollemund e al. (2020) 1-yea su i al p edic ion, classi ying pa ien s in o high, in e media e, and low su i al a es g oups UMAP BAcc: 91% F1 Sco e: 96% Tabula : Clinical, demog aphic, ALSFRS PRO-ACT (3971) T ophos (431) Exonhi (172) Pa is (646) Hold-ou ALSFRS, ALS unc ional a ing scale; ALSFRS-R, e ised ALS unc ional a ing scale; MRI, magne ic esonance image, FTD, on o empo al demen ia; FVC, o ced i al capaci y; SVC, slow i al capaci y; UMAP, uni o m mani old app oxima ion and p ojec ion; BAcc, balanced accu acy; FS, ea u e selec ion; MI, missing da a impu a ion. TABLE 8 | O e iew o ML app oach on need o suppo . Re e ences Ta ge p edic ion Bes algo i hm Pe o mance Bioma ke s e alua ed Da ase (Samples) Techniques Valida ion Pi es e al. (2018) Pa ien s need o NIV suppo a 3, 6, and 12 mon hs o h ee p og ession g oups (slow, neu al, and as ) RF Slow: (3/6/12 mon hs) - AUC: 81/87/91 - Sens: 70/72/78 - Spec: 76/83/86 Neu al: (3/6/12 mon hs) - AUC: 76/82/86 - Sens: 58/62/79 - Spec: 78/83/77 Fas : (3/6/12 mon hs) - AUC: 72/81/79 - Sens: 51/71/74 - Spec: 77/76/71 Tabula : Clinical, demog aphic, El Esco ial, BMI, C9o 72, VC, FVC, P0.1, SNIP, MIP, MEP, NIV, PNRA, PNRL, CE, CF, ALSFRS, ALSFRS-R Lisbon (1070) FS DB 10-Fold CV ALSFRS, ALS unc ional a ing scale; ALSFRS-R, e ised ALS unc ional a ing scale; NIV, non-in asi e en ila ion; BMI, body mass index; FVC, o ced i al capaci y; SVC, slow i al capaci y; VC, i al capaci y; P0.1, ai way occlusion p essu e; SNIP, sni nasal inspi a o y p essu e; MIP, maximum inspi a o y p essu e; MEP, maximum expi a o y p essu e; PNRA, ph enic ne e esponse ampli ude; PNRL, ph enic ne e esponse la ency; CE, ce ical ex ension; CF, ce ical lexion; RF, andom o es ; AUC, a ea unde he ROC cu e; Sens, sensi i i y; Spec, speci ici y; FS, ea u e selec ion; DB, da a balancing; CV, c oss- alida ion. 12-mon h ALSFRS slope using he only baseline and 3-mon hs da a. Fea u e Selec ion was pe o med using he Random Fo es and he Knocko Fil e me hods. A e combining he op- anked bioma ke s e u ned by bo h me hods, he bes p edic i e bioma ke s we e he ALSFRS sco e, he disease du a ion, he FVC, and he Absolu e Monocy e Coun . To p edic he FVC Pe cen age changes be ween 3 and 12 mon hs, Random Fo es models we e es ed in wo scena ios (ei he including he baseline FVC o no ). The bes esul s we e ob ained using he FVC a baseline da a, demons a ing he powe o his bioma ke , which inc eased he co ela ion om 0.67 o 0.83. The au ho s also applied unsupe ised classi ica ion (K-Means) o ind dis inc pheno ypes g oups, ounding ou balanced clus e s among he pa ien s. Howe e , i was conside ed imp ac ical o clea ly unde s and how he g oups di e due o he high numbe o bioma ke s de ined o each g oup du ing he clus e ing p ocess. Hadad and Le ne (2020) s udied p edic ion o he ALSFRS sco e in se e al ime in e als, a ying om 6 o 24 mon hs. Tempo al (Long Sho Te m Memo y—LSTM) and non- empo al (Random Fo es , XGBoos , and Mul ilaye Pe cep on) models we e e alua ed o e he PRO-ACT da ase (n=3,171). To be used by he non- empo al models, he longi udinal da a we e ans o med in o ec o s con aining agg ega ed alues (mean, s anda d de ia ion, slope, minimum, maximum). Each model was es ed using 60 di e en andomly gene a ed con igu a ions, and hei a e aged pe o mances we e compa ed (Roo Mean Squa e E o and Mean Absolu e E o ). The XGBoos model ob ained supe io pe o mance o he mos F on ie s in Compu e Science | www. on ie sin.o g 7Ap il 2022 | Volume 4 | A icle 869140 Papaiz e al. Machine Lea ning Applied o ALS P ognosis TABLE 9 | Mos ele an bioma ke s iden i ied, associa ed p edic ions, and e e ences. Type Bioma ke Associa ed p edic ions/Re e ences Disease p og ession Su i al ime Need o suppo Clinical Age a disease onse Halbe sbe g and Le ne , 2019 Kue ne e al., 2019 – Body Mass Index (BMI) Kue ne e al., 2019; Leão e al., 2021 Kue ne e al., 2019 – Disease du a ion Go don and Le ne , 2019; Halbe sbe g and Le ne , 2019; Kue ne e al., 2019; Tang e al., 2019; Leão e al., 2021 Kue ne e al., 2019 – Si e o onse Go don and Le ne , 2019 – – Imaging Magne ic Resonance Imaging – an de Bu gh e al., 2017 – Func ional ALSFRS Halbe sbe g and Le ne , 2019; Kue ne e al., 2019; Tang e al., 2019; Hadad and Le ne , 2020 Kue ne e al., 2019; G ollemund e al., 2020 Pi es e al., 2018 ALSFRS-R Leão e al., 2021 –Pi es e al., 2018 Respi a o y Fo ced Vi al Capaci y (FVC) Go don and Le ne , 2019; Kue ne e al., 2019; Tang e al., 2019 Kue ne e al., 2019 Pi es e al., 2018 Maximal expi a o y p essu e (MEP) Leão e al., 2021 – – Maximal inspi a o y p essu e (MIP) Leão e al., 2021 – – Slow i al capaci y (SVC) Kue ne e al., 2019 Kue ne e al., 2019 – Vi al capaci y (VC) – – Pi es e al., 2018 Labo a o y Absolu e monocy e coun Tang e al., 2019 – – Alanine ansaminase (ALT) Tang e al., 2019 – – Alkaline phospha ase Go don and Le ne , 2019 – – Calcium Tang e al., 2019 – – Chlo ide Go don and Le ne , 2019 – – Choles e ol—To al G eco e al., 2021 – – Choles e ol—high-densi y (HDL) G eco e al., 2021 – – C ea ine kinase (CK) Go don and Le ne , 2019 – – C ea inine Go don and Le ne , 2019; Kue ne e al., 2019 – – Hema oc i Tang e al., 2019 – – Phospho us Go don and Le ne , 2019 – – Po assium Tang e al., 2019 – – Segmen ed neu ophils Kue ne e al., 2019 – – U ine Ph Kue ne e al., 2019 – – Vi amin B12 G eco e al., 2021 – – ime in e als e alua ed (RMSE: 2.65−5.57, MAE: 1.98−4.42), being mo e p ecise o sho e han longe in e als. The ele an p edic i e bioma ke s we e he ALSFRS subsco es. In ano he expe imen , hese models we e e alua ed in wo scena ios: (i) ained wi h he PRO-ACT and es ed wi h he TASMC da ase (n =1,328), and (ii) ained and es ed using only he TASMC da a. The sho - e m p edic ions (up o 6 mon hs) we e mo e p ecise using models ained wi h he PRO-ACT, and he XGBoos ob ained he bes esul s again. The au ho s highligh ed ha he PROC-ACT con ains da a om clinical ials ha may no e lec he eali y p esen ed by he clinical en i onmen pa ien s due o he inclusion/exclusion c i e ia used. Thus, hei pa ien s end o be younge and o ha e a slowe disease p og ession, in addi ion o ha ing mo e isi s egis e ed han he usual clinical pa ien s. To add ess his p oblem, hey p oposed a inal expe imen applying he Domain Adap a ion app oach o de elop p edic i e models using he PRO-ACT da a and imp o e hei pe o mances using pa ien clinical da a. Fi s ly, LSTM and Mul ilaye Pe cep on models we e ained using only da a om he PRO-ACT. Then, he aining phase was complemen ed using he TASMC da a o ine- une he models o he clinical da a. The esul s demons a ed ha he use o domain adap a ion imp o ed he p edic i e pe o mance o bo h models. G ollemund e al. (2020) p esen ed a dimensionali y educ ion model o p edic 1-yea su i al a es. The bioma ke s analyzed we e gende , si e onse , age, weigh , disease du a ion, ALSFRS sco es, ALSFRS slopes, and i died o no a e one yea . They F on ie s in Compu e Science | www. on ie sin.o g 8Ap il 2022 | Volume 4 | A icle 869140 Papaiz e al. Machine Lea ning Applied o ALS P ognosis combined da a om ou da ase s (PRO-ACT, T ophos, Exonhi , and Pa is Te ia y Re e al Cen e ), o aling 5,220 samples. The ob ained da ase was u he di ided in o de elopmen and alida ion se s. A e , he high-dimensional da a om he de elopmen se we e educed and p ojec ed on o 2D space h ough he Uni o m Mani old App oxima ion and P ojec ion (UMAP) algo i hm. Thus, he au ho s we e able o p ojec in o ma ion abou he pa ien s in o a 2D g aph. The 2D da a we e di ided in o h ee 1-yea su i al p obabili y zones: High (90%), In e media e (80%), and Low (58%). Then, he alida ion se was used o e alua e he p oposed model, and he esul s we e compa ed wi h he Random Fo es and he Logis ic Reg ession models. The UMAP model ob ained be e classi ica ion esul s (F1 sco e: 96%, Balanced Accu acy: 91%) when compa ed o he a e age esul s o he o he models (F1 sco e: 50%, Balanced Accu acy: 60%). The adop ed app oach also helped iden i y he bioma ke s wi h highe o lowe co ela ion wi h he su i al p edic ion. Fo example, he age and ALSFRS sco e p esen ed a high co ela ion, while he gende and weigh showed a low co ela ion. Howe e , he o al comp ehension o he ela ionship be ween inpu and ou pu a iables canno be ob ained because he adop ed model is conside ed a black-box app oach, which deg ades i s in e p e abili y. Despi e G eco e al. (2021) aimed o ind blood analy es o dis inguish pa ien s who ha e ALS om hose wi h Lowe Mo o Neu on Disease (LMND), hey also s udied he classi ica ion o hese pa ien s wi h ela ion o hei disease p og ession a es (High o Low). They analyzed clinic, demog aphic, and blood (108 analy es) da a om 41 ALS pa ien s. An SVM model was de eloped, and he Recu si e-Fea u e-Elimina ion algo i hm was used as a ea u e selec ion me hod. This model ob ained an accu acy o 87.25% in classi ying ALS pa ien s in o he High and Low g oups using he i s 16 anked analy es, indica ing he po en ial o using blood da a as p edic o bioma ke s. Ele a ed le els o Vi amin-B12, To al Choles e ol, and HDL we e ela ed o a highe disease p og ession a e. Leão e al. (2021) p oposed a p edic i e model based on Dynamic Bayesian Ne wo ks (DBN), including bo h s a ic and longi udinal da a. They accessed da a om he Po uguese ALS da ase (n=1,214), and he a ge p edic ion was he disease p og ession (ALSFRS sco e and subsco es) ela ed o he need o NIV suppo . To be p ocessed by he DBN model, he longi udinal da a we e con e ed in o ime-se ies da a and hen di ided in o Be o e NIV and A e NIV subse s. Thus, hey we e able o de e mine he mos ele an bioma ke s ela ed o hese wo essen ial disease s ages. The au ho s de eloped a p edic i e model, e med s dDBN amewo k, which uses s a iona y DBNs o p edic disease p og ession and non-s a iona y DBNs o de e mine how he bioma ke s analyzed change o e ime in each subse . The a e age esul s o p edic ing disease p og ession we e abo e 80% o bo h subse s ega ding he Accu acy, Sensi i i y, and AUC me ics, demons a ing he po en ial o he p oposed me hodology. G aphs we e gene a ed o isualize how he bioma ke s change o e ime, displaying hei alues in di e en ime s eps o each s age (be o e and a e NIV). This app oach allowed iden i ying some in e es ing ela ionships, as ollowing men ioned. The Maximum Expi a o y P essu e (MEP) was conside ed he mos impo an espi a o y exam o p edic he pa ien en ila o y decline be o e he need o NIV suppo . The ALSFRS Bulba subsco e had mo e in luence on disease p og ession a e NIV han be o e NIV. The BMI and Disease Du a ion had a s onge in luence han he o he s a ic bioma ke s o bo h subse s. 4. DISCUSSION This s udy sys ema ically e iewed he li e a u e o iden i y ele an s udies ha used ML app oaches o assis ALS disease p ognosis. As explained be o e in Sec ion 2, we ocused on hose s udies comp ising bioma ke s commonly p esen in he daily ALS clinical p ac ice. We iden i ied 10 s udies and de ailed hei he a ge p edic ions, bes ML algo i hm, pe o mance, da ase s, samples size, echniques, alida ion s a egies, bioma ke s e alua ed, and he mos ele an bioma ke s iden i ied. 4.1. ALS Da ase s and Da a P ep ocessing No ably, he s udies accessed da ase s ha concen a e ALS pa ien s om Eu ope and he Uni ed S a es o Ame ica. Da a om o he egions we e no analyzed (e.g., Sou h Ame ica, A ica, o Asia). We conside his analysis essen ial o con i m (o no ) i he p edic i e ML solu ions can be b oadly gene alized and i di e en da ase s can be combined o compose an e en mo e ele an ALS da ase . Mos o he s udies (60%) analyzed da a om he PRO-ACT da ase . PRO-ACT is he la ges public ALS da ase a ailable, con aining o e 10,000 samples, se ing as a basis o se e al s udies on ALS disease, and sui able o de eloping ML solu ions. Howe e , some s udies included ad ised ha he PRO-ACT has limi a ions ha can inc ease he isk o c ea ing biased models (Tang e al., 2019; G ollemund e al., 2020; Hadad and Le ne , 2020). P e ious s udies also epo ed hese PRO-ACT limi a ions, and he isk o i does no ep esen he clinical pa ien popula ion due o he inclusion and exclusion c i e ia used in he clinical ials (Chio e al., 2011; A assi e al., 2014). Fo ins ance, hei pa ien s end o be younge and p esen ewe unc ional impai men s. In his sense, using a alida ion s a egy ha includes an ex e nal da ase ep esen s an al e na i e o dec ease bias isk and achie e a mo e eliable ML algo i hm e alua ion. This s a egy was u ilized by Hadad and Le ne (2020) and G ollemund e al. (2020).Hadad and Le ne (2020) c ea ed a aining da ase combining samples om he PRO-ACT (100%) and Tel A i (90%) da ase . The samples emaining (10%) o he Tel A i da ase we e used o es he model. G ollemund e al. (2020) pe o med he alida ion using he Pa is da ase , which was no used in he aining and es ing s ages. P e e ably, he ex e nal da ase should con ain da a om he clinical pa ien popula ion. When designing ML solu ions, we need o be awa e o issues ha can a ec he pe o mance and eliabili y o he model, such as missing alues o da a imbalance. The PRO- ACT da ase p esen ed a conside able amoun o missing alues wha caused ha only 32% o i s samples could be used in p ac ice. Thus, i is aluable o e alua e how he missing da a impu a ion me hods can help o inc ease he sample size. an de Bu gh e al. (2017) and Tang e al. (2019) used a mo e F on ie s in Compu e Science | www. on ie sin.o g 9Ap il 2022 | Volume 4 | A icle 869140