Ci a ion: Amo-Salas, J.; Oli a es-Gil, A.;
Ga cía-Bus illo, Á.; Ga cía-Ga cía, D.;
A naiz-González, Á.; Cubo, E.
Compu e Vision o Pa kinson’s
Disease E alua ion: A Su ey on
Finge Tapping. Heal hca e 2024,12,
439. h ps://doi.o g/10.3390/
heal hca e12040439
Academic Edi o : Tin-Chih Toly Chen
Recei ed: 22 Decembe 2023
Re ised: 1 Feb ua y 2024
Accep ed: 6 Feb ua y 2024
Published: 8 Feb ua y 2024
Copy igh : © 2024 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/).
heal hca e
Re iew
Compu e Vision o Pa kinson’s Disease E alua ion: A Su ey
on Finge Tapping
Ja ie Amo-Salas 1, Alicia Oli a es-Gil 1, Ál a o Ga cía-Bus illo 2, Da id Ga cía-Ga cía 1,
Ál a A naiz-González 1,* and Es he Cubo 3
1Escuela Poli écnica Supe io , Depa amen o de Ingenie ía In o má ica, Uni e sidad de Bu gos,
09001 Bu gos, Spain; dga [email p o ec ed] (D.G.-G)
2Facul ad de Ciencias de la Salud, Depa amen o de Ciencias de la Salud, Uni e sidad de Bu gos,
09001 Bu gos, Spain; [email p o ec ed]
3Se icio de Neu ología, Hospi al Uni e si a io de Bu gos, 09006 Bu gos, Spain
*Co espondence: [email p o ec ed]
Abs ac : Pa kinson’s disease (PD) is a p og essi e neu odegene a i e diso de whose p e alence has
s eadily been ising o e he yea s. Specialis neu ologis s ac oss he wo ld assess and diagnose pa ien s
wi h PD, al hough he diagnos ic p ocess is ime-consuming and a ious symp oms ake yea s o appea ,
which means ha he diagnosis is p one o human e o . The pa ial au oma iza ion o PD assessmen
and diagnosis h ough compu a ional p ocesses has he e o e been conside ed o some ime. One
well-known ool o PD assessmen is inge apping (FT), which can now be assessed h ough compu e
ision (CV). A i icial in elligence and ela ed ad ances o e ecen decades, mo e speci ically in he
a ea o CV, ha e made i possible o de elop compu e sys ems ha can help specialis s assess and
diagnose PD. The aim o his s udy is o e iew some ad ances ela ed o CV echniques and FT so as
o o e insigh in o u u e esea ch lines ha echnological ad ances a e now opening up.
Keywo ds: Pa kinson’s disease; inge apping; machine lea ning; compu e ision
1. In oduc ion
In 2015, he Global Bu den o Disease s udy es ima ed ha neu ological diso de s a e
he leading cause o disabili y wo ldwide. The incidence and p e alence o neu odegene a-
i e diseases such as Pa kinson’s disease (PD) inc ease conside ably wi h age. PD is he
second mos common neu odegene a i e disease wo ldwide. Acco ding o he li e a u e,
he numbe o cases es ima ed be ween 1990 and 2015 doubled, a ec ing 6.2 million people
wo ldwide, a igu e ha is likely o double again by 2040 [1].
PD is cha ac e ized by p og essi e p ima y mo o disabili ies ollowing he degene a-
ion o he dopamine gic neu ons loca ed in he subs an ia nig a and associa ed a eas o he
b ain [
2
]. The pa hophysiology o PD in ol es signa u e abno mali ies in se e al pa allels
and la gely seg ega ed basal ganglia halamoco ical ci cui s (i.e., he mo o ci cui ). The
a ailable e idence sugges s ha he a ied mo emen diso de s esul ing om dys unc-
ions wi hin ha ci cui esul om he p opaga ed dis up ion o downs eam ne wo k
ac i i y in he halamus, co ex, and b ains em, and neu o ansmi e s, including dopamine,
ace ylcholine, no ad enaline, and se o onin [3,4].
Al hough PD is mainly associa ed wi h mo o symp oms, cha ac e ized by he p es-
ence o mo o asymme y wi h b adykinesia (slowness), igidi y, es ing emo , and gai
di icul ies wi h pos u al ins abili y [
5
], i is also accompanied by o he non-mo o symp-
oms such as cogni i e impai men , beha io al dis u bances, sleep diso de s, hyposmia
o au onomic dys unc ion, among o he s [
6
,
7
]. As a esul , PD is a highly he e ogeneous
disease, bo h wi h espec o i s symp oms and i s p og ession o e ime [8].
The e is no cu e o PD, bu pha macological and non-pha macological ea men s a e
a ailable, p o iding symp oma ic elie and imp o ing he quali y o li e. In ha ega d,
Heal hca e 2024,12, 439. h ps://doi.o g/10.3390/heal hca e12040439 h ps://www.mdpi.com/jou nal/heal hca e
Heal hca e 2024,12, 439 2 o 14
le odopa is he mos e ec i e medica ion a ailable o ea ing he mo o symp oms o PD,
bu in ce ain ins ances, i can be associa ed wi h o he dopamine gic and non-dopamine gic
d ugs [
9
]. O he non-pha macological ea men s include deep b ain s imula ion o he
sub halamic nucleus o he in e nal globus pallidum and physical, occupa ional, and
neu opsychological in e en ions.
Ea ly diagnosis o PD has p o ound implica ions o pa ien s and hei amilies, and
despi e impo an ad ances, i is s ill a challenge [
10
]. Recen de elopmen s include he
alida ion o modi ied clinical diagnos ic c i e ia, he in oduc ion and es ing o esea ch
c i e ia o p od omal Pa kinson’s disease, and he iden i ica ion o gene ic sub ypes and a
g owing numbe o biological bioma ke s associa ed wi h Pa kinson’s disease isk [
10
]. In his
ega d, he In e na ional PD and Mo emen Diso de Socie y (MDS) has published clinical
c i e ia o he diagnosis o PD ha a e in ended o use in clinical esea ch and clinical p ac ice.
These c i e ia include wo le els o ce ain y: clinically es ablished PD (maximizing speci ici y,
bu wi h educed sensi i i y) and p obable PD (balancing sensi i i y and speci ici y) [11].
A p esen , he new MDS- e ised e sion o he Uni ied Pa kinson’s Disease Ra ing
Scale (MDS-UPDRS) is used o a e he p og ess o PD. I comp ises ou pa s: he non-
mo o aspec s o daily li e expe iences (Pa I); he mo o aspec s o daily li e expe iences
(Pa II); he mo o examina ion (Pa III); and, inally, he mo o complica ions (Pa IV) [
12
].
Howe e , adminis e ing he ull scale is e y ime-consuming, so in o de o op imize ime
in clinical p ac ice, se e al au ho s ha e ied o de elop sho e a ing scales o o use
speci ic i ems om Pa III o he MDS-UPDRS [13,14].
Among he mo o ea u es associa ed wi h PD, b adykinesia, cha ac e ized by hy-
pokinesia (i.e., educed mo emen ampli ude, hesi a ions/hal s, and sequence e ec ) has
a signi ican impac on PD- ela ed disabili y. Finge apping (FT), one i em o he mo o
examina ion included in he MDS-UPDRS, is a es in which he pa ien is asked o ap
hei index inge on hei humb as apidly as possible, sepa a ing bo h inge s as much as
possible. FT seems o be one o he mos sensi i e i ems, so i can be used o c ea e a as
clinical judgmen o mo o s a us. Consequen ly, he FT es could po en ially be used as he
gold s anda d o ideo-based analysis [
15
]. Howe e , in e p e ing objec i e b adykinesia
da a, ob ained h ough kinema ic echniques, is an especially challenging ask, pa icula ly
when u ilized o diagnos ic pu poses [16].
In ecen yea s, he e ha e been emendous de elopmen s in he ield o echnologies
which, coupled wi h he imp o ed capabili ies o machine lea ning (ML) algo i hms, has
led o inc eased esea ch ac i i y on he au oma ic moni o ing o PD mo o symp oms, he
moni o ing o he hands being o pa icula in e es [
17
]. Recen publica ions ha e explo ed
he use o AI in he diagnosis, p og ession, and assessmen s o PD mo o and non-mo o
symp oms; howe e , he e is limi ed esea ch on he applica ion o AI o ideo mo ion analyses.
In his e iew, he aim is o highligh de eloping uses o AI-based echnology o ideo mo ion
analysis o hand mo emen s, so as o acili a e he diagnosis, managemen , and empowe men
o pa ien s and o supe ise he p og ession o hei disease and hei esponse o medica ion.
A e his in oduc ion, he emaining sec ions o his pape a e as ollows. Fi s ly,
ela ed wo ks wi h au oma ic PD diagnosis a e e iewed in Sec ion 2. Secondly, all he
pape s on compu e ision (CV) o PD diagnosis a e ca e ully summa ized in Sec ion 3.
Finally, he discussion o he esul s is p esen ed in Sec ion 4, and he main conclusions and
u u e lines o esea ch in Sec ions 5and 6 espec i ely.
2. Rela ed Wo ks
O e he pas decade, esea che s ha e been ying o de ine a use ul and e icien
me hod o PD assessmen . This assessmen could include diagnos ic o /and PD a ings
in acco dance wi h UPDRS le els. As p e iously no ed, he ocus o his pape is on
publica ions ha use CV and FT; ne e heless, in he cu en sec ion, some o he app oaches
a e co e ed in a b ie e iew.
One o he ini ial a emp s a diagnosing PD using CV includes he handw i en aces
analysis [
18
]. In ha es , pa ien s ha e o ill ou spi als and meande s on a piece o pape .
Heal hca e 2024,12, 439 3 o 14
Subsequen ly, he empla e and he d awings a e iden i ied and au oma ically spli using
image p ocessing echniques; bo h we e compa ed o ea u e ex ac ion. Finally, o PD
diagnosis (i.e., bina y classi ica ion), hey used adi ional classi ie s: Naï e Bayes, Op imum-
Pa h Fo es , and he Suppo Vec o Machine (SVM) algo i hm wi h a Radial Basis Func ion.
His o ically, ano he common app oach o PD assessmen has been he use o ex e nal
wea ables [
19
] and senso s [
20
]. These de ices a e capable o eco ding mo emen - ela ed
da a in e ec i e and accu a e ways, ye hey a e a ely used and a e e y expensi e. I
is a e y in e es ing line o esea ch, due o he eliabili y o he da a cap u e me hods
ha can ex ac solid ea u es on he basis o a pa ien ’s mo emen s. In bo h cases [
19
,
20
],
adi ional classi ie s, such as SVMs,
k
-Nea es Neighbo s, and Decision T ees we e used o
PD diagnos is (i.e., classi ica ion). Jeon e al. [
19
] esea ched he use o a w is wa ch- ype
wea able de ice wi h an accele ome e and a gy oscope o cap u e PD pa ien s’ mo emen
(accele a ion, angula eloci y, displacemen , and angle) and, a e he classi ica ion p ocess,
he au ho s pe o med a compa ison o he UDPRS a ing assigned by wo neu ologis s.
Rela ed o Moshko a e al. [
20
], hey ocused hei a icle on cap u ing hand mo emen
signals om a LeapMo ion senso , which was placed a a dis ance o 15–30 cm om he
pa ien ’s hand. Thei main aim was o pe o m PD assessmen based on h ee di e en hand
mo emen s: FT, p ona ion–supina ion o he hand, and opening–closing hand mo emen s.
In some s udies [
21
–
23
], PD se e i y has been assessed h ough ideos, hanks o
he echnological ad ances wi h CV sys ems. Fo example, Zhang e al. [
21
] ocused
on emo s, while Lu e al. [
22
,
23
] sough o quan i y he se e i y o PD h ough he
analysis o ideos showing pa ien s pe o ming MDS-UPDRS. Zhang e al. [
21
] ocused
hei esea ch in analyzing emo se e i y. Fo his pu pose, hey used OpenPose [24] o
ex ac ing 2D skele on ea u es; a e his s ep, classi ica ion was pe o med using a g aph
neu al ne wo k wi h a spa ial a en ion mechanism. They also compa ed he esul s using
o he s anda d classi ie s such as decision ee, con olu ional neu al ne wo k, and SVM.
Meanwhile,
Lu e al. [22,23]
de eloped hei own classi ie , called O dinal Focal Double-
Fea u es Double-Mo ion Ne wo k. In he p e ious pape [
22
], hey only included gai
analysis, bu in he second one [
23
], FT was also e alua ed. Fo gai analysis, hey used
VIBE [
25
] ( ideo in e ence o human body pose and shape es ima ion) om ex ac ing he
3D skele on; o FT, hey used he OpenPose [24] de ec ion sys em.
A mo e ecen a icle [
17
] mus also be highligh ed, which demons a ed he excellen
co ela ion be ween da a ex ac ed using CV (mo e speci ically MediaPipe [
26
]) and da a
cap u ed wi h hand-held accele ome e s. Tha s udy showed ha non-in usi e me hods
such as CV can ex ac simila da a o physical de ices. Williams e al. [
27
] sough o p o e
a co ela i e ela ionship when assessing whe he o no sma phone ideo eco dings
could be enough o e alua ing b adikynesia. S anda d sma phone ideo eco dings
o pa ien s pe o ming FT we e acked wi h DeepLabCu [
28
]. Th ee ea u es such as
apping speed, ampli ude, and hy hm we e co ela ed wi h clinical a ings made by
22 mo emen
diso de neu ologis s using he Modi ied B adykinesia Ra ing Scale (MBRS)
and he Mo emen Diso de Socie y e ision o he MDS-UPDRS.
Wi h a simila pu pose in mind, Jabe e al. [
29
] sough o show how CV can be
a sui able amewo k o PD assessmen . The esea ch showed a way o cap u ing FT
mo emen s using CV lib a ies (i.e., LabelImg and YOLO [
30
]) and ans o ming hem in o
aluable me ics and ea u es ha can help wi h PD diagnos ics.
Addi ionally, wi h he ise and democ a iza ion o In o ma ion Technology (IT), o he
esea che s ha e app oached PD assessmen on he basis o IT in e ac ions; o example,
using mobile de ices [
31
,
32
] and web b owse s [
33
]. In bo h app oaches, PD pa ien s
a e in i ed o en e a gami ied si ua ion and a e p omp ed o pe o m mo emen s, ei he
by apping di ec ly on he mobile phone sc een [
31
,
32
] o by acking mouse ope a ions
and keyboa d inpu s on he web b owse [
33
]. Based on hy hm, accu acy, a igue, and
eac ion ime, among o he ac o s, da a a e ga he ed o use as inpu o machine lea ning
classi ie s pe o ming an assessmen o PD. Rega ding mobile phone usage, some o hese
in es iga ions also ook pa ien oice eco dings in o conside a ion [32].
Heal hca e 2024,12, 439 4 o 14
Finally, he use o a CV amewo k o he diagnosis o mo emen diso de s [
34
] o
speci ically PD [
35
] mus be men ioned. In bo h pape s, he way ha deep-lea ning-based
ma ke less mo ion acking echniques can imp o e PD diagnosis and assessmen we e
highligh ed. Tien e al. [
34
] e iewed and discussed he po en ial clinical applica ions and
echnical limi a ions o hese echniques, wi h a ocus on DeepLabCu [
28
]. To e alua e
he use o DeepLabCu o au oma ed mo emen diso de disease assessmen , hey buil a
mobile ame wi h h ee synch onized came as o eco ding hand mo emen s, including
heal hy con ol subjec s and mo emen diso de pa ien s wi h a ious diagnoses, such as
PD and essen ial emo . In his case, au ho s sha e he u ili y o DeepLabCu , men ion-
ing h ee ongoing s udies, bu wi hou p o iding addi ional de ails. On he o he hand,
Sibley e al. [35]
published an a icle whe ein hey e iewed he echniques, so wa e li-
b a ies, and comme cial app oaches o ideo analyses o PD. Addi ionally, hey iden i ied
challenges and possible solu ions associa ed wi h a ing mo o symp oms o PD using ideo.
As i has been p e iously no ed, adi ional classi ie s ha e been e alua ed his o ically
o app aising models’ pe o mance and o ca ying ou s a is ical compa ison wi h p e i-
ous a icles. To p o ide a summa y iew, Table 1shows he classi ie s which ha e been
used in he a icles men ioned in his sec ion. A e a i s insigh , decision ee, andom
o es , and SVM (in i s di e en a ia ions) seem o be he mos used classi ie s ac oss he
dissec ed a icles. The use o ad hoc algo i hms by some au ho s is also wo h no ing.
Table 1. Summa y o he ela ed wo ks and he classi ie s used in each s udy. When mo e han one
classi ie was used, he one ha achie ed he bes pe o mance is highligh ed in bold (SVM: suppo
ec o machine).
Re . Yea B ie Desc ip ion Classi ie s
[18] 2016 Handw i en aces Naï e Bayes
Op imum-Pa h Fo es
SVM-R
[19] 2017 T emo se e i y analysis using a w is wa ch- ype
wea able de ice
Decision ee
SVM-L
SVM-P
SVM-R
k-nea es neighbo s
Disc iminan Analysis
[31] 2018 On-sc een apping on a mobile phone (iPhone app)
Logis ic eg ession
Random o es
Deep Neu al Ne wo k
Con olu ional Neu al Ne wo k
[32] 2018 PD- ela ed ac i i ies ( oice, inge apping, gai ,
balance, and eac ion ime) assessmen (And oid app) Machine-Lea ning algo i hm [36]
[20] 2019 Hand mo emen s da a cap u ing using
LeapMo ion senso
Decision ee
SVM
k-nea es neighbo s
Random o es
[22] 2020 Gai analysis using ideos O dinal Focal Double-Fea u es Double-Mo ion Ne wo k
[23] 2021 Gai and FT analysis using ideos O dinal Focal Double-Fea u es Double-Mo ion Ne wo k
[21] 2022 T emo se e i y analysis using ideos
G aph Neu al Ne wo k
Decision ee
Con olu ional Neu al Ne wo k
SVM
[33] 2023
Gami ied websi e acking keyboa d and mouse inpu s
Random o es
Decision ee
SVM
Mul ilaye pe cep on
3. Finge Tapping and Compu e Vision on Pa kinson’s Disease E alua ion
In ecen yea s, sui able me hods ha e been p oposed in se e al wo ks o imp o ing
he diagnosis and/o he PD a ing using CV and FT. In his sec ion, he mos in e es ing
Heal hca e 2024,12, 439 5 o 14
s udies on CV and FT a e b ie ly e iewed and compa ed. To do so, a s a e-o - he-a e iew is
conduc ed wi h pape s published no ea lie han 2014 (e.g., [
37
]). As shown in Table 2, mos o
he pape s we e w i en o e he pas 4 yea s. The e a e wo main easons o his: i s o all,
he imp o emen in CV- ela ed de ices (came as, sma phones, e c.) o e s use ul capabili ies
a a o dable cos s; secondly, he inc easingly e ec i e pe o mance and accu acy o de ec ion
and pose es ima ion lib a ies [24,26,38] also make hem e ec i e choices.
All he s udies sha e common poin s (see Figu e 1): i s o all, he humb and he
index inge s a e au oma ically iden i ied o pe o m ea u e ex ac ion, be o e a da ase is
compiled. Then, one o mo e algo i hms a e ained o p oduce a model, which can inally
be used o es ing wi h unseen ins ances/examples.
Shoo ing and collec ion
o ideo ins ances
Landma k de ec ion by
compu e ision
Fea u e
ex ac ion
- Ampli ude
- F equency/Rhy hm
- Speed
- Fa igue
- e c.
Fo each
ame
Cons uc ion o
ain and es
da ase s
Machine
Lea ning
model
T aining and
alida ion
Figu e 1. Gene al p ocess o he de ec ion o PD using CV and FT.
Mos o he wo ks a e in ended o sol e classi ica ion p oblems: a ew o hem ace
he simples p oblem (bina y classi ica ion, i.e., ei he has PD o no PD), whe eas mos o
hem a e designed o p edic a class among mo e han wo alues (mul iclass classi ica ion);
o ha pu pose, he common app oach is o p edic he UPDRS a ing. The dis ibu ion
be ween classes is une en (bo h o bina y and mul iclass classi ica ion pape s), and a
summa y o he numbe o ideos o each class is shown in Figu e 2. I mus be no ed ha
UPDRS 3 and 4 le el se e i y a e especially unde ep esen ed in mos o he wo ks.
The main cha ac e is ics o he esea ch pape s unde analysis a e summa ized in Table 2.

Heal hca e 2024,12, 439 6 o 14
Table 2. Summa y o he main cha ac e is ics o he compu e ision (CV)- ela ed pape s e iewed in his esea ch. In he s udies wi h mo e han one classi ie , he one ha
achie ed he bes pe o mance is highligh ed in bold (PD: Pa kinson’s disease, HC: heal hy con ol, AUC: a ea unde he ROC, SVM: suppo ec o machine).
Re . Yea Cap u e Finge Iden i ica ion Classi ie s ML P oblem Da ase Pe o mance Measu es
[37] 2014 2D OpenCV SVM-PUK Bina y Mul iclass (3) 13 PD
6 HC Accu acy
[39] 2019 2D Con olu ional Neu al Ne wo k
Naï e Bayes
Logis ic eg ession
SVM-L
SVM-R
Bina y 20 PD
15 HC
Accu acy
Sensi i i y
Speci ici y
AUC
[40] 2019 3D Cus om-made acke s A i icial neu al ne wo k
SVM Bina y 16 PD
14 HC
Accu acy
Sensi i i y
Speci ici y
[41] 2021 2D OpenPose Con olu ional Neu al Ne wo k Mul iclass (5) 157 PD
0 HC
Accu acy
AUC
P ecision
Recall
F1-sco e
[42] 2021 2D Single Sho Mul iBox De ec o (SSD) + OpenPose
Logis ic Reg ession
Naï e Bayes
Random o es Bina y 22 PD
20 HC Accu acy
AUC
[43] 2021 2D OpenPose SVM-R Mul iclass (5) 55 PD
0 HC Weigh ed κ
In aclass co . coe .
[44] 2022 2D MMPose Deep Neu al Ne wo k Mul iclass (4) 300 PD
0 HC
P ecision
Recall
F1-sco e
[45] 2022 2D MediaPipe Fully Connec ed Ne wo k Mul iclass (5) 93 PD
27 HC
Accu acy
P ecision
Recall
F1-sco e
[46] 2022 3D Spa ial-Tempo al Ancho - o-Join Reg ession Ne wo k
(ST-A2J)
k-nea es neighbo s
Random o es
XGBoos
SVM-L
SVM-R
Mul iclass (5) 48 PD
11 HC Accu acy
[47] 2023 3D MediaPipe
k-nea es neighbo s
Random Fo es
XGBoos
SVM
Bina y 35 PD
60 HC
Accu acy
P ecision
Recall
F1-sco e
Heal hca e 2024,12, 439 7 o 14
123
163
101
0 0
70
397
196
67
14
917 25 32 27
99
247 229
36
0
54
115
52 30
2
0
26 19 38
7
0
50
100
150
200
250
300
350
400
UPDRS 0 UPDRS 1 UPDRS 2 UPDRS 3 UPDRS 4
Numbe o ideos
Se e i y
Khan e al., 2014 Li e al., 2021 Pa k e al., 2021 Yang e al., 2022 Li e al., 2022 Guo e al., 2022
Figu e 2. Ba plo wi h he numbe o ideos o each class (se e i y o UPDRS) o he pape s ha pe o m
mul iclass classi ica ion. Re e ences o he pape s p esen ed in he igu e: Khan e al. [
37
],
Li e al. [41]
,
Pa k e al. [43], Yang e al. [44], Li e al. [45], and Guo e al. [46].
3.1. Fea u e Ex ac ion
As migh be expec ed, he p e e ence o deep neu al ne wo ks o CV [
48
] mean
ha hey we e used o inge iden i ica ion in all
he s udies.
The mos popula ones a e
lis ed below:
•
Mediapipe [
26
] is an open-sou ce amewo k, de eloped by Google, which p o ides
eal- ime p ocessing o mul imedia da a, including ideo and audio. I includes se e al
modules o CV asks, including pose es ima ion, ace de ec ion, hand de ec ion, and
objec acking.
•
Openpose [
24
], de eloped by he Compu e Vision Cen e a he Au onomous Uni-
e si y o Ba celona, was eleased in 2016. I is a eal- ime mul i-pe son human pose
de ec ion lib a y wi h he capabili y o join ly de ec ing he human body, oo , hand,
and acial keypoin s on single images.
•
MMPose [
38
] is an open-sou ce oolbox o pose es ima ion based on PyTo ch. I
suppo s: mul i-pe son human pose es ima ion, 133 keypoin whole-body human
pose es ima ion, hand pose es ima ion, and 3D human mesh eco e y.
The inc easing use o Mediapipe mus be no ed in some o he mos ecen s udies ha
we e e iewed [
17
,
45
,
47
], eplacing OpenPose [
24
], which was he mos widely used in
p e ious yea s [41–43].
The mos common ea u es used in he esea ch pape s a e summa ized in Table 3. One
o he s udies [
41
] was in en ionally excluded om he able, due o a lack o in o ma ion
on he opic. Mo eo e , ea u es used in no mo e han one pape we e no included. A his
poin , i is impo an o men ion he linguis ic disc epancies be ween he no a ions in he
di e en wo ks, as well as he ag eemen s ha we e eached o he pu poses o his s udy
on co ec speci ica ion o he ea u es ha appea in he able.
•
Ampli ude and Speed: he wo mos common ea u es o be analyzed o PD as-
sessmen using FT. Ne e heless, hey a e no conside ed in qui e he same way in
all wo ks, al hough he e a e no seman ic di e ences ega ding he way ha hose
ea u es a e o be unde s ood in FT:
–Ampli ude: dis ance be ween humb and index inge s.
–Speed: ampli ude di e ence o e ime.
Fo example, a common app oach is o ob ain he alues du ing he ime se ies, bu
o he au ho s also compu e he mean o maximum alue, a maximum alue du ing
he opening o he closing phases, minimum, s anda d de ia ion, e c. In o he wo ds,
Heal hca e 2024,12, 439 8 o 14
once he ea u e is conside ed, se e al me ics could be ex ac ed, which will ob iously
di e ac oss he di e en s udies.
•
Fa igue: his ea u e is e alua ed in ew a icles, ye he app oaches used o i s
es ima ion a y. I should be no ed ha i is no a physical alue, such as ampli ude
and speed. The concep i sel is simila in he di e en a icles, bu essen ial nuances
in i s es ima ion we e iden i ied. Fo example:
–Di e ence be ween he highes and he lowes alues o ampli ude peaks [42].
–G adien in ampli ude acco ding o ime [43].
–O he au ho s [37] e alua ed a igue on he basis o di e en measu es:
*Di e ence be ween numbe o aps in wo ime slo s.
*Va ia ion coe icien (VC) in apping speed.
*
Di e ence be ween he a e age/VCs maximum ampli ude o inge aps in
wo ime slo s.
*VC in he maximum ampli ude o inge aps.
*Tapping accele a ion.
•
F equency/Rhy hm: wi hou a doub , he mos abs ac ea u e. Bo h concep s a e
used indis inc ly, bu no always o ep esen ing he same concep :
–
In some s udies [
17
,
27
], i s calcula ion is based on unde aking Fas Fou ie T ans o m.
–
Ano he common app oach [
37
,
39
] is o use a ea u e called “c oss-co ela ion
be ween he no malized peaks” (CCNP) o es ima ing consis ency and hy hm
in apping.
–
Buongio no e al. [
40
] used he a e aged alue o he di ision be ween he ampli-
ude peak eached in a single exe cise ial and he ime du a ion o he ial.
Table 3. Summa y o inge apping ea u es used in he CV pape s e iewed in his esea ch. The
symbol (✓) ep esen s ha he ea u e is used in he wo k.
Fea u e [37] [39] [40] [27] [42] [43] [29] [44] [45] [46] [17] [47]
Ampli ude ✓✓✓✓✓✓✓✓✓✓✓✓
F equency/Rhy hm ✓ ✓ ✓ ✓ ✓ ✓
Speed ✓ ✓✓✓✓✓✓✓✓ ✓
Fa igue ✓ ✓ ✓
3.2. Classi ie s
As p e iously men ioned, all he pape s ha we e e iewed had he common aim
o pe o ming a classi ica ion p edic ion (bina y o mul iclass). In o he wo ds, he main
a ge o he s udies was o implemen a PD p edic ion acco ding o UPDRS a ings o FT.
In no mo e han a couple o s udies [
37
,
44
] could he i e le els o UPDRS no be p edic ed,
due o he lack o enough examples du ing he aining phase.
Di e en classi ie s can be ained o pe o m he classi ica ion, once ea u e ex ac ion
has been comple ed, wi h SVM [
49
] (wi h i s di e en a ian s/ke nels) being by a he
mos popula [37,39,40,43,46,47].
B oadly used, especially in ecen yea s, deep neu al ne wo ks can also be applied as
classi ie s [
40
,
44
,
45
] (no only o inge iden i ica ion) wi h di e en con igu a ions and
a ia ions. Ad hoc designs and deploymen s ha e e en been p oposed [41].
On he o he hand, mul i-classi ie s [
50
] (a.k.a. ensembles) a e popula , he mos
widely used being andom o es (RF) [
42
,
46
,
47
,
51
] and XGBoos [
46
,
47
,
52
]. Las bu no
leas , con en ional classi ie s (such as Naï e Bayes [
39
,
42
],
k
-nea es neighbo s [
46
,
47
], and
logis ic eg ession [39,42]) ha e commonly been used as baselines.
No insigh could be gi en in o which classi ie was he bes o he ask o FT classi i-
ca ion wi hou u he expe imen a ion, due o he di e ences be ween he expe imen al
se ups, he da ase s, he classi ica ion asks, and he classi ie s ha we e used.
Heal hca e 2024,12, 439 9 o 14
3.3. Da ase s
One o he main p oblems ha se e ely complica es compa isons o he p oposals is
he lack o benchma k da ase s. A pa icula da ase is used in e e y single s udy, usually
con aining small numbe s o indi iduals: he smalles included 11 while he la ges had
300. I mus be no ed ha i can be ex emely di icul o assess he pe o mance o he
p oposals wi h such small-sized samples.
Deep analysis o his opic e eals g ea a iabili y. Commonly, mos o he esea ch pa-
pe s include PD pa ien s and heal hy con ols (HCs). The p oblem he e is ha , some imes,
he e a e la ge di e ences ela ed o class p opo ions. I is well known ha imbalanced
da ase s (i.e., when a class is unde - ep esen ed) a e challenging, as algo i hms will in a i-
ably igno e he unde ep esen ed class/es. The class p opo ion is usually measu ed by
means o he imbalance a io (IR) [53].
IR =
Numbe o majo i y
Numbe o mino i y
In his way, da ase s can be u he classi ied:
•
Fai ly balanced da ase s (
IR <
2): he p opo ions be ween Pa kinson’s disease
pa ien s and heal hy con ols a e balanced [27,39,40,42,47].
•
Imbalanced da ase s (
IR >
2): he p opo ions a e su icien o ake in o accoun he
imbalance p oblem [37,45,46].
In some s udies, only ideos om PD pa ien s we e aken in o conside a ion [
17
,
41
,
43
,
44
].
The e was no a emp o dis inguish be ween PD pa ien s and heal hy con ols, bu only
he di e ences be ween pa ien s we e conside ed, e.g., o a e hem acco ding o UPDRS.
3.4. Measu es
A e analyzing he a icles, simila pe o mance measu es based on ML asks we e
used. B ie explana ions appea below alongside no es on hei use in he pape s ha
we e e iewed.
•
Accu acy: by a he mos common measu e [
37
,
39
–
42
,
45
–
47
], he ounda ion and
common unde s anding o i s meaning is wha makes accu acy so popula (e.g., he
numbe o success ul ou comes di ided by he o al numbe o examples). E en a non-
amilia eade could de e mine he achie emen le el by in e p e ing he accu acy
pe cen age. Howe e , his measu e also has some d awbacks; a common complain
abou accu acy is ha i ails when he classes a e imbalanced.
•
A ea Unde ROC Cu e (AUC): also conside ed a popula measu e o classi ica ion
p oblems [
39
,
41
,
42
]. I is commonly used in ML and da a analy ics o assess he
pe o mance o models a p edic ing bina y ou comes and, in con as o s anda d
accu acy, i is pa icula ly use ul when p ocessing imbalanced da ase s, whe e accu a e
p edic ion o mino i y classes is o high impo ance.
•F1
-sco e [
41
,
44
,
45
,
47
] is he ha monic mean o p ecision ( he numbe o ue posi i e
di ided by he p edic ed as posi i es) and ecall ( he numbe o ue posi i es di-
ided by he numbe o all samples o he class o in e es ).
F1
-sco e gi es he same
impo ance o bo h p ecision and ecall, wha can be conside ed as i s main d aw-
back [
54
]. None heless, in eal-wo ld p oblems usually di e en cos s a e associa ed
o di e en e o s.
4. Discussion
Fi s o all, as p e iously s a ed, PD assessmen is s ill a majo challenge o clinical
neu ologis s. Few o he p oposed solu ions ha e been ce i ied by he Mo emen Diso de
Socie y o UPDRS a ing. Despi e he ac ha his ce i ica ion is mean o p o ide a
common amewo k and o objec i y a complex scena io, some oices a e skep ical o
i s gene alized use [
55
]. Fi s ly, p e ious knowledge is usually necessa y as well as a
case s udy wi h some clues on pa ien e olu ion, amily backg ound, and so on [
39
,
42
].