Assist-as-needed control for a smart orthotic system aiming a personalized gait rehabilitation

Uni e sidade do Minho
Escola de Engenha ia
Luís Ca los Rod igues Mo ei a
Assis -As-Needed Con ol o a Sma
O ho ic Sys em Aiming a Pe sonalized Gai
Rehabili a ion
Ou ub o de 2024
UMinho | 2024 Luís Ca los Rod igues Mo ei a Assis -As-Needed Con ol o a Sma O ho ic
Sys em Aiming a Pe sonalized Gai Rehabili a ion
i
i
Luís Ca los Rod igues Mo ei a
Assis -As-Needed Con ol o a Sma
O ho ic Sys em Aiming a Pe sonalized
Gai Rehabili a ion
Tese de Dou o amen o
P og ama Dou o al em Engenha ia Biomédica
T abalho e e uado sob a o ien ação de
P o esso a Dou o a C is ina Peixo o dos San os
P o esso Dou o João José Fe nandes Ca doso de A aújo
Ce quei a
Dou o a Joana So ia Campos Figuei edo
ou ub o de 2024
ii
DIREITOS DE AUTOR E CONDIÇÕES DE UTILIZAÇÃO DO TRABALHO POR TERCEIROS
Es e é um abalho académico que pode se u ilizado po e cei os desde que espei adas as eg as e
boas p á icas in e nacionalmen e acei es, no que conce ne aos di ei os de au o e di ei os conexos.
Assim, o p esen e abalho pode se u ilizado nos e mos p e is os na licença abaixo indicada.
Caso o u ilizado necessi e de pe missão pa a pode aze um uso do abalho em condições não p e is as
no licenciamen o indicado, de e á con ac a o au o , a a és do Reposi ó iUM da Uni e sidade do Minho.
A ibuição-NãoCome cial-SemDe i ações
CC BY-NC-ND
h ps://c ea i ecommons.o g/licenses/by-nc-nd/4.0/
iii
AGRADECIMENTOS
Após o é mino des e impo an e capí ulo, é ho a de ag adece .
Em p imei o luga , que o exp essa a minha p o unda g a idão aos meus pais. Eles p opo ciona am-me
apoio, ca inho, amo , con o o, es u u a e a o ça emocional necessá ia ao longo de oda es a
caminhada. São o meu po o-segu o e odo o sucesso académico, p o issional e pessoal de e-se a eles.
Ag adeço ambém à minha amília e amília da minha esposa pelo o gulho demons ado em cada e apa
do meu abalho e pelas conquis as alcançadas.
Um ag adecimen o aos meus o ien ado es (P o esso a C is ina, Dou o a Joana Figuei edo e Dou o João
Ce quei a) e ao co po clínico do Hospi al de B aga. À P o esso a C is ina, a minha g a idão po me
p opo ciona a opo unidade de abalha num p oje o ão en iquecedo e desa iado , do qual me o gulho
p o undamen e. Ag adeço pela o ien ação, disponibilidade, incen i o cons an e e po alimen a a minha
capacidade de sonha . À Dou o a Joana Figuei edo, ag adeço pela empa ia, amizade, pa ilha e
c escimen o pessoal e p o issional. A ua dedicação e amo pela causa são e iden es em udo o que
azes. Ob igado pelos alo es que me ansmi is e e po odo o apoio disponibilizado ao longo des a
jo nada.
Aos meus colegas de labo a ó io, ag adeço pelos momen os de pa ilha, ap endizagem, disponibilidade
e pelo companhei ismo. Um ag adecimen o especial aos amigos do Po a Abe a, que o na am es a
jo nada dou o al mais descon aída, eple a de aleg ias, e empe ada com o sabo das e as que nos
i am c esce .
Ag adeço ambém a odos os pa icipan es dos es udos expe imen ais que possibili a am a alidação das
ecnologias desen ol idas nes a ese. Em pa icula , o meu since o ag adecimen o ao S . João e à
Ma ga ida (e co esponden es amílias) pela dedicação, paciência, in e esse e disponibilidade
demons ados nas e apias conduzidas. Jun os, izemos ciência!
Um ag adecimen o ao melho naipe de cla ine es de Po ugal (Palhe as D’Ou o) e à Banda Musical de
Rio de Moinhos po odos os momen os musicais e de con a e nização.
À minha esposa, Ma iana Cos a, ag adeço po udo e po an o. A ua p esença ez des a jo nada algo
ainda mais especial e boni o. Ajudas e-me a nunca pe de o oco e os e essencial pa a aça e
conc e iza o caminho possí el que me le ou a a ingi es e obje i o.
Ob igado a odos! Jun os, con ibuímos pa a o a anço da ciência!

i
STATEMENT OF INTEGRITY
I he eby decla e ha ing conduc ed his academic wo k wi h in eg i y. I con i m ha I ha e no used
plagia ism o any o m o undue use o in o ma ion o alsi ica ion o esul s along he p ocess leading o
i s elabo a ion.
I u he decla e ha I ha e ully acknowledged he Code o E hical Conduc o he Uni e si y o Minho.
Con olo de assis ência con o me as necessidades pa a um sis ema o ó ico in eligen e
com is a à eabili ação pe sonalizada da ma cha
RESUMO
O aciden e ascula ce eb al (AVC) co esponde à e cei a p incipal causa de incapacidade
mo o a em adul os. A a és de e apias de eabili ação en ol endo um eino epe i i o e pe sonalizado
ao doen e, a eabili ação da ma cha é essencial pa a a ecupe ação da mobilidade. Pa a al, as o ó eses
a i as (OAs) e exosquele os êm sido apon ados como um meio pa a po encia os e ei os das e apias de
eabili ação. Con udo, é c ucial adap a a eabili ação da ma cha às necessidades e in enções de
locomoção de cada doen e, a a és do desen ol imen o de (i) no os senso es e algo i mos in eligen es
que pe mi am uma a aliação obje i a da ma cha; e (ii) es a égias de con olo pa a OAs que p omo am
uma e apia adap ada às necessidades de cada doen e, is o é,
assis -as-needed
(AAN). Es a ese em
como obje i o expandi o Sma Os, um sis ema o ó ico a i o, modula e es í el, pa a o e ece um eino
de ma cha adap ado às necessidades de doen es com AVC e a alia a locomoção a a és de dados
cinemá icos e muscula es.
A ese é es u u ada em cinco passos de in es igação. P imei o, oi expandida a es u u a
modula do Sma Os pa a in eg a no os sis emas senso iais, e amen as de análise da ma cha e
es a égias de con olo. Segundo, oi c iado um sis ema senso ial baseado em êx eis condu o es, capaz
de moni o iza , simul aneamen e, a a i idade muscula de á ios g upos muscula es. O benchma king
com sis emas come ciais demons ou o po encial des e sis ema pa a a a aliação obje i a da con ação
muscula . Te cei o, oi desen ol ida uma e amen a de
deep lea ning
pa a a descodi icação de modos
de locomoção, des acando-se pela sua classi icação p ecisa e an ecipa i a. Qua o, o am desen ol idas
ês es a égias de con olo AAN. A es a égia baseada em modos de locomoção o e ece assis ência de
aco do com as in enções de locomoção do u ilizado ; a es a égia baseada em ele omiog a ia con ibui
pa a o o alecimen o muscula e p omo e a pa icipação a i a do u ilizado na e apia; e a es a égia
baseada em ene gia isa adap a o con olo do disposi i o pa a eduzi o cus o me abólico do u ilizado .
Po im, o am conduzidos dois casos de es udo pa a de e mina os e ei os do sis ema Sma Os como
e amen a de eabili ação em pacien es que so e am um AVC. Em suma, os esul ados indicam que o
sis ema Sma Os es á ap o pa a aplicação em ambien e clínico, an o como uma solução pe sonalizada
de assis ência, como uma e amen a de a aliação da ma cha em pacien es que so e am um AVC.
Pala as-Cha e: assis ência e eabili ação da ma cha, descodi icação de modos de locomoção,
es a égias de con olo, exosquele os, in eligência a i icial, senso es es í eis.
i
ABSTRACT
S oke is he hi d leading cause o mo o disabili y in adul s. Th ough ehabili a ion he apies
in ol ing epe i i e and use -o ien ed aining, gai ehabili a ion is essen ial o egain mobili y. To his
end, ac i e o hoses (AOs) and exoskele ons ha e been iden i ied as a means o enhance he e ec s o
ehabili a ion he apies. Howe e , i is c ucial o adap gai ehabili a ion o he speci ic needs and
locomo ion in en ions o each pos -s oke pa ien h ough he de elopmen o (i) new senso s and
in elligen algo i hms ha allow an objec i e assessmen o gai ; and (ii) con ol s a egies o AO ha
p omo e he apy ailo ed o he needs o each pos -s oke pa ien , i.e., assis -as-needed (AAN). This Ph.D.
hesis aims o ex end Sma Os, an ac i e, modula , and wea able o ho ic sys em, o p o ide gai aining
adap ed o he needs o pos -s oke pa ien s and o assess locomo ion h ough kinema ic and muscula
da a.
The hesis is di ided in o i e esea ch s eps. Fi s ly, he modula s uc u e o Sma Os was
ex ended o in eg a e new senso y sys ems, gai analysis ools, and con ol s a egies. Secondly, a senso y
sys em based on conduc i e ex iles was de eloped, capable o simul aneously moni o ing he muscula
ac i i y o se e al muscle g oups. Benchma king wi h comme cial sys ems demons a ed he po en ial o
his sys em o objec i e assessmen o muscle con ac ion. Thi dly, a deep lea ning ool was de eloped
o decode locomo ion modes, which is cha ac e ized by i s accu a e classi ica ion and p edic i e na u e.
Fou h, h ee AAN con ol s a egies we e de eloped. The AAN Locomo ion Mode-d i en ajec o y con ol
p o ides assis ance acco ding o he use 's locomo ion in en ions. The AAN EMG-based con ol
con ibu es o muscle s eng hening and p omo es he use 's ac i e pa icipa ion in he apy. The AAN
Human-in- he-Loop con ol aims o adap he con ol o he de ice o educe he use 's ene gy expendi u e.
Finally, wo case s udies we e conduc ed o de e mine he e ec s o he Sma Os sys em as a
ehabili a ion ool o pos -s oke pa ien s. In sho , he esul s indica e ha he Sma Os sys em is sui able
o use in a clinical en i onmen , bo h as a pe sonalized assis ance solu ion and as a gai assessmen
ool o pa ien s who ha e su e ed a s oke.
Keywo ds: Gai assis ance and ehabili a ion, locomo ion mode decoding, con ol s a egies,
exoskele ons, a i icial in elligence, wea able senso s
ii
Table o Con en s
Lis o Figu es ...................................................................................................................x
Lis o Tables .................................................................................................................. xi
Lis o abb e ia ions and ac onyms ................................................................................ x i
1. In oduc ion .............................................................................................................. 1
1.1 Mo i a ion ................................................................................................................... 2
1.2 P oblem s a emen ....................................................................................................... 4
1.3 Goals ........................................................................................................................... 5
1.4 Resea ch ques ions ...................................................................................................... 8
1.5 Con ibu ions o knowledge.......................................................................................... 9
1.6 Publica ions ............................................................................................................... 11
1.7 Manusc ip ou line ..................................................................................................... 12
2. Li e a u e Resea ch ................................................................................................ 14
2.1 In oduc o y Insigh ................................................................................................... 15
2.2 Re iew o he E ec s o Lowe Limb Assis i e De ices on Pos -S oke Pa ien s ........... 17
2.2.1 Sea ch Me hodology ................................................................................................................. 17
2.2.2 Resul s ..................................................................................................................................... 18
2.2.3 Discussion ................................................................................................................................ 36
2.3 Conclusions ............................................................................................................... 39
3. Sma Os Sys em O e iew ...................................................................................... 40
3.1 In oduc o y Insigh ................................................................................................... 41
3.2 Concep ual Design and Func ionali ies ....................................................................... 41
3.2.1 Ac ua o s .................................................................................................................................. 43
3.2.2 Wea able Mo ion Lab ................................................................................................................ 43
3.2.3 Gai Analysis Tools .................................................................................................................... 45
3.2.4 Hie a chical Con ol A chi ec u e ............................................................................................... 46
3.2.5 G aphical Applica ion ................................................................................................................ 47
3.3 Conclusions ............................................................................................................... 49
4. MuscLab Sys em ..................................................................................................... 50
4.1 In oduc o y Insigh ................................................................................................... 51
4.2 C i ical Analysis o Rela ed Wo k ................................................................................ 52
4.3 Me hodology .............................................................................................................. 54
xi
Lis o Tables
Table 2.1 – Technical de ails ex ac ed om he selec ed s udies (A, P, and NI means ac i e, passi e,
and no indica ed, espec i ely) ......................................................................................................... 21
Table 2.2 – Clinical de ails ex ac ed om he selec ed s udies .......................................................... 30
Table 3.1 – Di e en ypes o da a collec ed by each senso sys em .................................................. 44
Table 3.2 – Di e en con ol s a egies a ailable in he Sma Os sys em ............................................ 47
Table 4.1 – MuscLab equi emen s................................................................................................... 54
Table 4.2 – A e age delay (s anda d de ia ion) be ween he MuscLab and EMG Delsys signals, and
MuscLab and Xsens Awinda signals, in milliseconds, o each mo ion cadence (40, 70, 105 bpm) .... 66
Table 4.3 – Spea man Co ela ion coe icien (
) be ween he MuscLab and EMG sys ems. A
p
- alue below
0.05 was e i ied in all co ela ions ................................................................................................... 67
Table 5.1 – Iden i ica ion o he ype o da a used o decode LMs. The join angles, magni ude o 3D aw
accele ome e and gy oscope, segmen angles, and EMG da a a e shaded in g een, yellow, g ay, and ed,
espec i ely ...................................................................................................................................... 78
Table 5.2 – Iden i ica ion o he condi ions o eal- ime es s pe o med by he able-bodied pa icipan s
and a pos -s oke pa ien .................................................................................................................. 83
Table 5.3 – LOSOCV me ics o all inpu combina ions and DL models. The bes and he wo s esul s
a e colo ed in blue and ed, espec i ely. The highes pe o mance achie ed is shaded in g een colo 85
Table 5.4 – O line es me ics ......................................................................................................... 88
Table 5.5 – Online es me ics o able-bodied pa icipan s ............................................................... 89
Table 5.6 – Online es me ics o he s oke pa ien ........................................................................ 89
Table 5.7 – Success a e (%) pe LM ................................................................................................. 90
Table 5.8 – A e age
p edic ion ime
in milliseconds (o ange) and as a pe cen age o he gai cycle (blue)
........................................................................................................................................................ 91
Table 6.1 – Online es me ics o able-bodied pa icipan s ............................................................. 104
Table 6.2 – A e age
upda e ime in ad ance
in milliseconds (o ange) and as a pe cen age o he gai
cycle (blue) ..................................................................................................................................... 105
Table 6.3 – DL models pe o mance when EMG signals we e/we e no included as inpu s .............. 116
Table 6.4 – RMSE, NMSE, and
me ics o LOSOCV p ocedu e ..................................................... 118
Table 6.5 – RMSE, NMSE, and
du ing he LOSOCV and model es p ocedu es ............................. 119
Table 6.6 – CNN’s compu a ional load ............................................................................................ 121

x
Table 6.7 – Va ia ion o he EMG signals om
ibialis an e io
and
gas ocnemius la e alis
, ange o mo ion
(ROM) o he hip join , and human ankle join o que be ween
uncondi ioned
and
condi ioned
asks du ing
AAN EMG-based con ol s a egy (a nega i e and a posi i e sign means a educ ion and an inc ease,
espec i ely, when compa ing he a iable measu ed a he
condi ioned
ask o he
uncondi ioned
ask)
...................................................................................................................................................... 124
Table 6.8 – Va ia ion o he do si lexion and plan a lexion mo o o ques be ween
uncondi ioned
and
condi ioned
asks (a nega i e and a posi i e sign means a educ ion and an inc ease, espec i ely, when
compa ing he a iable measu ed a he
condi ioned
ask o he
uncondi ioned
ask). Maximum
do si lexion, and plan a lexion angles measu ed a he
condi ioned
ask ........................................ 125
Table 6.9 – Va ia ion o he EMG signals om
ibialis an e io
and
gas ocnemius la e alis
, and do si lexion
and plan a lexion mo o o ques be ween
uncondi ioned
and
condi ioned
asks (a nega i e and a posi i e
sign means a educ ion and an inc ease, espec i ely, when compa ing he a iable measu ed a he
condi ioned
ask o he
uncondi ioned
ask). Maximum do si lexion, and plan a lexion angles measu ed
a he
condi ioned
ask ................................................................................................................... 126
Table 7.1 – Cha ac e is ics o he ec ui ed pos -s oke pa ien s ...................................................... 145
Table 7.2 – P ocedu e adop ed du ing p e-, pos - aining, and ollow-up sessions............................. 147
Table 7.3 – P ocedu e adop ed du ing he amilia iza ion session wi h he Sma Os sys em ............. 147
Table 7.4 – P ocedu e adop ed du ing in e en ion sessions wi h he Sma Os sys em .................... 151
x i
Lis o abb e ia ions and ac onyms
A
AAN
Assis -As-Needed
ACC
Accu acy
ADC
Analog- o-Digi al Con e e
AI
A i icial In elligence
AO
Ac i e O hosis
API
Applica ion P og amming In e ace
B
BBS
Be g Balance Scale
BI
Ba el Index
C
CAN
Con ol A ea Ne wo k
CCU
Cen al Con ol Uni
CNN
Con olu ional Neu al Ne wo k
CPG
Con ol Pa e n Gene a o
D
DL
Deep Lea ning
E
EGPR
Exponen ial Gaussian P ocess Reg ession
EMG
Elec omyog aphy
F
FAC
Func ional Ambula ion Classi ica ion
FEL
Feedback-E o Lea ning
FIM
Func ional Independence Measu e
5MWT
5-Me e Walk Tes
FMA-LE
Fugl-Meye Assessmen – Lowe Ex emi y
FSR
Fo ce Sensing Resis o
H
HAI
Hause Ambula ion Index
HIL
Human-in- he-Loop
x ii
I
IMU
Ine ial Measu emen Uni
K
KPI
Key Pe o mance Indica o
L
LGW
Le el-G ound Walking
LM
Locomo ion Mode
LOSOCV
Lea e-One-Subjec -Ou C oss-Valida ion
LSTM
Long Sho -Te m Memo y
M
MAS
Modi ied Ashwo h Scale
MCC
Ma hew’s Co ela ion Coe icien
MI
Mo ici y Index
MMG
Mechanomyog aphy
MMSE
Mini-Men al S a e Examina ion
MRCS
Medical Resea ch Council Scale
MSD
Musculoskele al Diso de
MSE
Mean Squa e E o
MVC
Maximum Volun a y Con ac ion
N
NIHSS
Na ional Ins i u e o Heal h S oke Scale
O
OB
Objec i e
P
PAFO
Powe ed Ankle-Foo O hosis
PID
P opo ional In eg al De i a i e
PKO
Powe ed Knee O hosis
R
Spea man Co ela ion coe icien
R2
Coe icien o De e mina ion
RA
Ramp Ascen
x iii
RD
Ramp Descen
RMI
Ri e mead Mobili y Index
RMS
Roo Mean Squa e
RMSE
Roo Mean Squa e E o
ROM
Range o Mo ion
RQ
Resea ch Ques ion
S
SA
S ai Ascen
SD
S ai Descen
SIAS
S oke Impai men Assessmen Se
Si
Si ing
6MWT
6-Minu e Walk Tes
Sma Os
SMAR con ol o a sTand-alone ac i e O ho ic Sys em
S
S anding
STD
S anda d De ia ion
T
10MWT
10-Me e Walk Tes
TCP/IP
T ansmission Con ol P o ocol/In e ne P o ocol
TCT
T unk Con ol Tes
TRL
Technology Readiness Le el
TUG
Timed Up and Go Tes
2MWT
2-Minu e Walk Tes
U
UART
Uni e sal Asynch onous Recei e /T ansmi e
1
1. INTRODUCTION

Chap e 1
2
This Ph.D. hesis p esen s he esea ch de eloped in he las ou yea s in he scope o he
doc o al p og am in Biomedical Enginee ing. The in es iga ion was de eloped in he Biomedical Robo ic
De ices Labo a o y (BiRDLAB) a he Cen e o Mic oElec oMechanical Sys ems (CMEMS) es ablished
a he Uni e si y o Minho, oge he wi h he Cen o Clínico Académico de B aga (2CA – B aga), a
pa ne ship be ween he Uni e si y o Minho and B aga Hospi al, which p o ides he expe ise,
equipmen , acili ies, and access o pa ien s and clinicians.
The de eloped ac i i ies ha e been included in he p ojec Sma Os – SMAR con ol o a sTand-
alone ac i e O ho ic Sys em (POCI-01-0247-FEDER-039868). In addi ion, his esea ch ex ends beyond
a esea ch g an (UMINHO/BI/83/2020), a Mas e ’s hesis [1], and a Ph.D. hesis [2] by p oposing
Assis -As-Needed (AAN) con ol s a egies o a wea able o ho ic sys em in an a emp o es o e
he unc ional lowe limb mo o abili ies o s oke su i o s. This Ph.D. hesis add esses he nexus o
assis i e con ol s a egies, wea able assis i e de ices, a i icial in elligence (AI), and human gai analysis
o pe sonalize he assis ance o he wea able o ho ic sys em acco ding o he use ’s needs
and locomo ion in en ions o p omo e pos -s oke eco e y.
1.1 MOTIVATION
Walking is one o he mos daily pe o med human mo o asks. Howe e , human gai can be
comp omised due o neu ological diseases, such as s oke. Acco ding o he Wo ld Heal h O ganiza ion,
s oke e en s co espond o he second leading cause o dea h and he hi d leading cause o
disabili y globally [3]. S udies ha e epo ed a ound 13.7 million new inciden cases o s oke in he
wo ld in 2016 (be ween 121 and 151 pe 100.000 people only in Po ugal) [4]. The annual economic
cos o s oke in Eu ope is a ound 60 billion eu os, ep esen ing 0.36% o he Eu opean G oss Domes ic
P oduc (nea ly 0.40% o Po ugal) [5], [6].
Based on he s udy [7], mo e han 80% o s oke su i o s p esen gai dys unc ion due
o muscle weakness, asymme ical gai pa e n, pain, spas ici y, o a loss o mo o con ol o he lowe
limbs. Consequen ly, he pa ien ’s quali y o li e is a ec ed since hey canno pe o m hei daily
locomo ion ac i i ies (e.g., walking, unning, s anding, si ing, u ning, and climbing s ai s/ amps), o hei
accomplishmen is di icul . These limi a ions commonly esul in social and wo k exclusion, cos ly
medical assis ance, and ea ly e i emen [8].
The e is an ex eme necessi y o imp o ing he quali y o li e o pos -s oke pa ien s.
Con en ional physical ehabili a ion p o ided by physio he apis s has been widely applied o ace long-
Chap e 1
3
e m mo o disabili ies o hese neu ologically inju ed pa ien s [9]. None heless, lowe limb ehabili a ion
in he las yea s has acknowledged he need o deal wi h (i) he disad an ages associa ed wi h he in e -
and in a- he apis a iances; (ii) he dependency on he malleabili y o he pa ien ’s join (commonly
a ec ed by spas ici y); and, (iii) he absence o p ecise, use -o ien ed, and pe sonalized
assis ance du ing he apy [10]. In addi ion, ecen s udies ha e iden i ied he cos s o ehabili a ing
pa ien s who ha e su i ed a s oke [11], [12]. These cos s a e di ided in o inpa ien and ou pa ien
ehabili a ion cos s. Inpa ien s a e pa ien s who s ay in he hospi al and ecei e medical ea men , as
well as ood and accommoda ion in a hospi al. Ou pa ien s a e pa ien s who do no equi e
hospi aliza ion. An ou pa ien isi s a hospi al, clinic, o simila acili y o a diagnosis, ea men , o
p ocedu e and may lea e. Acco ding o he s udy [11], he cos o ehabili a ion he apies o ou pa ien s
is ypically h ee imes lowe han o inpa ien s. While he la ges con ibu o o inpa ien ehabili a ion
cos s is he cos o s ay (34%), he second la ges con ibu o is he cos o physical
ehabili a ion (16%). In he case o ou pa ien s, physical ehabili a ion co esponds o he
la ges con ibu o o he ehabili a ion cos s (26%). Conside ing he men ioned opics, in he
ehabili a ion a ea, obo ic assis ance d i en by wea able assis i e de ices, such as exoskele ons
and ac i e o hoses (AOs), has s eadily gained impo ance [10].
Acco ding o he s udy [2], ehabili a ion he apies d i en by exoskele ons may imp o e he
pa ien ’s muscula s eng h, mo emen coo dina ion, and balance con ol, os e ing he pa ien ’s
ambula ion and pe o mance o success ul locomo ion. Acco dingly, he use o wea able assis i e
de ices as a complemen a y ehabili a ion ool o con en ional physical ehabili a ion
s a egies may empowe he long- e m unc ional mo o eco e y o pa ien s wi h lowe limb
impai men s, os e ing hei con idence and independence by p o iding daily assis ance [13]–
[18]. Fu he mo e, when i comes o compa ing he cos o ehabili a ion be ween con en ional he apy
and obo ics-guided he apy, a ecen s udy concluded ha in 80% o he cases analyzed, obo ics-
guided he apy was mo e cos -e ec i e han con en ional he apy [12]. Al hough some
wea able assis i e de ices can cos se e al hund ed eu os, hei use in ehabili a ion he apies can
inc ease he e iciency o he he apis 's wo k, meaning ha mo e pa ien s can be ea ed, leading o an
o e all educ ion in he cos o ea men pe pa ien .
Chap e 1
4
1.2 PROBLEM STATEMENT
Despi e becoming a p ominen in e en ion o ackle he necessi y o physical ehabili a ion
he apies, mos o he a ailable exoskele ons p esen poo usabili y and do no deli e
pe sonalized assis ance acco ding o he pa ien ’s needs and in en ions o mo e [19]–[26].
Addi ional impe a i e de elopmen s include explo ing a ian s o AAN con ol s a egies o
ace he limi a ions o he ypically adop ed ajec o y- acking con ol s a egies [27]–[31]. Despi e he
epe i i e na u e o gai aining imposed by ajec o y- acking con ol s a egies, hey end o igno e he
human- obo in e ac ion and he needs o each use . As a esul , hese s a egies may limi mo o
elea ning and end up being abandoned [27], [28]. The de elopmen o adap i e and complian con ol
s a egies is undamen al o imely pe sonalize assis i e ajec o ies acco ding o he use ’s
mo o needs, and consequen ly, assis he use s as much and when needed. A his le el, AAN con ol
s a egies a e needed o dynamically adjus he le el o assis ance based on he use 's eal- ime muscula
pe o mance and pa icipa ion abili y, ensu ing ha he wea able assis i e de ice assis s mo emen
wi hou o e compensa ing. The goal is o encou age he use 's ac i e pa icipa ion and muscle
engagemen , he eby p omo ing imp o emen s in s eng h and mo o unc ion o e ime [29]–[31]. In
o de o cap u e he use ’s muscula pe o mance and pa icipa ion abili y, elec omyog aphy (EMG)
signals ha e been employed in AAN EMG-based con ol s a egies. Al hough p omising, mos o he AAN
EMG-based con ol s a egies we e de eloped o uppe limbs [32]–[34]. And hose de eloped o he
lowe limbs we e no designed o assis he en i e gai cycle, bu a he ocus on isola ed lexion and
ex ension mo emen s [28], [35]–[38].
In addi ion, cu en challenges in pe sonalized obo ics-based assis ance a e ela ed o decoding
di e en locomo ion modes (LMs) wi h a non-in usi e senso se up o imely igge he assis ance
deli e ed by wea able assis i e de ices acco ding o he use ’s locomo ion in en ions (i.e., AAN LM-d i en
con ols). Despi e ecen ad ancemen s, mos o he cu en AAN LM-d i en con ols in eg a ed in o
wea able assis i e de ices (i) add ess a limi ed numbe o daily LMs (a e non-gene ic ools) [19], [39]; (ii)
p esen high ecogni ion delays, classi ying and assis ing he LM only a e he use has al eady
ansi ioned o he new LM [20], [21], [40]–[44]; (iii) do no accoun o he ypically slowe gai speeds
o indi iduals wi h lowe limb disabili ies, since s udies add essed sel -selec ed and/o ixed speeds o
heal hy pa icipan s ( ypically abo e 2.7 km/h) [20], [21], [45], [46]; and (i ) do no p esen clinical
e idence [19]–[22]. I is o he u mos impo ance ha wea able assis i e de ices ackle hese limi a ions
by including algo i hms and AAN LM-d i en con ols capable o accu a ely and imely decoding
Chap e 1
5
di e en LMs o p o ide pe sonalized assis ance. I he use ’s LMs canno be p ope ly well-
de e mined, he exoskele on assis ance may be a ec ed [42], [47].
Fu he mo e, ecen a ia ions o AAN app oaches include human-in- he-loop (HITL) con ol
s a egies, using he use 's ene gy expendi u e as a means o adap ing he assis ance p o ided by he
wea able assis i e de ice [31], [48]. By ailo ing he assis ance o he use 's cu en physical s a e and
ac i i y le el, hese con ol s a egies aim o imp o e o e all com o , pe o mance, and
endu ance du ing he wea able assis i e de ice’s use [31], [49]–[51]. Howe e , despi e he p omising
pe o mance, he a ailable me hod o es ima ing ene gy expendi u e is indi ec calo ime y, which uses
non-po able equipmen , is ime-consuming, p oduces noisy es ima es, and is imp ac ical o eal-wo ld
applica ions [52].
In addi ion, mos wea able assis i e de ices p esen a poo clinical e idence base, cha ac e ized
by a lack o s udies e alua ing he longi udinal and ollow-up e ec s o obo ic he apies. The e is a high
numbe o s udies ha do no ollow a andomized con olled p ocedu e and do no comp ehensi ely
assess he e ec s o wea able assis i e de ices a kinema ic, physiological, and unc ional le els [53]–
[60].
This Ph.D. hesis in ends o ackle he challenges men ioned abo e using a sma o ho ic
sys em – Sma Os sys em, in an a emp o p o ide new insigh s and inno a i e esea ch di ec ions
o pos -s oke ehabili a ion using wea able assis i e de ices.
1.3 GOALS
This inno a i e mul idisciplina y Ph.D. hesis ocuses on de eloping AAN con ol s a egies
and hei in eg a ion in o he Sma Os sys em, o p omo e pe sonalized assis ance au oma ically
adap ed acco ding o he pa ien ’s physiological needs and locomo ion in en ions.
Fo his pu pose and conside ing he main goal o his hesis, he e a e six objec i es o be
pu sued:
• Objec i e 1: To e iew he e ec s o lowe limb assis i e de ices in pos -
s oke gai ehabili a ion. This e iew aims o iden i y he wea able assis i e de ices
al eady applied in he ehabili a ion o pos -s oke pa ien s, as well as he join s ypically
assis ed, he senso s used, and he con ol s a egies employed. I also ocuses on
ex ac ing in o ma ion on he cha ac e iza ion o pos -s oke pa ien s, s udy design,
p o ocols used, and clinical ou comes. The e iew concludes by highligh ing he bene i s
o using wea able assis i e de ices in ehabili a ion sessions o pos -s oke pa ien s. The
Chap e 1
12
Con e ence Pape s
• Luís Mo ei a, Robe o Ma ins Ba bosa, Joana Figuei edo, Ped o Fonseca, João Paulo
Vilas-Boas, C is ina P. San os, “Real-Time To que Es ima ion Using Human and Senso
Da a Fusion o Exoskele on Assis ance”,
The Six h Ibe ian Robo ics Con e ence
(ROBOT2023), Coimb a, 2023.
• Luís Mo ei a, Joana Figuei edo, João Ce quei a, C is ina P. San os, “A Real- ime
Kinema ic-based Locomo ion Mode P edic ion Algo i hm o an Ankle O hosis”,
24 h IEEE
In e na ional Con e ence on Au onomous Robo Sys ems and Compe i ions
(ICARSC),
Pa edes de Cou a, 2024.
• Luís Mo ei a, Joana Figuei edo, João Ce quei a, C is ina P. San os, “Assis -As-Needed
Elec omyog aphy-based Con ol o a Wea able Ankle Robo ic O hosis”,
IEEE RAS
EMBS 10 h In e na ional Con e ence on Biomedical Robo ics and Biomecha onics
(BioRob 2024), Heidelbe g, 2024.
1.7 MANUSCRIPT OUTLINE
This Ph.D. hesis is o ganized in o eigh chap e s, as illus a ed in Figu e 1.2. Chap e 2 p o ides
a comp ehensi e e iew o he clinical e idence o using lowe limb assis i e de ices in pos -s oke gai
ehabili a ion. I co e s he join s ypically assis ed, he con ol s a egies employed, he senso s used,
pos -s oke pa ien cha ac e iza ion, s udy design, p o ocols, and bo h senso -based and clinical
ou comes. Chap e 3 de ails he design o he Sma Os sys em, highligh ing he p oposed ad ances in
he a chi ec u e de eloped in he p e ious s udy [2]. This includes he in eg a ion o a new wea able
senso sys em, gai analysis ools, and AAN con ol s a egies in o he Sma Os amewo k. Chap e 4
p esen s he de elopmen and alida ion o a wea able senso sys em ( he MuscLab sys em) o moni o
muscle con ac ion, u ilizing e- ex ile senso s embedded in a lexible and elas ic wea able band. Chap e
5 ocuses on he de elopmen , in eg a ion, and alida ion o an LM decoding ool o decode ou LMs (S ,
LGW, SA, and SD), in eal- ime. I also includes a benchma k analysis o iden i y he bes usion o senso
ea u es and DL algo i hms o e ec i e decoding o daily LMs. Chap e 6 p esen s he h ee p oposed
AAN con ol s a egies, de ailing he design, me hodology, and alida ion o each con olle de eloped o
he Sma Os sys em. Chap e 7 in oduces he clinical p o ocol de eloped and s a es he physiological,
kinema ic, and unc ional e ec s o using he Sma Os sys em in ehabili a ion he apies o pos -s oke

Chap e 1
13
pa ien s. Chap e 8 summa izes he main indings and conclusions o he Ph.D. hesis, oge he wi h
di ec ions o u u e esea ch and oppo uni ies o echnical imp o emen .
Figu e 1.2 – Ph.D. hesis o ganiza ion. OB. and RQ. ep esen objec i e and esea ch ques ion, espec i ely.
Chap e : In oduc ion
Chap e : Li e a u e Resea ch
Chap e : Sma Os Sys em O e iew
Chap e : MuscLab Sys em
Chap e : Locomo ion Mode Decoding Tool
Chap e : Assis As Needed Con ol S a egies
Chap e : Sma Os Clinical alida ion
Chap e : Conclusions
14
2. LITERATURE RESEARCH
Chap e 2
15
2.1 INTRODUCTORY INSIGHT
S oke is one o he leading causes o mo o disabili y in adul s, and hose a ec ed o en ace
p o ound and las ing neu ological consequences [3]. Acco ding o he s udy [68], 50% o pos -s oke
pa ien s a e ini ially unable o walk, 12% can walk wi h assis ance, and 38% can walk independen ly. Pos -
s oke pa ien s who a e able o walk, expe ience a gai pa e n ha is commonly di e en om ha o
heal hy indi iduals and ha is associa ed wi h a highe isk o alling. This phenomenon is highly
associa ed wi h hemipa esis allied o join spas ici y [69], [70]. Spas ici y is a common sequela in
s oke su i o s, o en leading o a ious de o mi ies and changes in pos u e and mo emen pa e ns due
o he abno mal inc ease in muscle one and s i ness [69], [70].
Ankle spas ici y is among he mos p e alen mo emen diso de s ollowing a s oke ypically
esul ing in equinus, a us, o equino a us de o mi ies [69], [70]. Equinus oo de o mi y causes
he ankle and oo o be in a plan a lexed posi ion, making olun a y do si lexion o he ankle join
di icul . This de o mi y is p ima ily caused by spas ici y in he ankle plan a lexo s (
gas ocnemius
and
soleus
muscles). Va us oo de o mi y is a condi ion in which he oo is held in an in e ed
posi ion. This abno mal alignmen is p ima ily caused by spas ici y in bo h he
ibialis an e io
and
ibialis pos e io
muscles. On he o he hand, an equino a us de o mi y is cha ac e ized by
ankle plan a lexion and in e sion. I is p ima ily caused by spas ici y in he plan a lexo s
and
ibialis pos e io
muscle (in e o muscle), wi h minimal o no con ibu ion om he
ibialis
an e io
(do si lexo muscle) [69], [70].
In addi ion o hese di e en ankle de o mi ies, pos -s oke pa ien s also e eal changes in he
physiological (EMG), kinema ic, and spa io empo al pa ame e s o he lowe limb join s du ing
locomo ion [71]–[75]. A he physiological le el, he e is high he e ogenei y in he EMG signals o
pos -s oke pa ien s. This he e ogenei y is ound no only be ween indi iduals (in e -indi idual a ia ions)
bu also in he same indi idual (in a-indi idual a ia ions). In e -indi idual a ia ion may be explained by
he na u e o he s oke, which can a ec di e en egions o he b ain and esul in di e en mo o
disabili ies [7]. On he o he hand, in a-indi idual a ia ions may occu due o empe a u e a ia ions
[72]. As s a ed in he s udy [72], olde and emale pos -s oke pa ien s a e mo e ulne able o wea he
condi ions. Al hough he e is high a iabili y, he e a e also common ac o s a he EMG le el, namely a
educ ion in he magni ude o he EMG signals ob ained om he muscles o he pa e ic limb,
ea ly onse o a igue, and p olonged du a ion o i ing du ing he gai cycle [71], [76].
Chap e 2
16
Lowe limb join kinema ics a e a ec ed as a esul o i egula EMG pa e ns. Pos -s oke pa ien s
a e commonly cha ac e ized by educed hip ex ension du ing he s ance phase [73]. This
phenomenon is a consequence o he o e ac i a ion o he ankle plan a lexo s. The excessi e
ac i i y o he ankle plan a lexo s does no allow he ankle o do si lex as much as equi ed. Besides,
a e a s oke, he leng h o plan a lexo s is commonly dec eased, which educes he abili y o hese
muscles o p oduce enough o ce o pe o m plan a lexion mo emen s a he e minal s ance phase
[73]. As a esul , he o wa d mo emen o he uppe body is es ic ed and he hip join does no ex end
as expec ed [71]. Mo eo e , du ing he swing phase, hip lexion ends o be educed as a
consequence o he inabili y o ac i a e hip lexo s and/o o e ac i i y o hip ex enso s [71]. Rega ding
he knee join , i ypically p esen s a educed lexion abili y du ing he ea ly-s ance phase. This
phenomenon is ollowed by knee hype ex ension used as a compensa o y mechanism o achie e s abili y
[74]. A he mid-swing phase, he knee join ypically exhibi s a dec eased knee lexion due o
o e ac i i y o he
ec us emo is
and/o weakness o he
biceps emo is
[71]. A he same
phase o he gai cycle, he ankle join p esen s a educed do si lexion due o o e ac i i y o he
plan a lexo s. Thus, a neu al posi ion o he ankle join (a ound 0º in he sagi al plane) is no
commonly e i ied [75]. In summa y, es ic ed hip and knee lexion, along wi h dec eased ankle
do si lexion, elonga es he leg leng h du ing he swing phase. This elonga ion dec eases he oo 's
clea ance om he loo , leading o oe d agging o compensa o y leg ci cumduc ion [71].
As a consequence o EMG and kinema ics being a ec ed, he spa io empo al pe o mance o pos -
s oke pa ien s is also comp omised. A his le el, he locomo ion o pos -s oke pa ien s is ypically
cha ac e ized by a educed gai speed (below 2.7 km/h [77]) and cadence, inc eased s ide ime
o he non-pa e ic limb, and double suppo ime. Mo eo e , he pa e ic limb commonly p esen s a
s ance phase wi h a lowe du a ion and a swing phase wi h a highe du a ion, compa ed o
he non-pa e ic limb [71].
To deal wi h he abo e-men ioned mo o disabili ies, physical ehabili a ion in e en ions ha
p omo e b ain plas ici y may include use -o ien ed, ask-o ien ed, and epe i i e gai aining ha
encou ages ac i e pa icipa ion in he apy. Wea able assis i e de ices, such as exoskele ons and
AOs, in conjunc ion wi h con en ional ehabili a ion he apies, may help achie e hese goals [9].
These de ices ha e been employed o es o e o modi y human mo o unc ion, enhancing walking abili y
in indi iduals wi h impai ed gai due o neu ological and/o mo o diseases o inju ies, such as pos -
s oke pa ien s [78]. Recen ly, esea ch in o lowe limb exoskele ons and AOs has su ged exponen ially,
es ablishing hem as p ominen physical ehabili a ion in e en ions. These de ices educe he
Chap e 2
17
physical bu den o he apis s and may ep esen a powe ul ool o imp o ing mobili y and
eco e y o pos -s oke pa ien s [79].
2.2 REVIEW OF THE EFFECTS OF LOWER LIMB ASSISTIVE DEVICES ON
POST-STROKE PATIENTS
As wea able assis i e de ices ha e been inco po a ed in o ehabili a ion he apies o pos -s oke
pa ien s, i is no able o conside he bene i s o hei use. Conside ing his, his e iew aims o iden i y
he wea able assis i e de ices ha ha e been es ed in pos -s oke pa ien s, as well as he join s ypically
assis ed, he con ol s a egies adop ed, and he senso s used. This e iew also ocuses on ex ac ing
in o ma ion ega ding he cha ac e iza ion o pos -s oke pa ien s, s udy design, p o ocols used, and
senso -based and clinical ou comes. The e iew concludes by iden i ying he bene i s o using
wea able assis i e de ices in ehabili a ion sessions o pos -s oke pa ien s.
2.2.1 SEARCH METHODOLOGY
The li e a u e sea ch was conduc ed om Decembe 2020 o June 2024 in he Scopus, PubMed,
and Coch ane da abases using he ollowing keywo ds: (exoskele on OR exoskele ons OR o hos?s) AND
("pos -s oke") AND ("lowe -limb" OR "lowe -limbs" OR "lowe limb" OR "lowe limbs"). This sea ch was
limi ed o i les, keywo ds, and abs ac s.
Manusc ip s we e e alua ed based on he ollowing inclusion c i e ia: (i) o iginal s udies; (ii) clinical
in e en ion including pos -s oke pa ien s; (iii) lowe limb ac i e exoskele ons/o hoses applied o gai
ehabili a ion. The ollowing exclusion c i e ia we e applied: (i) unused exoskele ons/o hoses (3 pape s
ejec ed); (ii) only one aining session applied (12 pape s ejec ed); (iii) only p o ocol p esen ed (6 pape s
ejec ed); and (i ) no pos -s oke pa ien s (3 pape s ejec ed).
Technical and clinical in o ma ion da a we e ex ac ed om he selec ed s udies. Technical
in o ma ion includes (i) he exoskele on used, ac ua o s employed, and join assis ed; (ii) senso sys ems
used o adap he assis ance; and (iii) assis i e con ol s a egies adop ed. On he o he hand, clinical
in o ma ion includes (i) cha ac e is ics o pos -s oke pa ien s; (ii) he s udy design; (iii) he clinical p o ocol
ollowed; (i ) senso -based and clinical ou comes; and ( ) clinical e idence o wea able assis i e de ices
on pos -s oke eco e y.

Chap e 2
18
2.2.2 RESULTS
The li e a u e sea ch iden i ied 123 s udies, o which 78, 21, and 19 we e ound in he Scopus,
PubMed, and Coch ane da abases and 5 pape s we e iden i ied manually om he e e ence sec ions o
o he s udies. A e emo ing duplica es, 100 s udies emained o sc eening. Based on hei i les and
abs ac s, 58 pape s we e excluded. Consequen ly, 42 ull- ex a icles we e assessed o eligibili y.
Acco ding o he inclusion and exclusion c i e ia, 18 s udies we e inally included. Figu e 2.1 depic s he
PRISMA lowcha de ailing his selec ion p ocess.
Figu e 2.1 – PRISMA lowcha .
S udies included in e iew
(n = 18)
Iden i ica ion o s udies ia da abases and egis e s
Sc eening
Reco ds sc eened
(n = 100)
Reco ds excluded
(n = 58)
Repo s sough o e ie al
(n = 42)
Repo s no e ie ed
(n = 0)
Repo s assessed o eligibili y
(n = 42)
Repo s excluded:
The s udy does no use
exoskele ons/o hoses (n =
3)
The s udy applies a single
session (n = 13)
The s udy only p esen s he
p o ocol (n = 6)
The s udy does no include
pos -s oke pa ien s (n = 3)
Included
Iden i ica ion
Reco ds emo ed
be o e sc eening
:
Duplica e eco ds emo ed (n = 23)
Reco ds ma ked as ineligible by
au oma ion ools (n = 0)
Reco ds emo ed o o he
easons (n = 0)
Reco ds iden i ied om:
Scopus (n = 78)
PubMed (n = 21)
Coch ane (n = 19)
Addi ional Reco ds iden i ied
h ough o he sou ces (n = 5)
Chap e 2
19
A. Technical In o ma ion
Table 2.1 summa izes he echnical in o ma ion ex ac ed om he selec ed s udies, namely,
(i) he de ice, ac ua o ype, and assis ed join s; (ii) senso sys ems; and (iii) assis i e con ol
s a egies.
De ice, Ac ua o Type, and Assis ed Join s
Acco ding o Table 2.1, he e ec s o wea able assis i e de ices in pos -s oke pa ien s we e
s udied o di e en de ices. The mos explo ed wea able assis i e de ices we e he EksoNR (six
s udies (33.3%) [16], [54], [56], [57], [60], [80]) and he HAL ( i e s udies (27.6%) [17], [58], [59],
[81], [82]). Two s udies explo ed he ExoA le [14] [15]. The e ec s o F eeWalk [13], H2-Exo [53],
Healbo T [83], SMA [18], and a sel -made ac i e ankle- oo o hosis [55] we e explo ed in only one
s udy each.
O he eigh een selec ed s udies, se en een used an elec ic mo o -based ac ua o wi h gea -
based ansmission [13]–[18], [53], [54], [56]–[60], [80]–[83] (94.4%). In o ma ion abou he
ac ua o o he sel -made ac i e ankle- oo o hosis o he s udy [55] was no de ailed.
Addi ionally, he lowe limb join ypically add essed by wea able assis i e de ices was he
hip join (se en een s udies (94.4%) [13]–[18], [53], [54], [56]–[60], [80]–[83]), ollowed by he knee
join (six een s udies (88.9%) [13]–[17], [53], [54], [56]–[60], [80]–[83]). These wo join s we e
assis ed oge he in six een s udies (88.9%) [13]–[17], [53], [54], [56]–[60], [80]–[83]. The ankle
join was (i) ac i ely assis ed in wo s udies (11.1%) [53], [55]; (ii) passi ely assis ed in ou een
s udies (77.8%) [13]–[17], [54], [56]–[60], [80]–[82]; and (iii) no assis ed in wo s udies (11.1%)
[18], [83].
Senso Sys ems
Di e en senso sys ems we e employed in he con ol s a egy a chi ec u e o each s udy.
Se en s udies (41.2%) used Fo ce Sensing Resis o (FSR) senso s o de ec la e al and o wa d weigh
shi s [16], [54], [56]–[58], [60], [80] and o segmen gai cycles [81]. Fi e s udies (29.4%) employed
EMG sys ems, acqui ing EMG signals om
illiopsoas
,
glu eus maximus
,
semi endinosus
, and
as us
la e alis
muscles [17], [58], [59], [81], [82]. Th ee s udies (17.6%) used embedded encode s o
measu e he lowe limb join angles [18], [53], [83]. Single s udies ha e used an Ine ial
Measu emen Uni (IMU) senso a he unk o segmen gai cycles [57] and an in e ac ion o que
senso [53] o de e mine he human- obo in e ac ion in each lowe limb join .
Chap e 2
20
Assis i e Con ol S a egies
Se e al assis i e con ol s a egies we e adop ed by di e en wea able assis i e de ices. O
he eigh een selec ed s udies, eigh s udies (44.4%) used a he apis -led s ep-by-s ep con ol [13]–
[15], [54], [56], [57], [60], [80]. This con ol s a egy is d i en by a physical he apis who commands
he s eps ha he use should pe o m by using a push bu on in he wea able assis i e de ice.
Six s udies (33.3%) explo ed he po en iali ies o a ajec o y- acking con ol based on la e al
and o wa d weigh shi [16], [54], [56], [57], [60], [80], all applied wi h he EksoNR. This con ol
s a egy ac s as a hyb id app oach in which each s ep is igge ed by he la e al and o wa d weigh
shi o he use . Once igge ed, he use ollows he ajec o y imposed by he wea able assis i e
de ice. In i e s udies (27.8%), his con ol s a egy was adop ed a e he he apis -led s ep-by-s ep
con ol, when he use lea ns o weigh -shi o a s ance posi ion [54], [56], [57], [60], [80].
Fi e s udies (27.8%) employed an EMG-based con ol [17], [58], [59], [81], [82], all applied
wi h he HAL. This con ol s a egy ies o ind a co ela ion be ween EMG and o que signals. Once
ound, he con olle se s assis i e o que commands p opo ional o he EMG signals o he pa ien .
Two s udies (11.1%) explo ed a ajec o y- acking con ol [17], [83]. Wi h his con ol
s a egy, he wea able assis i e de ice p o ides ixed assis ance le els ha a e equi ed o comple e
he desi ed posi ion ajec o ies. The use ollows he ajec o y imposed by he wea able assis i e
de ice.
One s udy (5.6%) explo ed he impedance con ol o H2 exoskele on [53]. When using his
con ol ype, he aim is o adjus he deg ee o eedom in he pa ien 's mo emen s by a ying he
in e ac ion s i ness. Low le els o in e ac ion s i ness allow g ea e eedom o mo emen , while high
le els esul in a mo e igid beha io by he wea able assis i e de ice o s ic ly en o ce he e e ence
ajec o y. Adjus ing he in e ac ion s i ness con ols he o que exe ed by he wea able assis i e
de ice, he eby in luencing he pa ien 's in e ac ion and he e o equi ed o main ain he desi ed
gai pa e n.
One s udy (5.6%) used a Con ol Pa e n Gene a o (CPG)-based con ol. This con ol sys em
uses neu al oscilla o s in addi ion o he use 's CPG o synch onize wi h he use 's mo emen s. The
join angles a e measu ed in eal- ime, ac ing as inpu in he con olle , which assesses he join
angles' symme y. Based on his analysis, he con olle gene a es assis i e o ques a speci ic poin s
o he gai cycle o imp o e i s symme y.
Chap e 2
21
Table 2.1 – Technical de ails ex ac ed om he selec ed s udies (A, P, and NI means ac i e, passi e, and no indica ed, espec i ely)
S udy
De ice
Ac ua o
Ac ua ed join
Senso sys ems used in con ol s a egies
Con ol s a egy
Hip
Knee
Ankle
Senso
Measu emen
Lee
e al.
[13]
F ee Walk
Elec ic wi h
gea -based
ansmission
A
A
P
Embedded encode
The apis -led s ep-by-s ep
exe cise
Ko o
e
al.
[14]
ExoA le
Elec ic wi h
gea -based
ansmission
A
A
P
The apis -led s ep-by-s ep
exe cise
Ko alenko
e al.
[15]
ExoA le
A
A
P
The apis -led s ep-by-s ep
exe cise
Louie
e
al.
[80]
EksoNR
Elec ic wi h
gea -based
ansmission
A
A
P
Embedded Foo P essu e
Senso (FSR)
La e al and o wa d weigh
shi
The apis -led s ep-by-s ep
exe cise and T ajec o y-
acking con ol based on
la e al weigh shi
Calab ò
e
al.
[16]
EksoNR
A
A
P
Embedded FSR
La e al and o wa d weigh
shi
T ajec o y- acking con ol
based on la e al weigh shi
Zhu
e al.
[60]
EksoNR
A
A
P
Embedded FSR
La e al and o wa d weigh
shi
The apis -led s ep-by-s ep
exe cise and T ajec o y-
acking con ol based on
la e al weigh shi
In a ina o
e al.
[57]
EksoNR
A
A
P
Embedded FSR
La e al and o wa d weigh
shi
The apis -led s ep-by-s ep
exe cise and T ajec o y-
Chap e 2
28
The pos -s oke s age e ec s we e e i ied by he Fugl-Meye Assessmen – Lowe Ex emi y
(FMA-LE) (87.5%) [17], [53], [55], [58], [80]–[82], and he S oke Impai men Assessmen Se (SIAS)
(12.5%) [55].
The men al e ec s we e analyzed in single s udies by he 36-i em [80] and 12-i em Sho
Su ey [13], Mon eal Cogni i e Assessmen [80], and Pa ien Heal h Ques ionnai e [80].
To e alua e he men ioned e ec s, all selec ed s udies ca ied ou a p e- and pos - aining
e alua ion. The s udies [18], [54], [59], [81], [82] pe o med addi ional e alua ions in in e media e
ime poin s be ween he beginning and end o he aining. Mo eo e , he s udy [18] and he s udies
[80], [82] pe o med a 3-mon h and 6-mon h ollow-up, espec i ely, o e alua e he e en ion o he
e ec s.
Clinical E idence o Wea able Assis i e De ices in Pos -S oke Pa ien s
Fo andomized con olled and non- andomized con olled s udies, he clinical e idence was
p esen ed as imp o emen s achie ed by he expe imen al g oup ( he g oup ha pe o med
con en ional and obo ic-based aining sessions) o e he con ol g oup ( he g oup ha ca ied ou
con en ional he apy), conside ing he p e- and pos - aining ime poin s. In he case o uncon olled
s udies, he clinical e idence was e alua ed by compa ing he ou comes o he expe imen al g oup
a he p e- and pos - aining ime poin s.
Gai speed imp o emen s we e he ou come mos ly e e enced (27.5%) [13]–[18], [53]–[56],
[58]–[60], [83], ollowed by imp o emen s in he balance (11.8%) [14]–[17], [54], [59], dis ance
walked (9.8%) [13], [17], [53], [54], [60], gai symme y (9.8%) [16], [18], [58], [60], [81], muscle
s eng h o he quad iceps (7.8%) [13], [14], [16], [54], and pos u al s abili y (5.9%) [14], [15], [54].
O he imp o emen s we e epo ed once (2.0%), such as s ide leng h [18], s ep leng h [18], walking
ime [56], ime o e icaliza ion [56], numbe o s eps [56], s ep cadence [58], oo clea ance [60],
knee lexion [16], ange o mo ion o he hip, knee, and ankle o he non-pa e ic limb [58], ange o
mo ion o he hip and knee o he pa e ic limb [58], muscle s eng h o
ibialis an e io , gas ocnemius,
and
biceps emo is
[57]. Fu he , he s udy [14] also epo ed a dec ease in he hemipa esis deg ee
and ene gy expendi u e.
Addi ionally, he esul s o he pos - aining o wo s udies [80], [82] did no epo any
imp o emen s in he expe imen al g oup o e he con ol g oup. Unlike he s udy [14], he s udy [60]
did no show imp o emen s in he ene gy expendi u e. The pos - aining esul s o he s udy [57] did
no e eal imp o emen s in gai speed, spas ici y le el, and pos u al s abili y. Despi e he pos - aining

Chap e 2
29
esul s e ealing imp o emen s in he gai speed, balance, gai symme y, knee lexion, and knee
muscle s eng h, he s udy [16] did no ind signi ican imp o emen s in he muscle ac i i y o he
ibialis an e io
and
gas ocnemius
muscles o he expe imen al g oup o e he con ol g oup.
Mo eo e , he s udy [58] also epo ed a dec ease in he ange o mo ion o he ankle join on he
pa e ic side.
Chap e 2
30
Table 2.2 – Clinical de ails ex ac ed om he selec ed s udies
S udy
Sample
Size
(EG/CG*)
S udy
design
Clinical
p o ocol
Ou comes
E alua ion
ime poin s
Clinical e idence
Senso -based
Clinical
Lee
e al.
[13]
38
(17/21)
RCT
3 imes/week
un il 12
sessions
Quad iceps isokine ic muscle
s eng h
Timed Up and Go Tes
(TUG); 6-min Walk Tes
(6MWT); 12-i em sho
o m su ey
P e- and
pos - aining
• Imp o emen o knee
muscle s eng h, dis ance
walked, gai speed, and
quali y o li e
Ko o
e
al.
[14]
42
(21/21)
RCT
30-min 5
imes/week
un il 10
sessions
Ve ical s abili y; Ene gy
consump ion
Medical Resea ch Council
(MRC); Modi ied Rankin
Scale; Ba el Index (BI);
Hause Ambula ion Index
(HAI); Be g Balance Scale
(BBS); 10-Me e Walk Tes
(10MWT)
P e- and
pos - aining
• Imp o emen o pa e ic
limb muscle s eng h,
balance, pos u al s abili y,
and gai speed.
• Dec ease in hemipa esis
deg ee and ene gy
consump ion
Louie
e
al.
[80]
36
(17/19)
RCT
25-min
wice/week
un il 24
sessions
Func ional Ambula ion
Ca ego ies (FAC); Fugl-
Meye Assessmen – Lowe
Ex emi y (FGM-LE); 5-
Me e Walk Tes (5MWT);
6MWT; BBS; Mon eal
Cogni i e Assessmen
P e-, pos -
aining, and
6-mon h
ollow-up
No imp o emen s we e
ound when compa ing he
exoskele on-based he apy
wi h con en ional he apy
Chap e 2
31
S udy
Sample
Size
(EG/CG*)
S udy
design
Clinical
p o ocol
Ou comes
E alua ion
ime poin s
Clinical e idence
Senso -based
Clinical
(MCA); 36-i em sho o m
su ey
Ko alenko
e al.
[15]
62
(31/31)
RCT
1-h/day un il
10 sessions
Ta dieu Scale (TS);
Modi ied Ashwo h Scale
(MAS); MRCS; 10MWT;
Ri e mead Mobili y Index
(RMI); BBS; Modi ied
Rankin Scale
P e- and
pos - aining
Imp o emen o gai speed,
balance, and pos u al
s abili y
Calab ò
e
al.
[16]
40
(20/20)
RCT
45-min 5
imes/week
un il 40
sessions
RMS o
ibialis an e io , soleus,
ec us emo is,
and
biceps
emo is
; s ep cadence;
s ance/swing a io; gai quali y
index; gai cycle du a ion
10MWT; RMI; TUG;
P e- and
pos - aining
• Imp o emen o gai speed,
balance, gai symme y,
knee lexion, and knee
muscle s eng h.
• The
ibialis an e io
and
soleus
did no show
imp o emen s
Zhu
e al.
[60]
21
(10/11)
Non-RCT
50-min 3
imes/week
un il 15
sessions
Ene gy expendi u e; lowe limb
join angles (hips, knees, and
ankles); RMS o
ibialis an e io ;
soleus, gas ocnemius medialis,
10MWT; 6MWT, TUG;
P e- and
pos - aining
• Imp o emen o gai speed,
dis ance walked, oo
clea ance, and gai
symme y.
Chap e 2
32
S udy
Sample
Size
(EG/CG*)
S udy
design
Clinical
p o ocol
Ou comes
E alua ion
ime poin s
Clinical e idence
Senso -based
Clinical
as us medialis, ec us emo is,
biceps emo is, semi endinosus,
and
glu eus medius
; walking
speed; oo clea ance;
s ance/swing a io
• The ene gy expendi u e did
no show imp o emen s
Tan
e al.
[58]
8 (8/0)
Uncon olled
20-min 3
imes/week
un il 9
sessions
Walking speed, cadence, s ep
leng h, s ance/swing a io, ange
o mo ion o he lowe limb join s
(hips, knees, ankles); EMG om
as us medialis, semi endinosus,
ibialis an e io , gas ocnemius,
adduc o longus, and glu eus
maximus
FAC; Func ional
Independence Measu e
(FIM); FMA-LE;
P e- and
pos - aining
Imp o emen o gai speed,
s ep cadence, gai
symme y, hip and knee
ange o mo ion o he
pa e ic side, and hip, knee,
and ankle ange o mo ion o
he una ec ed side
Tan
e al.
[81]
20 (9/11)
RCT
S ance a io be ween pa e ic and
non-pa e ic limb; EMG om
as us
medialis, semi endinosus, ibialis
an e io , gas ocnemius, adduc o
longus, and glu eus maximus
FAC; FIM; FMA-
Locomo ion; FMA-Mo o ;
FMA-LE
P e-, a e
session 3,
a e session
6, and pos -
aining
Imp o emen o gai
symme y
Chap e 2
33
S udy
Sample
Size
(EG/CG*)
S udy
design
Clinical
p o ocol
Ou comes
E alua ion
ime poin s
Clinical e idence
Senso -based
Clinical
In a ina o
e al.
[57]
8 (8/0)
Uncon olled
5 imes/week
un il 15
sessions
EMG om
ibialis an e io ,
gas ocnemius medialis, ec us
emo is,
and
biceps emo is;
knee
join angles
MAS; Mo ici y Index (MI);
FAC; T unk Con ol Tes
(TCT); 10MWT
P e- and
pos - aining
• Imp o emen o muscle
s eng h a
ibialis an e io ,
gas ocnemius medialis,
and
biceps emo is
• No imp o emen s we e
ound in gai speed,
spas ici y le el, and
pos u al s abili y
Szu le
e
al.
[56]
19 (19/0)
Uncon olled
50-min 3
imes/week
un il 12
sessions
Numbe o s eps; walking ime;
ime o e icaliza ion
TUG
P e- and
pos - aining
Imp o emen o gai speed,
numbe o s eps, walking
ime, and ime o
e icaliza ion
Lee
e al.
[83]
43
(33/10)
RCT
30-
min/session
un il 10
sessions
10MWT; BBS; FAC; TUG;
MI
P e- and
pos - aining
Imp o emen o pos u al
s abili y, balance, and gai
speed
Tomioka
e al.
[55]
27 (27/0)
Uncon olled
40-min 6
imes/week
FMA-LE; TUG; 10MWT;
S oke Impai men
Assessmen Se (SIAS)
P e- and
pos - aining
Imp o emen o gai speed

Chap e 2
34
S udy
Sample
Size
(EG/CG*)
S udy
design
Clinical
p o ocol
Ou comes
E alua ion
ime poin s
Clinical e idence
Senso -based
Clinical
un il 24
sessions
Mol eni
e
al.
[54]
23 (23/0)
Uncon olled
60-min 3
imes/week
un il 12
sessions
MAS; MI; TCT; FAC;
10MWT; 6MWT
P e-, a e
session 6,
and pos -
aining
Imp o emen o muscle
s eng h, pos u al s abili y,
balance, gai speed, and
dis ance walked
Wa anabe
e al.
[17]
22
(11/11)
RCT
20-min 3
imes/week
un il 12
sessions
Isome ic muscle s eng h a hip
and knee lexion and ex ension
mo emen s
FAC; 10MWT; 6MWT; FMA-
LE
P e- and
pos - aining
Imp o emen o gai speed,
balance, and dis ance
walked
Yoshimo o
e al.
[59]
18 (9/9)
Non-RCT
20-min
once/week
un il 8
sessions
FAC; 10MWT; TUG; BBS
P e-, a e
session 4,
and pos -
aining
Imp o emen o gai speed
and balance
Bo ole
e
al.
[53]
3 (3/0)
Uncon olled
3 imes/week
un il 12
sessions
BBS; Ba hel Index (BI);
FAC; FMA-LE; TUG; 6MWT
P e- and
pos - aining
Imp o emen o dis ance
walked and gai speed
Chap e 2
35
S udy
Sample
Size
(EG/CG*)
S udy
design
Clinical
p o ocol
Ou comes
E alua ion
ime poin s
Clinical e idence
Senso -based
Clinical
Buesing
e
al.
[18]
50
(25/25)
RCT
45-min 3
imes/week
un il 18
sessions
Walking speed; cadence; s ep
ime; s ep leng h; s ide leng h;
s ance ime; double suppo ime
P e-, a e
session 9,
pos - aining,
and 3-mon h
ollow-up
Imp o emen o gai speed,
s ep ime, s ep leng h, s ide
leng h, and gai symme y
Wall
e al.
[82]
32
(16/16)
RCT
60-min 4
imes/week
un il 16
sessions
FAC; FMA-LE; 2-minu e
Walk Tes ; BI
P e-, pos -
aining, and
6-mon h
ollow-up
No imp o emen s we e
ound when compa ing he
exoskele on-based he apy
wi h con en ional he apy
*EG and CG mean expe imen al and con ol g oups, espec i ely
Chap e 2
36
2.2.3 DISCUSSION
A. Technical In o ma ion
Among a ious clinical ials, EksoNR and HAL we e he mos commonly used de ices o
e alua e he e ec s o wea able assis i e de ice he apy in pos -s oke pa ien s. Al hough he use o
exoskele ons o he ehabili a ion o pos -s oke pa ien s has been a opic o discussion o some
ime, hese esul s show ha only wo wea able assis i e de ices ha e ea u ed p ominen ly in
s udies designed o p o e he e ec i eness o hese de ices in pos -s oke pa ien s. Fu u e esea ch
should add ess his gap by conduc ing clinical ials wi h o he de ices o in es iga e he
bene i s o each de ice o in es iga e i he e a e di e ences among hem.
Acco ding o he s udy [53], (i) hyd aulic and pneuma ic ac ua o s o e high powe densi y
bu a e ypically bulky and p one o in e nal leakage and ic ion; (ii) se ies elas ic ac ua o s a e
limi ed by he ixed sp ing cons an o hei elas ic elemen s; and (iii) elec ic mo o -based ac ua o s
ha e been indica ed due o i s educed powe consump ion du ing gai . These esul s suppo he
indings o he s udies e iewed since elec ic mo o -based ac ua o s we e ypically adop ed
o he majo i y o he s udies.
In mos o he s udies, he hip and knee join s we e ac i ely assis ed by he wea able
assis i e de ice, while he ankle join was passi ely assis ed. Acco ding o he s udy [71], he
do si lexion o he ankle join du ing he s ance phase is es ic ed due o he o e ac i i y o he
plan a lexo muscles, speci ically he
gas ocnemius
muscle. This limi a ion a ec s he hip join by
p e en ing i om ex ending as a as necessa y. Ano he common kinema ic issue is educed ankle
do si lexion du ing he swing phase. No mally, he ankle eaches a neu al posi ion a he mid-swing
phase and main ains his posi ion un il ini ial con ac . Achie ing a neu al ankle posi ion a his poin
is c ucial because i b ings he oo close o he g ound, acili a ing limb mo emen o p e en oe
d ag. The lack o do si lexion du ing he swing phase and a he heel-s ike phase is likely caused by
o e ac i e plan a lexo muscles and insu icien do si lexion. As a esul , pos -s oke pa ien s end
o pe o m leg ci cumduc ion [71]. These phenomena sugges ha he ankle join is o en a ec ed
in pos -s oke pa ien s and ha some imes a kinema ic change in ano he lowe limb join (such as
he hip) is a consequence o he ankle join being a ec ed. The e o e, u u e esea ch should
add ess hese issues by de eloping wea able assis i e de ices ha in oke he ankle join o
play an ac i e ole in ehabili a ion he apies.
Chap e 2
37
Addi ionally, weigh shi de ec ions measu ed by FSRs, EMG signals om,
glu eus
maximus
,
semi endinosus
, and
as us la e alis
muscles, and lowe limb join angles measu ed
by embedded encode s we e he mos used ype o da a in he con ol s a egies o he
e iewed s udies. Al hough EMG senso s we e widely used, hei applica ion was limi ed o he HAL
de ice. Gi en he impo ance o hese senso s in disc imina ing he locomo ion in en ions and needs
o pos -s oke pa ien s, u u e esea ch should add ess he inclusion o hese senso s in he con ol
a chi ec u e.
The he apis -led s ep-by-s ep con ol was he mos used con ol when using wea able
assis i e de ices, ollowed by he ajec o y- acking con ol based on la e al and o wa d
weigh shi . This las con ol s a egy was ypically adop ed a e he he apis -led s ep-by-s ep
con ol when he pos -s oke pa ien lea ned o weigh -shi o a s ance posi ion. While in he he apis -
led s ep-by-s ep con ol, pos -s oke pa ien s a e commanded by he physical he apis , he ajec o y-
acking con ol based on la e al weigh shi in okes he pa ien s' pa icipa ion o con inuously
pe o m igh and le s eps. None heless, he use ’s pa icipa ion is only equi ed a he beginning
o each s ep and no o he comple e gai cycle. In addi ion, he e ec o EMG-based con ol
s a egies was also explo ed. Fo he wea able assis i e de ice o mo e, he pos -s oke pa ien
mus ac i ely pa icipa e h oughou he gai cycle, as he de ice p o ides assis ance ha is
p opo ional o he le el o muscle ac i a ion o he moni o ed muscles. The e o e, non-pa icipa ion
by pos -s oke pa ien s means ha he wea able assis i e de ice does no assis [37]. Despi e being
a aluable con ibu ion o muscle s eng hening, hese EMG-based o que con ols do no conside
di e en le els o mo o disabili ies, no assis he use when and as much as needed. Fu u e
esea ch should add ess his limi a ion by de eloping AAN con ol s a egies o p o ide he
obo ic assis ance needed o pos -s oke pa ien s o comple e a mo emen , aking in o accoun hei
mo o impai men s and in oking hei pa icipa ion.
B. Clinical In o ma ion
Clinical ials in es iga ing he e ec s o wea able assis i e de ices on pos -s oke eco e y
had, on a e age, a di e se sample o pa icipan s, anging om 3 o 62, which p o ides a benchma k
o u u e esea ch. Pa icipan s in hese ials we e p ima ily cha ac e ized by age ( om 18 o 80
yea s old), gende (balanced gende dis ibu ion), ime pos -s oke ( a ying be ween sub-acu e
( om 25 days o 6 mon hs) and ch onic ( om 6 mon hs o 9.5 yea s) s ages), hemiplegic side,
absence o ca diopulmona y disease, and s oke e iology. To inc ease he eliabili y o u u e
Chap e 3
44
senso sys ems we e de eloped in he s udy [2]. Howe e , in his Ph.D. hesis, a new e sion o he
MuscLab sys em was p ojec ed and conc e ized (de ailed in Chap e 4).
The Foo Lab consis s o ins umen ed shoes wi h wo IMUs (one in he uppe o he shoe
(LSM6DS3) and he o he in he insole (LSMDSOX)) and an insole ins umen ed wi h 8 FSRs. Bo h senso s
enable he acquisi ion o he oo kinema ics and he c ea ion o a p essu e map on he su ace o he
oo , espec i ely. The sys em p esen s Blue oo h wi eless echnology and a echa geable ba e y.
The Ine ialLab is composed o se en IMUs (MPU6050) posi ioned a he pel is, igh /le highs,
shanks, and ee , measu ing he kinema ics o lowe limb segmen s.
The MuscLab (de ailed in Chap e 4) is an e- ex ile designed o moni o he kinema ics and he
muscle con ac ion o muscles in a human segmen . The sys em is made o an IMU (LSM6DS3) and
piezo esis i e ex ile s ips (Shieldex® Technik- ex P130+B), in eg a ing Blue oo h wi eless echnology
and a echa geable ba e y.
The EMG T igno A an i sys em in eg a es eigh senso s ha enable he acquisi ion o EMG
muscle ac i i y du ing di e en asks. Each EMG senso has a buil -in iaxial IMU, composed o an
accele ome e and gy oscope wi h a ansmission ange o 40 me e s and a echa geable ba e y. Table
3.1 summa izes he da a collec ed by each senso sys em.
Table 3.1 – Di e en ypes o da a collec ed by each senso sys em
Senso sys em
Senso da a
Foo Lab
➢ Fee kinema ics (angula speed and accele a ion)
➢ Plan a p essu e (p essu e map)
Ine ialLab
➢ Kinema ics (angula speed, accele a ion) o lowe limb segmen s ( igh and le oo ,
shank, high, unk)
MuscLab
➢ Kinema ics (angula speed, accele a ion) o he shank segmen
➢ Muscle con ac ion o he shank muscles
EMG T igno A an i
➢ Roo Mean Squa e up o 8 EMG signals
In his Ph.D. hesis, he EMG T igno A an i sys em was in eg a ed in o Sma Os sys em by using
he T igno SDK, o enable he eal- ime acquisi ion o EMG signals du ing he o hosis’ use [86]. Fo ha ,
a T ansmission Con ol P o ocol/In e ne P o ocol (TCP/IP) was implemen ed o communica e be ween
bo h sys ems (CCU o Sma Os sys em and he Base S a ion o T igno A an i sys em). To dec ease he
amoun o da a o be sen ia TCP/IP and o wo k wi h cleane EMG signals, he Roo Mean Squa e (RMS)
mode o he T igno A an i sys em was con igu ed (A an i-Only Modes: 83). This mode enables he

Chap e 3
45
acquisi ion o ec i ied EMG signals wi h a equency o 148 Hz, by applying he RMS me hod. By de aul ,
he RMS mode sends 27 EMG samples a e e y 0.0135 s (74 Hz). These samples a e ep ocessed in
Sma Os sys em by using he RMS a 74 Hz [88]. Figu e 3.2 depic s he EMG sample ansmission
be ween 1 EMG senso and he CCU o Sma Os. As men ioned, o each EMG senso , i is applied he
RMS me hod o he EMG aw da a collec ed om a speci ic muscle, e u ning a single EMG sample wi h
a equency o 148 Hz. These ec i ied EMG samples a e sen in packe s o 2 samples o he T igno A an i
Base S a ion ia a Blue oo h p o ocol in eg a ed in o he T igno A an i sys em, being hen sen ia UART
o an ex e nal compu e unning he T igno SDK. I is no ewo hy ha his ex e nal compu e is only
equi ed because he T igno SDK uns in he Windows ope a ing sys em, and no in he Ubun u ope a ing
sys em, which is he ope a ing sys em unning in he Sma Os’ CCU. Subsequen ly, he wo ecei ed
EMG samples a e esampled o 27 EMG samples inside he T igno SDK, being hen sen by Wi-Fi o he
CCU ha applies again he RMS me hod o e ain one EMG sample wi h a equency o 74 Hz.
Figu e 3.2 – Flowcha o acqui ing RMS EMG samples wi h he EMG T igno A an i sys em.
3.2.3 GAIT ANALYSIS TOOLS
In he Sma Os amewo k, all signals measu ed wi h he wea able mo ion lab can be subsequen ly
used in gai analysis ools. In he s udy [2], se e al gai analysis ools we e de eloped, namely, he
es ima ion o (i) spa io empo al pa ame e s (walking speed, s ep leng h, s ep wid h, s ide leng h,
dis ance); (ii) o ce pa ame e s (cen e o p essu e); (iii) lowe limb join angles (ankle, knee, hip) and
Chap e 3
46
segmen angles ( oo , shank, high, unk). This Ph.D. hesis ad ances by p esen ing mo e h ee gai
analysis ools, i.e., he abili y o decode LMs (Chap e 5), and o es ima e ankle join o ques
(Chap e 6.3), and ene gy expendi u e (Chap e 6.4), in eal- ime.
3.2.4 HIERARCHICAL CONTROL ARCHITECTURE
The Sma Os sys em endows a hie a chical con ol a chi ec u e o ganized in o h ee
le els, as sugges ed in he s udy [9]. Inspi ed by he human mo ion con ol sys em, his a chi ec u e
in eg a es bo h s uc u al and unc ional aspec s o con ol and senso eedback sys ems. A he high-
le el, known as he pe cep ion laye , he sys em gene a es use -o ien ed ajec o ies. The mid-le el ac s
as he ansla ion laye , being esponsible o ans o ming he use -o ien ed ajec o ies in o e e ence
ajec o ies o he AO in acco dance wi h walking speed. A las , he low-le el gene a es assis i e
commands, ensu ing ha he s a e o he AO e ec i ely acks he desi ed assis ance ajec o y in a imely
manne . Cu en ly, he con ol a chi ec u e includes low-le el posi ion-based and o que-based acking
con olle s using P opo ional In eg al De i a i e (PID) and Feedback-E o Lea ning (FEL) con ol,
espec i ely [89].
In e ms o con ol equency, he high- and mid-le el equencies we e se a 100 Hz, a su icien
a e o human-machine gai analysis. Con e sely, he low-le el ope a es a 1 kHz o allow high- equency
ope a ion, which is conduci e o e ec i e human-machine acking con ol loop dynamics. The so wa e
ou ines con olling he low- and mid-le el con olle s we e coded in C and implemen ed on he
STM32F407VGT mic ocon olle . In con as , he high-le el con olle s, coded in C++, a e execu ed wi hin
he CCU, which is housed wi hin a UDOO X86. This a chi ec u al amewo k embodies a modula design,
acili a ing scalabili y o inco po a e addi ional assis i e con ol s a egies as equi ed o expand Sma Os
in o a e sa ile obo -based gai aining solu ion.
Cu en ly, he Sma Os sys em includes se en use -cen e ed, closed-loop assis i e con ol
s a egies, as p esen ed in Table 3.2. These s a egies ha e been ca e ully designed o add ess he
he apeu ic pu poses iden i ied o he Sma Os sys em in pos -s oke gai aining. Taken oge he , hese
s a egies make Sma Os adap able o di e en he apeu ic app oaches and add ess bo h immedia e and
pe manen changes in mo o abili y. Each s a egy allows o gai speed adjus men s in he ange o 0.5
o 1.6 km/h, aking in o accoun he mechanical limi a ions o he AOs, he eby acili a ing gai aining
ac oss he di e en challenges inhe en in each s a egy. This Ph.D. hesis ad ances by p oposing he
AAN LM-d i en ajec o y (Chap e 6.2), he AAN EMG-based (Chap e 6.3), and he AAN HITL
(Chap e 6.4) con ols.
Chap e 3
47
Table 3.2 – Di e en con ol s a egies a ailable in he Sma Os sys em
Con ol
S a egy
The apeu ic
Pu pose
Bene i s
Ze o- o que
Con ol
Rehabili a ion
➢ Familia iza ion pe iod
➢ Muscle s eng hening
T ajec o y-
acking Posi ion
Con ol
Assis ance and/o
Rehabili a ion
➢ Use -o ien ed epe i i e gai aining
➢ Reco e y o he use ’s gai pa e n ( ange o mo ion (ROM)
and symme y)
Adap i e
Impedance
Con ol
Rehabili a ion
➢ Reco e y o he use ’s gai pa e n (ROM and symme y)
➢ In oke he use ’s ac i e pa icipa ion
➢ Muscle s eng hening
➢ Manual assis ance le el adjus men
➢ Long- e m eco e y o unc ional mo o abili ies
EMG-based
Con ol
Rehabili a ion
➢ In oke he use ’s ac i e pa icipa ion
➢ Muscle s eng hening
➢ Long- e m eco e y o unc ional mo o abili ies
AAN LM-d i en
T ajec o y
Con ol
Assis ance and/o
Rehabili a ion
➢ Use -o ien ed epe i i e gai aining in di e en LMs
➢ In oke he use ’s ac i e pa icipa ion
AAN EMG-based
Con ol
Rehabili a ion
➢ Reco e y o he use ’s gai pa e n (ROM and symme y)
➢ In oke he use ’s ac i e pa icipa ion
➢ Muscle s eng hening
➢ Au oma ic assis ance le el adjus men
➢ Long- e m eco e y o unc ional mo o abili ies
AAN HITL
Con ol
Assis ance
➢ Ene ge ic-e icien mo o eco e y
3.2.5 GRAPHICAL APPLICATION
The mobile g aphical applica ion acili a es he in ui i e con igu a ion o all Sma Os modules,
allowing he sys em o be se up o moni o ing and/o assis ance in aining sessions. This applica ion
mee s equi emen s such as (i) simpli ied and guided in e ac ion o as , na u al, and easy na iga ion,
and (ii) he use o explici g aphical componen s o be e use unde s anding, as ini ially de ined in he
s udy [2]. De eloped o he And oid ope a ing sys em, all messages a e ansmi ed ia he Blue oo h
Chap e 3
48
p o ocol o he Sma Os CCU. In his Ph.D. hesis, he mobile g aphical applica ion has been ex ended o
allow he selec ion o new senso s (MuscLab, and EMG T igno A an i sys em), algo i hms (ankle join
o que es ima ion, LM decoding, and ene gy expendi u e es ima ion), and con ol s a egies (AAN LM-
d i en ajec o y, AAN EMG-based, and AAN HITL con ols), as shown in Figu e 3.3.
Figu e 3.3 – Mobile g aphical applica ion.
Chap e 3
49
3.3 CONCLUSIONS
Fo deli e ing pe sonalized and use -o ien ed assis ance o pos -s oke pa ien s, he modula and
hie a chical a chi ec u e o he Sma Os sys em was ex ended. The MuscLab sys em (Chap e 4) was
de eloped and in eg a ed in o he Sma Os sys em o enable he moni o iza ion o mul iple lowe
limb muscles simul aneously, in eal- ime. Fu u e esea ch on his senso may allow (i) muscle a igue
es ima ion; (ii) LMs decoding; and (iii) lowe limb join and segmen angle es ima ion.
Addi ionally, he in eg a ion o an LM decoding ool (Chap e 5) and he de elopmen o an
AAN LM-d i en ajec o y con ol (Chap e 6.2) enables he adap a ion o he sys em dynamics
acco ding o he use ’s locomo ion in en ions.
The in eg a ion o he EMG T igno A an i sys em and he de elopmen o a join o que
es ima ion algo i hm acili a ed he design o an AAN EMG-based con ol s a egy (Chap e 6.3)
o assis he use as much and when needed.
Finally, he de elopmen o AAN HITL con ol (Chap e 6.4), based on eal- ime es ima ion o
ene gy expendi u e, makes i possible o op imize he con ol pa ame e s o he AOs o help use s educe
hei me abolic cos s.
The mobile g aphical applica ion was upda ed wi h hese new unc ionali ies o enable he
applica ion o hese s a egies du ing ehabili a ion he apies.

50
4. MUSCLAB SYSTEM
Chap e 4
51
This chap e desc ibes he MuscLab sys em, an elas ic and lexible ex ile band ha
simul aneously moni o s muscle con ac ion in ex enso and lexo muscles and he kinema ics o
he segmen in which i is loca ed. The chap e begins wi h an in oduc o y o e iew o he MuscLab
sys em, ollowed by he concep ual design and unc ionali ies o he sys em. In addi ion, his chap e
p esen s he echnological solu ions implemen ed, aking in o accoun he ha dwa e and so wa e
in e aces. The chap e ends wi h he sys em alida ion agains gold-s anda d acking echnologies.
4.1 INTRODUCTORY INSIGHT
Acco ding o he Wo ld Heal h O ganiza ion, musculoskele al diso de s (MSDs) a e he
wo ld's leading cause o human mo o disabili y. An es ima ed 1.71 billion people wo ldwide li e
wi h MSDs. Mos MSDs a e associa ed wi h he wo k con ex , seden a y li es yle, and p ac ice o spo s,
a a p o essional o ama eu le el [90]. As in he case o s oke, MSDs make i di icul o ca y ou daily
asks a home and wo k and a e he main eason o absen eeism and ea ly e i emen [91]. This
aises he need o de eloping solu ions o p omo e objec i e, non-in asi e, and eal- ime
moni o ing o muscle con ac ion and elaxa ion. Ideally, his solu ion should be e sa ile o use
in di e en con ex s, such as (i) heal h, o suppo he clinical diagnosis o he e olu ion o muscle
con ac ion o o se e as a ool o gai disabili y le el assessmen in pos -s oke pa ien s; (ii) spo s, o
analysis o muscle pe o mance and p edic ion o he isk o MSDs; (iii) wo k, o suppo he e gonomics
assessmen o muscle condi ion and p edic ion o he isk o MSDs.
Typically, su ace muscle con ac ion moni o ing is conduc ed by expensi e bu high-p ecision
sensing EMG equipmen . The gold-s anda d su ace EMG solu ions include he non-in asi e EMG
elec odes om Ul ium EMG (No axon, To on o, Canada) [92] and Pico EMG (Come a Sys ems, Newbu g,
USA) [93], o he d y EMG senso s such as T igno A an i (Delsys Inco po a ed, Na ick, USA) [87].
Howe e , he long- e m use o hese senso s can be a ec ed by (i) swea ing; (ii) empe a u e a ia ions;
and, (iii) mo emen s be ween he skin and he elec odes [42]. Fu he mo e, i se e al agonis and
an agonis muscles need o be moni o ed simul aneously, i will need as many senso s as he numbe o
muscles o moni o . This may esul in a mo e in usi e and ime-consuming solu ion o donning and
do ing, na owing i s p ac ical use.
Cu en challenges cen e on he de elopmen o easily wea able and cos -e ec i e senso s ha
enable he eal- ime moni o ing o se e al muscles simul aneously in a p ac ical way. In his connec ion,
his chap e p esen s a second e sion o he MuscLab sys em (ad ancing [94]), which co esponds o a
Chap e 4
52
low-cos , non-in usi e, sel -calib a ed, and s and-alone p o o ype o moni o he muscle con ac ion in
se e al muscles o a human body segmen , simul aneously.
4.2 CRITICAL ANALYSIS OF RELATED WORK
Di e en de ices ha e been p oposed o he non-in asi e e alua ion o muscle con ac ion. The
s udies [95], [96] p oposed he de elopmen o ga men s co e ed wi h d y EMG elec odes. In he
s udy [95], wo ga men s ( ouse s and a shi ) moni o ed he elec ical ac i i y o he muscles p esen in
he lowe limbs, uppe limbs, and o so, making a maximum o 19 muscles. In he s udy [96], a lexible
and elas ic ex ile band wi h h ee elec odes sewn pa allel o each o he was esponsible o measu ing
EMG signals in inge s, w is s, elbows, shoulde s, back, hips, knees, ankles, and neck. Conside ing he
wo pa en documen s p esen ed abo e [95], [96], i was e i ied ha he inclusion o EMG elec odes in
ex iles wi h lexible cha ac e acili a es he moni o ing o se e al muscle g oups simul aneously.
None heless, d y EMG elec odes need o be always in con ac wi h he use ’s skin.
As an al e na i e o o e come he limi a ions o using EMG, MMG has eme ged o muscula
analysis [97]–[103]. MMG senso s, al hough hey do no measu e elec ical signals om muscles, can
de ec whe he muscles a e con ac ing o elaxing, wi hou he need o be in di ec con ac wi h
he use 's skin. The p inciple unde lying he MMG senso s is ha muscle con ac ion is ypically
associa ed wi h muscle sho ening, causing an inc ease in muscle c oss-sec ional a ea, s i ness, ension,
and mechanical ib a ion. In his connec ion, an MMG senso (e.g., o ce, capaci i e, and ine ial senso )
posi ioned abo e he muscle may de ec i s con ac ion [97].
The s udies [97], [98] used a piezo esis i e o ce senso placed on a m muscles o moni o
hei con ac ion. Due o i s piezo esis i e p ope y, he e is a shape a ia ion o he o ce senso as he
muscle unde he senso is con ac ed; hus, a ying he senso esis ance. Bo h s udies e i ied ha he
MMG signals ha e a simila pa e n o he en elopes o he EMG signals. Mo eo e , a pa en documen
[99] desc ibed ano he MMG-based de ice o moni o ing muscle ac i i y, cap u ing muscle mechanical
ib a ions h ough an ine ial senso . The pa en documen [100] used a capaci i e senso o e he
muscle o measu e capaci ance a ia ions acco ding o muscle con ac ion and elaxa ion. F om hese
s udies [97], [98] and pa en documen s [99], [100], i was ound ha he use o o ce, ine ial, and
capaci i e senso s allows he moni o ing o muscle ac i i y o he muscles whe e he senso s a e loca ed.
Howe e , hese solu ions may become non-p ac ical o daily use when i is necessa y o moni o mul i-
Chap e 4
53
muscle g oups o a body segmen simul aneously since he numbe o MMG senso s inc eases
acco ding o he numbe o muscles o moni o .
S udies [101], [102], [103] p oposed a de ice ha can moni o he muscle con ac ion o se e al
muscles o muscle g oups simul aneously. The s udy [101] p esen ed a co d-shaped senso wi h
piezo esis i e p ope ies o measu e muscle con ac ion in he o ea m muscles. To do his, he co d
was placed a ound he o ea m segmen . The con ac ion o agonis and an agonis muscles changes he
leng h o he co d; hus, a ying i s esis ance. Despi e simul aneously moni o ing he con ac ion o
agonis and an agonis muscles, his solu ion canno disc imina e con ac ions o bo h muscles,
since bo h con ac ions cause an inc ease in he leng h o he co d.
On he o he hand, he s udy [102] disclosed a lexible ex ile band (72% nylon and 28%
spandex) o moni o ing he muscle con ac ion o he muscles in he shank segmen , while disc imina ing
which muscle is con ac ed. A lexible, piezo esis i e, and e- ex ile (conduc i e ex ile) has been
in eg a ed in o an ex e nal ex ile band and sewn in o a ma ix shape. The ex e nal ex ile band was
designed o he shank segmen and, due o i s ma ix con igu a ion, i is possible o disc imina e which
muscles o he shank egion a e con ac ed/ elaxed h ough he p essu e hey make on he di e en
zones o he ma ix. Simila ly, in he pa en documen [103], a ci cula band wi h p essu e senso s
is desc ibed o he simul aneous moni o ing o muscle con ac ion/ elaxa ion o muscles esponsible o
hand, w is , o a m mo emen s. Al hough hese wo app oaches [102], [103] can moni o he muscle
ac i i y o se e al muscles simul aneously, he ex ile bands do no ha e elas ic cha ac e is ics.
As such, hese solu ions equi e he de elopmen o a use -speci ic moni o ing de ice o ensu e he p ope
senso placemen acco ding o he use ’s an h opome y.
Conside ing all he s udies and pa en documen s p esen ed abo e, i becomes impe a i e o
de elop a solu ion based on MMG senso s sewn on o a lexible and elas ic ex ile band, so ha he same
de ice can be used o di e en an h opome ies and on di e en pa s o he human body. Fo his
pu pose, his Ph.D. hesis ad ances wi h he MuscLab sys em – a s and-alone MMG solu ion
based on e- ex ile (piezo esis i e ex ile) senso s sewn on o a lexible and elas ic ex ile
band. The MuscLab sys em was designed as a p oo o concep o simul aneously moni o and
disc imina e he muscle con ac ion o he shank segmen muscles (mainly, he
ibialis an e io
and
gas ocnemius la e alis
) in indi iduals wi h di e en an h opome ies.
Chap e 4
60
esis ance o each piezo esis i e ex ile s ip. Du ing his calib a ion s ep, he use mus emain in a si ing
posi ion (wi h he knee join s pe o ming 90º) and s a ic o 10 seconds, while he so wa e ou ine inds
he op imal posi ion among 256 possible p og ammable posi ions o he digi al po en iome e un il he
ou pu o he ADC eaches 0.75 V o each piezo esis i e ex ile s ip.
Once calib a ed, he algo i hm mo es on o a calib a ion ou ine o he second digi al po en iome e
(AD8400 50 kΩ). This po en iome e is esponsible o ampli ying he signal ead o each s ip, allowing
gains be ween 1 and 250. To calib a e i , he use was asked o pe o m dynamic con ac ions (maximum
olun a y con ac ions a e also alid) o 10 seconds. This calib a ion aimed o de e mine he gain
necessa y o ensu e ha he maximum ol age is 3.3 V, he maximum alue allowed by he A duino.
A e calib a ing he digi al po en iome e s, he A duino’s p og am wai s o ecei e a command
message om he MuscLab API. These commands a e ecei ed by he Nina p ocesso and sen ia UART
se ial communica ion o he Co ex-M0 32-bi SAMD21 o he A duino Nano 33 IoT. I he command sen
by he API is a s a command, i will ini ia e he collec ion o ine ial da a, piezo esis i e ex ile da a, and
ba e y ol age le el. This is done by eading he ba e y da a e e y 5 minu es and he ine ial and
piezo esis i e da a e e y 10 ms. The da a a e sen o he MuscLab API a 100 Hz h ough a da a message,
o ganized as ollows:
• by e o indica e he s a o heade o he packe (indica ed wi h he alue o he ‘h’
cha ac e );
• 1 by e comp ising he message numbe and he ba e y alue ( i s 4 bi s o he message
numbe , 0 o 15; and he las 4 bi s o he ba e y alue, (0, 10, 20, ..., 100 %) di ided by
10;
• The senso da a by es ( a iable size acco ding o he numbe o senso s eques ed by he
API, hese can be ine ial da a and/o da a om he piezo esis i e ex ile);
o Ine ial senso : 2 by es pe signal (accele ome e and gy oscope in X, Y, and Z
componen s) - a o al o 12 by es;
o Piezo esis i e ex ile senso : 2 by es pe s ip - a o al o 10 by es.
• 2 by es indica ing he end o he packe o he ail o he packe (indica ed wi h he alue o
he ‘ ’ cha ac e ).
Once he da a has been ecei ed by he API, he da a a e pa sed in o package ID, ba e y le el,
ime s amp, 3-axis accele ome e and gy oscope, and ol age ead o each piezo esis i e ex ile s ip. The
piezo esis i e da a is il e ed in he API by using a low-pass exponen ial il e wi h a cu o equency o 5

Chap e 4
61
Hz. The API hen sa es all da a in o a ex ile. The code is con inuously unning un il he use s ops i
h ough a s op command sen by he API.
Figu e 4.4 – Flowcha o he MuscLab acquisi ion sys em.
DP1
and
DP2
ep esen he Digi al
Po en iome e o he Whea s one b idge and summing in e ing ampli ie , espec i ely. The blue and
g een blocks ep esen he low diag am o he A duino and API codes, espec i ely.
S a
Swi ch on he sys em
Calib a e DP
Calib a e DP
DP calib a ed
DP calib a ed
Wai o a Blue oo h connec ion
Blue oo h connec ed
Wai o S a Command
S a Command
ecei ed
Collec ine ial, pie o esis i e, and
ba e y da a
Send collec ed da a
S op Command
ecei ed
S op da a collec ion
S a
Run API
Inpu di ec o y
De ices ound
Sea ch de ices o connec
Wai o a S a Command by he
use
S a Command
in oduced
Send S a Command o clien
Wai o new da a
New da a ecei ed
Sa e collec ed da a
Wai o a S op Command by he
use
S op Command
in oduced
S op da a collec ion
Connec o clien
yes
yes
yes
yes
yes
yes
yes
yes
yes
no
no
no
no
no
no
no
no
no
Chap e 4
62
4.3.5 EXPERIMENTAL VALIDATION
Two expe imen al es s we e ca ied ou o in e he ope abili y o he MuscLab sys em. Fi s ,
bench es s we e conduc ed o check i he equi emen s de ined o MuscLab we e me . Mo eo e ,
human es s we e pe o med o e alua e he eliabili y o he MuscLab signals agains EMG and
kinema ic signals du ing do si lexion and plan a lexion mo emen s o he ankle join .
A. Bench Tes s
A benching p o ocol was ca ied ou o e alua e he MuscLab sys em equi emen s, s a ed
in Table 4.1, ega ding he (i) acquisi ion equency (100 Hz); (ii) pe cen age o packe loss
o wi eless communica ion (below 5%); and (iii) au onomy (> 2h). The MuscLab mass and
dimensions we e also measu ed. The p o ocol consis ed o wo ials, whe e he MuscLab sys em
acqui ed and ansmi ed he ine ial and piezo esis i e ex ile da a a 100 Hz. The p o ocol s a ed
wi h a ba e y le el o 100%. The ba e y le el was measu ed wi h a mul ime e a he beginning o
he p o ocol and in 20-minu e in e als un il he sys em swi ched o due o lack o ba e y powe .
Du ing he da a collec ion, he ime s amp o each ecei ed sample a he API was egis e ed. The
ial du a ion was eco ded by a ch onome e o compu e he ba e y au onomy.
A e he da a collec ion, he ime be ween consecu i e samples was compu ed. To in e he
acquisi ion equency, he a e age ime be ween consecu i e samples was de e mined. Fu he , all
samples ecei ed wi h a ime be ween consecu i e samples o less han 9.9 ms o mo e han 10.1
ms we e conside ed los samples. The pe cen age o los packe s was calcula ed aking in o accoun
he ecei ed and los samples.
B. Human es s
Ten heal hy pa icipan s (5 males and 5 emales) we e in ol ed in he alida ion o he
MuscLab sys em (a e age age o 26.5 ± 2.8 yea s old, an a e age body mass o 65.4 ± 11.2 kg, and
an a e age body heigh o 169.8 ± 10.9 cm). The lowes and highes shank pe ime e s
measu ed we e 35.0 and 42.1 cm, espec i ely, achie ing an a e age alue o 37.6 ± 1.9 cm.
P io o he expe imen s, all pa icipan s ga e w i en in o med consen in acco dance wi h he e hical
guidelines es ablished by he E hics Commi ee o he Uni e si y o Minho (CEICVS 006/2020).
The p o ocol s a ed by equipping he pa icipan s, as illus a ed in Figu e 4.5. Fi s , he
pa icipan s we e ins umen ed wi h he EMG T igno A an i senso s [T igno A an i (Delsys
Inco po a ed, Na ick, USA)] o cap u e EMG signals om wo lowe limb muscles: he
ibialis an e io
Chap e 4
63
and he
gas ocnemius la e alis
o he igh leg. The senso s we e posi ioned abo e he e e ed
muscles, ollowing SENIAM ecommenda ions [105]. Fu he , he sys em was con igu ed o compu e
he RMS o he EMG signals a 100 Hz. Mo eo e , pa icipan s we e equipped wi h 7 IMUs ( o so,
highs, shanks, and ee ) om he Xsens Awinda sys em (Mo ella – Hende son, USA) o collec lowe
limb join angles a 100 Hz.
Figu e 4.5 – A pa icipan ins umen ed wi h he (i) T igno A an i and Xsens Awinda sys ems ( op iew);
(ii) T igno A an i, Xsens Awinda, and MuscLab (wi hou ou e band) sys ems (middle iew); and (iii) T igno
A an i, Xsens Awinda, and MuscLab (wi h ou e band) sys ems (bo om iew).
Chap e 4
64
Then, he pa icipan s we e ins umen ed wi h he MuscLab sys em a he igh shank. The
MuscLab sys em was posi ioned abo e he senso s om he T igno A an i and Xsens Awinda
sys ems, as depic ed in Figu e 4.5.
Pa icipan s we e hen asked o pe o m he N-pose +Walk calib a ion p ocedu e om he
Xsens Awinda sys em. Pa icipan s s ood in he N-pose o 3 seconds, hen walked 5 me e s o wa d
and backwa d in a s aigh line. A e wa d, pa icipan s we e asked o si o pe o m he MuscLab
calib a ion: emain s a ic o 10 seconds and hen pe o m epe i i e do si lexion and plan a lexion
mo emen s o 5 seconds. Once calib a ed, pa icipan s comple ed h ee independen es s pe
each mo emen ype: do si lexion and plan a lexion mo emen s. In all es s, he ime was con olled
by an ex e nal esea che .
The i s es consis ed o he ollowing sequence: 1) s a ed in a elaxed posi ion o 5
seconds (Figu e 4.6 – a)); 2) ca ied ou i e do si lexion epe i ions (Figu e 4.6 – b)) a h ee
cadences imposed by a me onome, wi h each cadence sepa a ed by a 5-second elaxed posi ion.
The chosen cadences we e 40, 75, and 105 bea s/minu e, which may co espond o walking speeds
o 1.0, 2.5, and 4.0 km/h [106]. This i s es aimed o analyze he sys em's esponsi eness o
di e en mo ion cadences. Fo ha , he delay be ween he MuscLab signals and he signals
collec ed by he T igno A an i and Xsens Awinda sys ems was compu ed, o each cadence.
The second es included wo cons an do si lexion mo emen s o 3 seconds each,
sepa a ed by a 3-second elaxed posi ion. This es aims o e alua e he MuscLab’s measu ing
epea abili y. In his es , o bo h he MuscLab and T igno A an i sys ems, he pe cen age o he
a e age magni ude a ia ion be ween he i s and he second do si lexion mo emen s was compu ed.
In he hi d es , he pa icipan s we e ins uc ed o execu e h ee cons an do si lexion
mo emen s o 2 seconds each, a di e en ankle do si lexion angles, i.e., (i) a small do si lexion
mo emen ; (ii) a do si lexion mo emen a an in e media e posi ion be ween he elaxed and he
maximum posi ions; and (iii) a maximum do si lexion mo emen . This magni ude es was pe o med
o e i y i he sys em can disc imina e di e en muscle con ac ion le els a di e en ankle
join angles. In his connec ion, he Coe icien o De e mina ion (
R2
) was compu ed be ween he
MuscLab and he Xsens Awinda signals.
Subsequen ly, he pa icipan s we e asked o pe o m he same h ee es s bu o plan a
lexion mo emen s (Figu e 4.6 – c)).
In addi ion, he co ela ion o he MuscLab sys em wi h he EMG signals o he
ibialis an e io
and
gas ocnemius la e alis
muscles was assessed unde he desc ibed do si lexion and plan a
Chap e 4
65
lexion mo emen s, espec i ely. Fo ha , he Spea man Co ela ion coe icien (
) was used o
e alua e he s eng h and di ec ion o he mono onic ela ionship be ween MuscLab and EMG signals.
A
p
- alue o 0.05 was used o de e mine whe he he obse ed co ela ion be ween he wo sys ems
was s a is ically signi ican . Acco ding o he s udy [107], he ela ionship can be ca ego ized as (i)
negligible i 0.0 <
< 0.19; (ii) weak i 0.20 <
< 0.29; (iii) mode a e i 0.30 <
< 0.39; (i ) s ong i
0.40 <
< 0.69; and ( ) e y s ong i
≥ . .
Figu e 4.6 – A pa icipan pe o ming a) a elaxed mo emen ; b) a do si lexion mo emen ; and c) a plan a
lexion mo emen .
4.4 RESULTS
4.4.1 BENCH TESTS
The MuscLab sys em showed an a e age (i) acquisi ion equency o 100.0 ± 0.007 Hz (≅
equi emen o 100 Hz); (ii) pe cen age o packe loss o 0.0972 ± 0.0318% (< 5%); and (iii) ba e y li e
o 2 hou s and 9 minu es (> 2h).
Rega ding he MuscLab dimensions, he inal leng h, wid h, and heigh o he case con aining he
elec onic boa d we e 5.4 cm (< 5.5 cm), 4.3 cm (< 4.5 cm), and 2.2 cm (< 2.5 cm), espec i ely. The
lexible ex ile band p esen ed a leng h and a wid h o 33.5 cm and 18.0 cm, espec i ely, allowing he
moni o ing o shank segmen s wi h pe ime e s anging om 33.5 o 48.7 cm.
The mass o he case and lexible ex ile band we e 40 g and 49 g, espec i ely, comp ising a o al
o 89 g (< 100 g).
a) Relaxed b) Do si lexion c) Plan a Flexion

Chap e 4
66
4.4.2 HUMAN TESTS
The MuscLab sys em was able o moni o he muscle con ac ion o indi iduals wi h a shank
pe ime e anging om 35.0 and 42.1 cm. Table 4.2 p esen s he esul s o he MuscLab sys em's
esponsi eness o di e en mo ion cadences. Acco ding o Table 4.2, all MuscLab signals we e, on
a e age, (i) delayed ega ding he EMG T igno A an i signals (135.8 ± 78.0 ms); and (ii) an icipa ed
ega ding he Xsens Awinda signals (-36.8 ± 112.2 ms). Mo eo e , he a e age delay o he MuscLab
signals measu ed by each s ip was consis en ac oss di e en cadences, since i a ied om (i) 135.5
and 136.2 ms, ega ding he EMG T igno A an i signals; and (ii) 34.5 and 38.8 ms ega ding he Xsens
Awinda signals.
Table 4.2 – A e age delay (s anda d de ia ion) be ween he MuscLab and EMG Delsys signals, and MuscLab and
Xsens Awinda signals, in milliseconds, o each mo ion cadence (40, 70, 105 bpm)
MuscLab
S ip
T igno A an i Sys em
Xsens Awinda Sys em
40 bpm
75 bpm
105 bpm
40 bpm
75 bpm
105 bpm
S1
138.0
(82.3)
125.0
(84.1)
134.0
(73.2)
- 5.0
(128.4)
- 4.5
(127.1)
- 8.5
(121.3)
S2
136.7
(71.6)
124.0
(79.3)
134.0
(94.3)
- 8.0
(121.4)
- 3.5
(126.7)
- 4.0
(122.1)
S3
129.0
(66.2)
134.0
(90.7)
120.0
(93.4)
- 66.1
(72.1)
- 69.4
(83.5)
- 60.0
(89.4)
S4
127.0
(92.5)
128.0
(95.0)
113.0
(93.2)
- 76.7
(90.1)
- 61.1
(104.4)
- 67.8
(90.8)
S5
146.7
(68.0)
170.0
(43.2)
178.3
(43.7)
- 38.0
(118.2)
- 46.7
(167.4)
- 32.0
(120.4)
A e age ±
STD*
135.5
(76.1)
136.2
(78.4)
135.9
(79.5)
- 38.8
(106.0)
- 37.0
(121.8)
- 34.5
(108.8)
To al
A e age ±
STD
135.8 (78.0)
-36.8 (112.2)
*STD means s anda d de ia ion
The esul s o epea abili y es s ocus on he compa ison o he piezo esis i e ex ile s ips numbe
1 and 2 (S1 and S2, espec i ely, o Figu e 4.1) wi h
ibialis an e io
signals o do si lexion mo emen s.
On he o he hand, o plan a lexion mo emen s, he piezo esis i e ex ile s ips numbe 3, 4, and 5 (S3,
Chap e 4
67
S4, and S5, espec i ely, o Figu e 4.1) we e compa ed o he
gas ocnemius la e alis
signals. In his
con ex , he
ibialis an e io
EMG, S1, and S2 signals e ealed close a e age a ia ions o muscle
con ac ion be ween wo pe o med do si lexion mo emen s, namely 3.17 ± 11.7%, 2.97 ± 4.32%, and
2.41 ± 3.77%. Mo eo e , he
gas ocnemius la e alis
EMG, S3, S4, and S5 signals e ealed an a e age
a ia ion o 4.88 ± 13.7%, 5.30 ± 8.98%, 3.47 ± 8.68%, and -20.2 ± 94.9%.
In ela ion o he magni ude es esul s, an
R2
o (i) 0.91 ± 0.12 and 0.92 ± 0.08 was ound
be ween he do si lexion ankle angle and he S1 and S2 s ips, espec i ely; and (ii) 0.92 ± 0.15, 0.85 ±
0.18, and 0.69 ± 0.22 was ound be ween he plan a lexion angle and he S3, S4, and S5 s ips,
espec i ely.
Fu he mo e, Table 4.3 p esen s he
alues ob ained be ween he MuscLab and EMG T igno
A an i signals, pe pa icipan . The esul s p esen ed in Table 4.3 indica ed e y s ong co ela ions
be ween he MuscLab signals a S1, S2, S3, and S4 s ips and EMG T igno A an i signals, epo ing
a e age
alues abo e 0.78 ± 0.08 (
p
- alue < 0.05). In de ail, e y s ong co ela ions we e achie ed
be ween (i)
ibialis an e io
signals and bo h S1 and S2 signals (
= 0.78 ± 0.07); and (ii)
gas ocnemius
la e alis
signals and bo h S3 and S4 (
= 0.78 ± 0.09 and 0.77 ± 0.07, espec i ely). The S5 signals we e
mode a ely co ela ed wi h he
gas ocnemius la e alis
signals (
= 0.38 ± 0.23). Mo eo e , low s anda d
de ia ion alues we e e i ied o S1, S2, S3, and S4 (< 0.09), while he S5 p esen ed he highes s anda d
de ia ion alue (0.23).
Addi ionally, Figu e 4.7 depic s he EMG T igno A an i, Xsens Awinda, and MuscLab signals o a
andom pa icipan execu ing he alida ion p o ocol, i.e., esponsi eness, epea abili y, and magni ude
es s.
Table 4.3 – Spea man Co ela ion coe icien (
) be ween he MuscLab and EMG sys ems. A
p
- alue below 0.05
was e i ied in all co ela ions
Pa icipan
ID
Shank
Pe ime e
(cm)
Shank
Leng h
(cm)
S1
S2
S3
S4
S5
1
42.1
33.0
0.84
0.83
0.80
0.76
0.33
2
38.2
33.0
0.70
0.74
0.77
0.84
0.48
3
38.8
35.8
0.86
0.88
0.58
0.71
0.49
4
37.9
31.5
0.74
0.71
0.66
0.74
0.75
5
35.0
41.5
0.84
0.87
0.87
0.63
0.75
6
36.5
36.4
0.67
0.70
0.81
0.77
0.22
Chap e 4
68
Pa icipan
ID
Shank
Pe ime e
(cm)
Shank
Leng h
(cm)
S1
S2
S3
S4
S5
7
37.4
35.0
0.74
0.75
0.79
0.83
0.17
8
37.9
35.0
0.70
0.72
0.81
0.75
0.41
9
35.3
34.4
0.88
0.85
0.89
0.84
0.07
10
37.2
33.2
0.83
0.71
0.77
0.86
0.15
A e age ±
STD
37.6 ± 1.9
34.9 ± 2.6
0.78 ±
0.07
0.78 ±
0.07
0.78 ±
0.09
0.77 ±
0.07
0.38 ±
0.23
Figu e 4.7 – A pa icipan pe o ming do si lexion ( op iew) and plan a lexion (bo om iew) mo emen s.
Boxes 1, 2, and 3 ep esen he speed, epea abili y, and magni ude es s, espec i ely. Fo easie
isualiza ion, he magni ude o EMG signals was scaled o MuscLab signal magni udes using a ac o o
x10,000.
Chap e 4
69
4.5 DISCUSSION
This s udy p oposes a lexible and elas ic ex ile band ins umen ed wi h piezo esis i e ex ile
senso s o moni o he muscle con ac ion o indi iduals wi h di e en an h opome ies. In an a emp o
add ess he s a e-o - he-a limi a ions, he p oposed sys em ad ances (i) s udies [95]–[100] by
moni o ing muscle con ac ion o di e en muscle g oups, simul aneously; (ii) s udy [101] by
disc imina ing he muscle con ac ion o he agonis and an agonis muscles wi h a single senso ; and
(iii) s udies [102], [103] by moni o ing he muscle con ac ion o indi iduals wi h di e en an h opome ies
( anging om 35.0 o 42.1 cm) due o he elas ic na u e o he p oposed solu ion.
In e ms o esponsi eness pe o mance, he MuscLab signals we e delayed ega ding EMG
signals (a e age delay o 135.8 ± 78.0 ms). Acco ding o he s udy [97], he e is a ime delay
be ween he onse o EMG and he onse o o ce gene a ion, de ined as he elec omechanical delay.
This delay ypically a ies be ween 9 and 130 ms [108]. Since an MMG, when posi ioned in a muscle,
measu es he esponse o a o ce gene a ion, i s measu ed signals a e expec ed o be delayed wi h
espec o he EMG signals. In his con ex , he a e age delays epo ed by he MuscLab sys em (135.8
± 78.0 ms) a e in line wi h wha would be expec ed because he MuscLab sys em is an MMG senso ; i ,
he e o e, de ec s muscle con ac ion only a e he muscle elec ical ac i a ion (signal measu ed by
EMG).
Con e sely, an a e age delay o -36.8 ± 112.2 ms was ound be ween he MuscLab and
Xsens Awinda signals. Despi e he high s anda d de ia ion alue, hese esul s sugges ha he
MuscLab signals may be an icipa ed ega ding ankle join kinema ics. These esul s may be jus i ied by
he ac ha muscle con ac ions occu om 20 o ms be o e he use ’s lowe limb kinema ic mo ion
[108].
Fu he mo e, esul s indica ed ha he a e age delay o he MuscLab signals emained
consis en ac oss he es ed cadences. These esul s a e aligned wi h hose p esen ed in he s udy
[104] since om a compa ison be ween six een e- ex iles, he Shieldex® Technik- ex P130+B (used in
he MuscLab sys em) showed he quickes esponses be ween elaxing and s e ching mo emen s. Thus,
he MuscLab sys em can be employed o moni o muscle con ac ions pe o med a cadences be ween
40 and 105 bpm.
In addi ion, MuscLab demons a ed sensing epea abili y when measu ing simila muscle
con ac ions. Conside ing he epea abili y es s, i was ound ha (i) he a ia ion in he magni ude o he
signals measu ed in s ips S1 and S2 (2.97 ± 4.32% and 2.41 ± 3.77%.) was, on a e age, close o he
a ia ion in he magni ude o he
ibialis an e io
signals (3.17 ± 11.7%); and (ii) he a ia ion in he
Chap e 5
76
Figu e 5.1 – Sma Os’ mobile g aphical applica ion (le iew) and a male pa icipan ins umen ed wi h
he Sma Os, T igno A an i, and Ine ialLab sys ems ( igh iew).
C. Expe imen al P o ocol
Once ins umen ed wi h he eigh EMG and se en IMU senso s, pa icipan s pe o med wo
Maximum Volun a y Con ac ions (MVCs) o each muscle o no malize all EMG signals, ollowing he
p o ocol desc ibed in he s udy [121]. Subsequen ly, pa icipan s we e equipped wi h he Sma Os
sys em and p o ided a 10 min- amilia iza ion session wi h he de ice ope a ing unde ze o- o que
con ol.
The pa icipan s hen comple ed a con inuous ci cui o nine ask sequences while he ankle
o hosis ope a ed in ze o- o que con ol (Figu e 5.2). The sequences, p esen ed in Figu e 5.2,
consis ed o : 1. s anding o 10 seconds (S ); 2. walking in a s aigh line o 2.5 me e s (LGW); 3.
descending s ai s wi h 13 s eps (SD); 4. walking again in a s aigh line o 3.5 me e s (LGW); 5.
s opping in he s anding posi ion o 5 seconds (S ); 6. walking in a s aigh line o 3.5 me e s (LGW);
7. ascending s ai s wi h 13 s eps (SA); 8. walking in a s aigh line o an addi ional 2.5 me e s (LGW);
and 9. s opping in he s anding posi ion o 5 seconds (S ). This ci cui was epea ed i e imes pe
pa icipan . Du ing dynamics asks, pa icipan s main ained a ixed cadence dic a ed by a me onome
a 40 bpm. Depending on he pa icipan 's heigh and subsequen s ep leng h, his cadence
co esponded o gai speeds anging om 1.0 o 1.5 km/h, as alida ed in he s udy [106]. This ange
o alues alls in he slow speeds ypically adop ed by indi iduals wi h mo o disabili ies [77].
Addi ionally, all pa icipan s we e able o pe o m each ansi ion wi h he sel -selec ed limb,
excep in he SD ask. A his ask, pa icipan s we e ins uc ed o ini ia e he SD ask wi h hei igh
oo , making i he leading limb o ad ance o he nex s ep. Following he ad ancemen wi h he igh

Chap e 5
77
oo , he le oo was hen mo ed o he same s ep. This SD p ocedu e was ollowed due o wo
easons. Fi s , i aligns wi h he ad ice o pa ien s wi h lowe limb impai men s o descend s ai s wi h
hei impai ed leg as he leading limb [122]. Second, he ankle o hosis is mechanically limi ed o a
maximum do si lexion angle alue o 20º (a alue easily exceeded when descending s ai s wi h he
non-ins umen ed leg (le leg) being he i s o ad ance).
Figu e 5.2 – A andom male pa icipan pe o ming he ou LMs (S , LGW, SD, SA) in he i s scena io.
Posi ions (1) and (9) deno e he s a ing and ending posi ions, espec i ely.
5.3.2 DATA PREPARATION
Du ing he p o ocol, a esea che anno a ed he ins an o ansi ioning be ween wo consecu i e
LMs. This ins an co esponded o a momen be ween he pa icipan ha ing li ed hei leading oo o
he g ound and he c i ical momen ( he ins an when he pa icipan placed he oo on he new LM [44]).
This ins an was eco ded using he Sma Os’ mobile g aphical applica ion (Figu e 5.1) o be synch onous
wi h he emaining collec ed da a. A e da a collec ion, all ials unde wen manual inspec ion by wo
Chap e 5
78
addi ional esea che s o ensu e eliabili y in labeling he ansi ion ins an s. A his le el, ou classes
we e de ined, namely, 0 – S , 1 – LGW, 2 – SD, and 3 – SA.
Conside ing he e iewed s udies [41], [43]–[46], [62], [114]–[117], se e al ypes o da a we e
employed in decoding LMs, namely ea u es om EMG signals [114], [115], and segmen [43]–[46],
[116] and join angles [41], [117]. The e is s ill no consensus on wha ype o da a yielded mo e accu a e
LM decoding. Thus, he LM decoding pe o mance was compa ed using ou ypes o senso da a: (i)
RMS EMG no malized by MVC (one o he mos used EMG ea u es [114], [115]); (ii) he segmen
angles ( o so, highs, shanks, and/o ee ) and hei i s and second-o de de i a i es, co esponding o
he angula eloci y and angula accele a ion; (iii) he lowe limb join angles (hips, knees, and/o
ankles), and hei i s and second-o de de i a i es; and, (i ) he magni ude o 3D accele a ion and
angula eloci y o he o so, high, shank, and/o oo segmen s. In he segmen o join angle da a
ypes, hei i s and second-o de de i a i es we e compu ed, since hese a iables imp o ed he model’s
pe o mance in empi ical es s. The ou h da a ype aims o explo e he po en ial o di ec ly using IMUs’
da a, which was no ye explo ed in he s a e-o - he-a o LM decoding. Equa ion 5.1 p esen s an example
o compu ing he magni ude o 3D accele a ion da a (‖
a
‖), in which
ax, ay, a
ep esen he accele a ion
measu ed in he x-axis, y-axis, and z-axis, espec i ely.
‖
a
‖
=
√
ax
ay
a
(5.1)
O e all, a o al o 38 inpu combina ions we e c ea ed and compa ed. They a e iden i ied by an
ID in Table 5.1. The da a was o ganized in o sequences made up o
X
columns and
Y
ows. While
X
ep esen s he numbe o samples o ganized sequen ially by ime and pa icipan s, he
Y
ows ep esen
he numbe o inpu s.
Table 5.1 – Iden i ica ion o he ype o da a used o decode LMs. The join angles, magni ude o 3D aw
accele ome e and gy oscope, segmen angles, and EMG da a a e shaded in g een, yellow, g ay, and ed,
espec i ely
ID
Inpu s
ID
Inpu s
1
Ankles JAD*
20
MAG (pel is and highs)
2
Hips JAD
21
MAG (pel is)
3
Ankles and hips JAD
22
Fee SAD***
4
Knees JAD
23
Fee and shanks SAD
5
Ankles and knees JAD
24
Shanks SAD
6
Knees and hips JAD
25
Fee and highs SAD
7
Ankles, knees, and hips JAD
26
Shanks and highs SAD
Chap e 5
79
ID
Inpu s
ID
Inpu s
8
MAG** ( ee )
27
Thighs SAD
9
MAG ( ee and shanks)
28
Pel is and ee SAD
10
MAG (shank)
29
Pel is, shanks, and ee SAD
11
MAG ( ee and highs)
30
Pel is and shanks SAD
12
MAG (shanks and highs)
31
Pel is, highs, and ee SAD
13
MAG ( highs)
32
Pel is, highs, shanks, and ee SAD
14
MAG (pel is and ee )
33
Pel is, highs, and shanks SAD
15
MAG (pel is, shanks, and ee )
34
Pel is and highs SAD
16
MAG (pel is and shanks)
35
Pel is SAD
17
MAG (pel is, highs, and ee )
36
EMG (TA, GL, RF, and BF)
18
MAG (pel is, highs, shanks, and ee )
37
EMG (RF, and BF)
19
MAG (pel is, highs, and shanks)
38
EMG (TA, and GL)
*JAD means Join Angles and De i a i es (angula eloci y and accele a ion).
**MAG means magni ude o 3D aw accele ome e and gy oscope.
***SAD means Segmen Angles and De i a i es (angula eloci y and accele a ion).
5.3.3 CLASSIFICATION MODELS
In he scope o his hesis, Long Sho -Te m Memo y (LSTM), CNN, and T ans o me s we e
explo ed since hese models e ealed high gene aliza ion abili y and high pe o mances in p e ious
s udies and a e adequa e o ime se ies [123]–[125]. All models we e implemen ed and e alua ed in
Ma lab® (2023b, The Ma hwo ks, MA, USA), unning in a Hewle -Packa d compu e (In el® Co e™ i7-
4710MQ CPU @ 2.50 GHz p ocesso and 16.0 GB andom access memo y).
Se e al empi ical analyses we e conduc ed o ind he bes LM decoding ool. Rega ding LSTM, i
was s udied (i) he numbe o neu ons (5, 10, 15, 20, 25, 30, 35, 40); and (ii) he numbe o
LSTM laye s
(1, 2, 3). In he case o CNN, he hype pa ame e s s udied we e (i) he numbe o il e s (8, 16, 32, 64,
128); (ii) he ke nel size (10 o 100 wi h inc emen s o 5); (iii) he numbe o con olu ional laye s (1, 2,
3); and (i ) he pooling laye (a e age and max pooling laye ). The hype pa ame e s explo ed in he
T ans o me we e (i) he numbe o encode laye s (1 and 2); (ii) he numbe o heads (2, 4, 8, 16, 32);
and (iii) he embedding size (128, 256).
Fo h ee DL models, he ollowing pa ame e s we e explo ed: (i) d opou a e (be ween 0% and
80% wi h inc emen s o 10%); (ii) he ba ch size (8, 16, 32, 64, 128, and 254); (iii) he sequence leng h
(80, 100, 120, 140, 160, 180, and 200); and (i ) he no maliza ion me hod ( obus , max-min, and z-
Chap e 5
80
sco e). Fu he mo e, i was u ilized (i) he adap i e momen es ima ion op imiza ion algo i hm o op imize
he pa ame e s o he neu al ne wo ks, including weigh s and biases [126]; (ii) he ocal c oss-en opy
loss o add ess po en ial class imbalance [127]; and (iii) a ious da a augmen a ion echniques such as
ji e ing, scaling, magni ude wa ping, and a combina ion o scaling and magni ude wa ping. These da a
augmen a ion echniques we e employed o inc ease he di e si y o he a ailable da a while p ese ing
accu a e labels [128], [129].
To de e mine he op imal hype pa ame e s and he bes LM decoding ool, a lea e-one-subjec -ou
c oss- alida ion (LOSOCV) was pe o med. Two pa icipan s, selec ed andomly om he i een
pa icipan s in he s udy (a 25-yea -old able-bodied emale wi h a body mass o 68.3 kg and a heigh o
1.65 m, and a 27-yea -old able-bodied male wi h a body mass o 83.0 kg and a heigh o 1.70 m), we e
chosen o alida e and es he model, o line, espec i ely. Subsequen ly, he model was ained using
da a om he emaining hi een pa icipan s.
5.3.4 EVALUATION METRICS
Six e alua ion me ics we e compu ed o assess he pe o mance o he LM decoding ool, namely,
he ACC, Ma hew’s Co ela ion Coe icien (MCC), F1-sco e, success a e pe class,
compu a ional load
, and
p edic ion ime
.
The
compu a ional load
(in ms) deno es he ime necessa y o he LM decoding ool o classi y he
LM upon ecei ing new da a. The
p edic ion ime
(in ms and as a pe cen age o he gai cycle) is
de e mined by Equa ion 5.2, in which (i)
Tc i ical
ep esen s he c i ical ins an ; and (ii)
Tp edic ion
is ins an in
which he LM decoding ool pe o ms a classi ica ion. Thus,
p edic ion ime
was compu ed and a e aged
o six ansi ions: S -LGW, LGW-S , LGW-SD, SD-LGW, LGW-SA, and SA-LGW. The
p edic ion ime
was
also assessed as a pe cen age o he gai cycle o unde s and a wha phase o he gai cycle he new LM
is decoded.
P edic ion ime = Tc i ical
Tp edic ion
(5.2)
5.3.5 EXPERIMENTAL VALIDATION
Once ained, he model was con e ed o he ONNX o ma , in o de o be in eg a ed in o he high-
le el o he Sma Os sys em. This model con e sion was done since he ONNX pla o m inc eases he
Chap e 5
81
po abili y and in e ope abili y o AI models [130]. A e in eg a ing he ained model in o he Sma Os
sys em, he LM decoding ool was e alua ed unde eal- ime p ocedu es, as ollows.
A. Pa icipan s
To alida e he LM decoding ool in eal- ime, h ee pa icipan s we e included: wo able-bodied
pa icipan s (one male and one emale wi h an a e age age, body mass, and body heigh o 26.0 ±
1.00 yea s old, 76.5 ± 6.5 kg, and 167.5 ± 2.5 cm) and one emale pa icipan ha su e ed a s oke
(age: 22 yea s old; body mass: 51.0 kg; body heigh : 160.0 cm; pa e ic side: igh ; s oke ype:
ischemic; s oke ime: 9 mon hs; FMA-LE: 26). I is no ewo hy ha he model aining p ocess did no
in ol e hese pa icipan s. The p oposed con ol s a egy was alida ed wi h he wo able-bodied
pa icipan s. All pa icipan s p o ided w i en and in o med consen , conside ing he Uni e si y o
Minho E hics Commi ee (CEICVS 006/2020).
B. Expe imen al P o ocol
In he beginning, pa icipan s we e ins umen ed wi h he senso con igu a ion ha p o ided
he bes classi ica ion pe o mance, along wi h he ankle o hosis. To ensu e amilia i y wi h he de ice,
a 10-minu e walking ial was pe o med wea ing he Sma Os sys em ope a ing unde ze o- o que
con ol.
Once amilia ized wi h he de ice, expe imen s we e conduc ed wi h he Sma Os sys em
ope a ing unde ze o- o que con ol. In hese expe imen s, able-bodied pa icipan s we e asked o walk
a di e en gai speeds, while he pos -s oke pa ien was asked o walk a a sel -selec ed speed, as
summa ized in Table 5.2. Fi s ly, pa icipan s walked a sel -selec ed speeds. Then, pa icipan s we e
asked o walk a ou di e en cadences, s a ing wi h 35 bea s/minu e un il 50 bea s/minu e, wi h
inc emen s o 5 bea s/minu e. The alues o he chosen cadences we e highe and lowe han he
alues chosen du ing he da a collec ion pe o med o ain he DL models. This was done o s udy
how he LM decoding ool pe o ms in speed condi ions di e en om hose he ool was ained.
Du ing hese expe imen s, he able-bodied pa icipan s walked in h ee di e en scena ios and,
o each scena io, walked a di e en gai speeds. This alida ion aims o explo e he obus ness o
he de eloped LM decoding ool in di e en en i onmen s. The i s scena io (Figu e 5.2) consis s o
he same en i onmen whe e da a we e collec ed o ain he DL model. I consis s o an indoo
scena io composed o 13 s eps wi h dimensions o 18.0 cm (heigh ), 110 cm (wid h), and 31.0 cm
(dep h). The second scena io (Figu e 5.3 – le iew) ep esen s ano he indoo scena io composed o

Chap e 5
82
8 s eps wi h he same dimensions as scena io 1. The hi d scena io (Figu e 5.3 – igh iew) is an
ou doo scena io consis ing o 13 s eps wi h dimensions o 16.0 cm (heigh ), 500 cm (wid h), and
30.0 cm (dep h).
The p o ocol adop ed o eal- ime expe imen s in he i s scena io was he same pe o med
du ing he da a collec ion o aining he DL models. Fo he second and hi d scena ios, pa icipan s
comple ed a con inuous ci cui o nine ask sequences, as depic ed in Figu e 5.3, namely: 1. s anding
o 10 seconds (S ); 2. walking in a s aigh line o 3.5 me e s (LGW); 3. ascending s ai s (SA); 4.
walking again in a s aigh line o 2.5 me e s (LGW); 5. s opping in he s anding posi ion o 5 seconds
(S ), u ning 180º, and s opping again in he s anding posi ion o mo e 5 seconds (S ); 6. walking in
a s aigh line o 2.5 me e s (LGW); 7. descending s ai s (SD); 8. walking in a s aigh line o an
addi ional 3.5 me e s (LGW); and 9. s opping in he s anding posi ion o 5 seconds (S ). A ask
sequence numbe 5, he da a co esponding o he u ning ask we e pos e io ly dele ed, since he LM
decoding was no ained unde u ning asks. Du ing his p o ocol, a esea che anno a ed he ins an
o ansi ioning be ween wo consecu i e LMs using he Sma Os’ mobile g aphical applica ion (Figu e
5.1). These e en s we e hen inspec ed by wo addi ional esea che s o ensu e accu acy in labeling
he ansi ion ins an s.
Figu e 5.3 – A heal hy es pa icipan pe o ming he ou LMs (S , LGW, SD, SA) in he second scena io
(le iew) and he hi d scena io ( igh iew). Posi ions (1) and (9) deno e he s a ing and ending
posi ions, espec i ely.
Chap e 5
83
Table 5.2 – Iden i ica ion o he condi ions o eal- ime es s pe o med by he able-bodied pa icipan s and a pos -
s oke pa ien
Pa icipan
Con ol Mode
Scena io
Cadence
Able-bodied
Ze o- o que mode
1
Sel -selec ed
35 bea s/minu e
40 bea s/minu e
45 bea s/minu e
50 bea s/minu e
2
3
Pos -s oke
Ze o- o que mode
2
Sel -selec ed
5.4 RESULTS
5.4.1 DATA BALANCING
A o al o 954,298 samples we e collec ed o ain he DL models, dis ibu ed as ollows: 211,381
o he S ask, 290,710 o he LGW, 221,928 o he SD, and 230,279 o he SA. This analysis is
complemen ed by Figu e 5.4, which shows he mean and s anda d de ia ion o each class, conside ing
he numbe o pa icipan s. Figu e 5.4 shows ha he S and SD asks ep esen he mino i y classes,
while he LGW is he majo i y class. Ne e heless, i was obse ed ha he mean numbe o samples pe
class is ela i ely consis en , anging om 22% o 30%. Fu he mo e, he amoun o each class pe
pa icipan shows minimal a iabili y, as indica ed by low s anda d de ia ion alues anging om 1.4% o
2.3%. Based on hese indings, he da ase is conside ed o be balanced.
Figu e 5.4 – Mean and s anda d de ia ion alues o he sample dis ibu ion pe class ac oss subjec s. S ,
LGW, SD, and SA mean S anding, Le el-G ound Walking, S ai Descen , and S ai Ascen , espec i ely.
Chap e 5
84
5.4.2 INPUT DATA ANALYSIS
Table 5.3 shows he esul s ob ained by he LOSOCV me hod o each DL model and o each se
o inpu s. The achie ed esul s indica ed ha he lowe pe o mances o all DL models (LSTM, CNN, and
T ans o me s) we e e i ied when using he RMS EMG as inpu (ACC, MCC, and F1-sco e below 0.794 ±
0.078, 0.561 ± 0.131, and 0.664 ± 0.098, espec i ely). Among hese da a, i was also obse ed ha
using RMS EMG signals om he
ibialis an e io
and
gas ocnemius la e alis
muscles p o ided simila
pe o mances o he ones achie ed when combining hese muscles wi h
ec us
, and
biceps emo is
muscles.
Wi hin he join angle da a ype, all DL models showed hei weakes pe o mance (a e age ACC,
MCC, and F1-sco e be ween 0.915 and 0.954, 0.796 and 0.883, and 0.846 and 0.912, espec i ely)
when using he angles o bo h ankles oge he wi h hei de i a i es (angula eloci ies and accele a ions)
as inpu . Con e sely, he models pe o med bes when he angles, angula eloci ies, and angula
accele a ions o he knees and hips we e used as inpu (a e age ACC, MCC, and F1-sco e be ween 0.976
and 0.980, 0.937 and 0.950, and 0.953 and 0.962, espec i ely).
Conce ning he da ase composed o he magni ude o 3D da a de i ed om accele ome e s and
gy oscopes, he h ee DL models showed hei wo s pe o mance when using hese inpu da a om
pel ic IMU only (a e age ACC, MCC, and F1-sco e be ween 0.482 and 0.900, 0.306 and 0.765, and
0.318 and 0.820, espec i ely). Con e sely, he bes pe o mance was obse ed wi h he magni ude o
3D accele a ion and angula eloci y om he pel ic, high and shank IMUs o he LSTM and T ans o me
models (a e age ACC, MCC, and F1-sco e be ween 0.968 and 0.979, 0.917 and 0.947, and 0.938 and
0.960, espec i ely), and om he high and shank IMUs speci ically o he CNN (a e age ACC, MCC,
and F1-sco e o 0.975 ± 0.007, 0.935 ± 0.019, and 0.952 ± 0.014, espec i ely).
Finally, when conside ing he da a ype combining he segmen angles along wi h hei de i a i es,
o all DL models, i was e i ied ha using (i) he pel ic segmen angles and hei de i a i es educed he
models' pe o mance (a e age ACC, MCC, and F1-sco e be ween 0.732 and 0.794, 0.471 and 0.571,
and 0.581 and 0.662, espec i ely); and (ii) he high and shank segmen angles and hei de i a i es
imp o ed he model pe o mance (a e age ACC, MCC, and F1-sco e be ween 0.979 and 0.982, 0.946
and 0.953, and 0.960 and 0.965, espec i ely).
O e all, he ype o da a ha p o ided he bes pe o mance ac oss all DL models was achie ed
using he high and shank segmen angles along wi h hei de i a i es as inpu da a.
Chap e 5
85
Table 5.3 – LOSOCV me ics o all inpu combina ions and DL models. The bes and he wo s esul s a e colo ed
in blue and ed, espec i ely. The highes pe o mance achie ed is shaded in g een colo
Inpu
ID
LSTM
CNN
T ans o me
ACC*
MCC*
F1-
sco e*
ACC*
MCC*
F1-
sco e*
ACC*
MCC*
F1-
sco e*
1
0.954
(0.023)
0.883
(0.051)
0.912
(0.040)
0.940
(0.023)
0.853
(0.051)
0.889
(0.040)
0.915
(0.030)
0.796
(0.064)
0.846
(0.051)
2
0.967
(0.023)
0.917
(0.054)
0.937
(0.042)
0.967
(0.019)
0.915
(0.046)
0.936
(0.035)
0.931
(0.039)
0.837
(0.079)
0.874
(0.066)
3
0.978
(0.013)
0.945
(0.032)
0.958
(0.025)
0.971
(0.024)
0.928
(0.055)
0.946
(0.043)
0.969
(0.019)
0.921
(0.046)
0.941
(0.036)
4
0.973
(0.010)
0.931
(0.026)
0.949
(0.019)
0.966
(0.014)
0.913
(0.034)
0.935
(0.026)
0.950
(0.024)
0.877
(0.052)
0.907
(0.042)
5
0.974
(0.016)
0.934
(0.039)
0.950
(0.030)
0.968
(0.024)
0.920
(0.055)
0.940
(0.043)
0.965
(0.001)
0.913
(0.059)
0.934
(0.046)
6
0.980
(0.010)
0.950
(0.024)
0.962
(0.019)
0.979
(0.013)
0.945
(0.031)
0.959
(0.024)
0.976
(0.010)
0.937
(0.027)
0.953
(0.020)
7
0.978
(0.011)
0.944
(0.028)
0.958
(0.022)
0.975
(0.015)
0.935
(0.036)
0.951
(0.028)
0.974
(0.012)
0.933
(0.031)
0.950
(0.024)
8
0.959
(0.011)
0.895
(0.028)
0.921
(0.021)
0.951
(0.017)
0.875
(0.039)
0.907
(0.030)
0.898
(0.025)
0.757
(0.052)
0.816
(0.041)
9
0.975
(0.011)
0.936
(0.029)
0.952
(0.022)
0.973
(0.010)
0.930
(0.026)
0.948
(0.019)
0.954
(0.011)
0.883
(0.026)
0.909
(0.020)
10
0.974
(0.011)
0.933
(0.028)
0.950
(0.021)
0.969
(0.011)
0.922
(0.027)
0.941
(0.021)
0.917
(0.021)
0.799
(0.045)
0.849
(0.036)
11
0.971
(0.009)
0.926
(0.024)
0.945
(0.018)
0.971
(0.009)
0.926
(0.023)
0.944
(0.018)
0.965
(0.009)
0.910
(0.023)
0.932
(0.017)
12
0.978
(0.007)
0.944
(0.019)
0.958
(0.015)
0.975
(0.007)
0.935
(0.019)
0.952
(0.014)
0.964
(0.013)
0.910
(0.032)
0.932
(0.025)
13
0.970
(0.007)
0.924
(0.019)
0.943
(0.014)
0.962
(0.010)
0.904
(0.025)
0.928
(0.019)
0.951
(0.018)
0.878
(0.042)
0.908
(0.032)
14
0.961
(0.014)
0.901
(0.035)
0.926
(0.027)
0.954
(0.022)
0.884
(0.053)
0.913
(0.040)
0.930
(0.021)
0.827
(0.047)
0.869
(0.037)
15
0.974
(0.014)
0.933
(0.034)
0.950
(0.026)
0.970
(0.013)
0.924
(0.032)
0.943
(0.025)
0.921
(0.032)
0.810
(0.065)
0.855
(0.054)
16
0.970
(0.015)
0.923
(0.037)
0.942
(0.029)
0.974
(0.007)
0.933
(0.018)
0.950
(0.014)
0.939
(0.016)
0.848
(0.037)
0.885
(0.029)
17
0.971
(0.012)
0.925
(0.031)
0.944
(0.023)
0.967
(0.016)
0.916
(0.038)
0.937
(0.030)
0.963
(0.014)
0.906
(0.034)
0.929
(0.026)
18
0.973
(0.014)
0.931
(0.033)
0.948
(0.027)
0.971
(0.011)
0.924
(0.027)
0.944
(0.020)
0.966
(0.016)
0.915
(0.038)
0.936
(0.029)
19
0.979
(0.008)
0.947
(0.020)
0.960
(0.015)
0.970
(0.014)
0.924
(0.035)
0.943
(0.027)
0.968
(0.013)
0.917
(0.031)
0.938
(0.024)
Chap e 5
92
de i a i es o LM decoding e ealed ACCs o 0.976. The use o ankle angles and hei de i a i es
p o ided he wo s pe o mances (0.915 < ACC < 0.954). This esul may be explained by he ac
ha he lexion/ex ension o he knee and hip, and consequen ly he ange o mo ion o hese join s,
p esen a highe a ia ion be ween he di e en LMs, whe eas he angle o he ankle is mo e iden ical
ac oss LMs.
Wi h espec o he da ase composed o he magni ude o 3D da a de i ed om
accele ome e s and gy oscopes, he use o pel ic, high, and shank IMU senso s e ealed
he bes esul s (0.970 < ACC < 0.979). Despi e his ype o da a being no used in he e iewed
s udies, he IMU combina ions ha p o ided he bes esul s (ACC anging om 0.932 o 0.984) we e
hose posi ioned in he same segmen s (pel is, highs, and shanks) used in s udies [41], [43]–[46], [116],
[117]. Mo eo e , he esul s indica ed he wo s LM decoding pe o mance when using he
magni ude o he 3D accele ome e and gy oscope da a om he pel ic IMU only (0.482 <
ACC < 0.900). This can be explained by he ac ha he pel ic segmen is he mos s able segmen
du ing locomo ion. Thus, i may no p esen signi ican changes in such a way ha i is easy o dis inguish
which LM he use is in.
Conside ing segmen angle da a ype, he use o pel ic segmen angles and hei
de i a i es as inpu showed he lowes pe o mances (0.732 < ACC < 0.794), as e i ied when
using he magni ude o 3D accele ome e and gy oscope da a. On he o he hand, he highes
pe o mance was e i ied when using he high and shank segmen angles and hei
de i a i es as inpu (0.979 < ACC < 0.982). These esul s a e aligned wi h hose achie ed in s udies
[43], [44], since he use o high and shank segmen angles and de i a i es e ealed ACCs anging om
0.930 and 0.983, espec i ely.
O e all, he e seems o be a endency o achie e high LM decoding pe o mances when using
kinema ic in o ma ion ex ac ed om IMU senso s placed a he shank and high segmen s (i.e., hip and
knee join angles, high and shank segmen angles, o magni ude o accele ome e and gy oscope da a
a hese segmen s). This endency could be a ibu ed o he high a ia ion in knee and hip lexion and
ex ension, as well as he ange o mo ion o hese join s, ac oss di e en LMs.
5.5.2 OFFLINE PERFORMANCE EVALUATION
F om he benchma k analysis be ween ypes o da a and DL algo i hms, he bes LOSOCV
esul s we e achie ed o an LSTM ed by high and shank segmen angles and de i a i es

Chap e 5
93
(a e age ACC o 0.982 ± 0.009). F om he li e a u e analysis, he LSTM model was no employed by
s udies [41], [43]–[46], [62], [114]–[117] in decoding LMs when using wea able assis i e de ices. A
di ec compa ison canno be made be ween he esul s o his s udy and he li e a u e, as he p o ocols
and he classi ied LMs di e among hem. None heless, he ob ained pe o mances a e close o he ones
ob ained in he s udy [45] (ACC = 0.984). The s udy [45] used a Mul ilaye Feed o wa d Neu al Ne wo k
o decode ST, LGW, SA, SD, RA, and RD asks. Despi e decoding RA and RD asks, he s udy [45] equi ed
a aining s ep o 18 minu es o each use . On he o he hand, he achie ed a e age ACC was p omising
since i was close o he highes ACC ound in he li e a u e (99.7%) [62].
In addi ion, bo h he LOSOCV and he model o line es pe o mances indica e ha he p oposed
LM decoding ool demons a es s ong gene aliza ion capabili ies ac oss pa icipan s.
5.5.3 ONLINE PERFORMANCE EVALUATION
O e all, he esul s ob ained du ing eal- ime es condi ions o bo h heal h and s oke
subjec s a e simila o he ones achie ed du ing he LOSOCV and o line es p ocedu es.
Despi e being sligh ly in e io , he esul s ob ained o he s oke pa ien (ACC = 0.972) a e compa able
o hose achie ed o he able-bodied pa icipan s (ACC = 0.986). This sligh educ ion in he e alua ion
me ics was also e i ied in he s udy [43] and i may be ela ed o he asymme ical and abno mal gai
pa e ns ypically pe o med by s oke pa ien s [7]. The ob ained model showed po en ial o decoding
LMs wi h a s oke pa ien , al hough i was ained on da a om able-bodied pa icipan s.
On he o he hand, none o he exis ing ools [41], [43]–[46], [116], [117] e alua ed he model
pe o mance a he p e e ed speeds o neu ologically impai ed use s (below 2.7 km/h). A his le el, he
p oposed LM decoding ool was demons a ed o be obus o di e en cadences,
co esponding o gai speeds anging om 1.0 o 1.6 km/h (0.988 < ACC < 0.991).
Addi ionally, in an a emp o ad ance s udies [41], [43]–[46], [116], [117], he obus ness o he
model was analyzed in ela ion o di e en scena ios. The DL decoding ool was es ed in di e en
en i onmen s (indoo and ou doo ), wi h a di e en numbe o s eps (8 and 13), wi h di e en s ep
dimensions (heigh : 16.0 - 18.0 cm; wid h: 110 – 500 cm; and dep h: 30.0 - 31.0 cm), and wi h a
di e en o de o ask execu ion. The p oposed LM decoding ool p o ed o pe o m consis en ly
in scena ios o he han hose adop ed du ing ool aining, e ealing ACCs anging om 0.978
o 0.991.
Chap e 5
94
Fu he mo e, he SD and SA asks p o ided he highes success a e o bo h able-bodied
and s oke subjec s, wi h alues anging om 95.7% o 99.1%. These esul s a e in line wi h hose poin ed
ou by s udies [44]–[46], [117]. The S ask p esen ed he lowes success a e (88.9% - 93.8%).
This esul may be explained by he ac ha in he ansi ion be ween S and LGW, a small in en ion o
s a walking (e.g., by bending he knee) is immedia ely de ec ed by he LM decoding ool as he LGW
ask, al hough he use s ill has bo h ee on he g ound ( he g ound u h is s ill S ask). This phenomenon
may educe he success a e in he S ask.
The p oposed LM decoding ool p esen ed a low compu a ional load (1.58 ± 0.42 ms) o
pe o m classi ica ion upon ecei ing new senso da a. This compu a ional load is wi hin he iming
equi emen s o he high-le el o he Sma Os sys em (10 ms, ope a ing a 100 Hz). Conside ing ha he
na u al equency o human mo ion du ing walking asks is less han 2 Hz (i.e., 500 ms), he p oposed
LM decoding ool wo ks as enough o decode human LMs [132].
I is no ewo hy ha he asse i eness o LM decoding ools is c i ical since misclassi ica ions can
cause he wea able assis i e de ice o adop inapp op ia e gai pa e ns, po en ially causing discom o
o he use s [14]. Howe e , he adop ion o inapp op ia e gai pa e ns can esul no only om
misclassi ica ions bu also om delays in decoding he new LMs [119]. Despi e he high accu acies
epo ed, mos o he e iewed s udies ecognize he new LM a e he use has al eady en e ed on i
[41], [45], [46], [62], [114]–[117]. The decoding delays ypically ange om 50.0 o 1897.9 ms.
Only wo s udies ([43], [44]) demons a ed he abili y o p edic speci ic LMs. The s udy [44]
success ully p edic ed he ansi ion om (i) SA o LGW in 10.7 ± 9.88 ms; (ii) SD o LGW in 78.7 ± 130.0
ms; and (iii) LGW o SA in 185.3 ± 56.9 ms. Howe e , he de ec ion o he LGW-SD ansi ion occu ed
40.0 ± 107.5 ms a e he c i ical ins an . In he s udy [43], he ollowing ansi ions we e p edic ed in
ad ance: (i) SD-LGW a 0.70 ± 58.7 ms, (ii) LGW-SA a 5.40 ± 74.6 ms, (iii) SA-LGW a 46.7 ± 46.6 ms,
and (i ) LGW-SD a 78.5 ± 25.0 ms. Ne e heless, he s udy [43] equi ed o line aining o c ea e a use -
speci ic decoding model o each pa icipan p io o eal- ime expe imen s. In ligh o hese indings, he
p oposed LM decoding ool ou pe o ms he a o emen ioned s udies ([41], [43]–[46], [62],
[114]–[117]) by p edic ing upcoming LMs wi h an a e age
p edic ion ime
o 482 ± 227 ms.
Fu he , he a e age
p edic ion ime
pe ansi ion was always posi i e ( anging om 241 o 860 ms),
which means ha , ac oss he six possible ansi ions (S -LGW, LGW-SA, SA-LGW, LGW-SD, SD-LGW,
and LGW-S ), he upcoming LM was decoded in ad ance, on a e age. T ansla ing hese esul s
as a pe cen age o he gai cycle ( om 72.3% o 93.1%), i can be assumed ha he nex LM was decoded,
on a e age, be ween he mid- and e minal swing o he gai cycle p eceding he upcoming LM [133].
Chap e 5
95
In addi ion, he S -LGW ansi ion was decoded wi h a highe
p edic ion ime
(860 ± 178 ms). This
esul could be explained by he possibili y ha du ing he S ask, sub le mo emen s could indica e he
imminen onse o he LGW ask. Con e sely, he LGW-S ansi ion p esen ed he lowes
p edic ion ime
(241 ± 294 ms). The high s anda d de ia ion alue sugges s ha , a his ansi ion, he pa icipan s may
be no imely assis ed by he Sma Os sys em in some cases, i.e., he assis ance could change a e he
c i ical ins an . This can be ela ed o he ac ha he LGW-S ansi ion may in ol e a mo e ab up shi
by he use , making i di icul o p edic he S ask.
5.6 CONCLUSIONS
This wo k o e s wo con ibu ions o he s a e o he a : (i) o in es iga e he bes usion o
senso da a and classi ica ion algo i hms o build a DL decoding ool capable o p edic ing in
ad ance ou LMs (S , LGW, SD, and SA) and ansi ions be ween hem; and (ii) o de elop an LM
decoding ool he imely decode ou LMs in ad ance.
This esea ch concludes ha , o decoding LMs, he use o (i) RMS EMG ea u es is no
su icien ; (ii) high and shank segmen angles as inpu o an LSTM p o ide he bes
pe o mances. Fu he mo e, he p oposed ool was ound o be accu a e in bo h o line and online
es ing wi h heal hy and s oke pa icipan s, p esen ing a low compu a ional load. In addi ion, he ool
p o ed o be obus o di e en walking cadences and di e en scena ios.
These conclusions highligh he sui abili y o he de eloped LM decoding ool o os e he
de elopmen o LM-d i en ajec o y con ol s a egies (de ailed in Chap e 6) ha imely adap he
assis ance o wea able assis i e de ices acco ding o he LM decoded.
Al hough he p oposed LM decoding ool showed p omising esul s, he e a e s ill oppo uni ies o
imp o emen . Speci ically, he e is a need o (i) conduc addi ional eal- ime alida ion p ocedu es wi h
mo e pa icipan s, including bo h able-bodied indi iduals and hose wi h a ying deg ees o neu ological
impai men s; (ii) ex end he decoding capabili ies o o he LMs, pa icula ly RA and RD, and ansi ions;
and, (iii) imp o e he p edic ion pe o mance o LGW-S ansi ions o inc ease he
p edic ion ime
.
96
6. ASSIST-AS-NEEDED
CONTROL STRATEGIES
Chap e 6
97
This chap e begins wi h an in oduc ion ega ding he ends o assis i e con ol s a egies applied
o obo -based gai aining. I hen desc ibes he implemen ed hie a chical con ol a chi ec u e, ocusing
on he de elopmen and alida ion o he assis i e con ol s a egies in es iga ed in his hesis: AAN LM-
d i en ajec o y con ol, AAN EMG-based con ol, and AAN HITL con ol. These s a egies we e
designed, implemen ed, and alida ed wi h heal hy pa icipan s using he Sma Os sys em. The po en ial
ehabili a ion bene i s o each assis ance s a egy a e e iewed and discussed. The chap e concludes
wi h a comp ehensi e o e iew o he p oposed assis i e con ol s a egies.
6.1 INTRODUCTORY INSIGHT
Bio-inspi ed con ol a chi ec u es ha e begun o eme ge, inspi ed by he p inciples and
o ganiza ion o he human mo ion con ol sys em [9]. These a chi ec u es aim o e icien ly implemen
use -o ien ed con ol s a egies and acili a e syne gis ic and con inuous in e ac ion be ween he use and
he wea able assis i e de ice, he eby enhancing b ain plas ici y [134]. Fo ha , hese a chi ec u es may
include low-, mid-, and high-le el con ols ha a e hie a chically o ganized o add ess he physical
in e ac ions be ween he use , he en i onmen , and he de ice. The design and de elopmen o he
hie a chical con ol a chi ec u e o he Sma Os sys em ollowed hese esea ch di ec ions and exploi ed
he po en ial o di e en assis i e con ol s a egies, as desc ibed in his chap e .
Acco ding o s udies [37], [134], ad anced echnology should be used a ionally. I seems ha
pos -s oke pa ien s may bene i mo e om aining wi h wea able assis i e de ices when
he ac i e pa icipa ion o he pa ien is in oked. These esul s may be due o he ac ha
spon aneous neu oplas ici y phenomena occu wi h g ea e in ensi y in he ini ial acu e and sub-acu e
phases o ehabili a ion a e s oke. In his sense, pa ien s' ac i e pa icipa ion should be
encou aged o achie e apid neu omuscula eco e y du ing ehabili a ion aining.
Recen ends sugges ha use -o ien ed assis ance ha encou ages use pa icipa ion may be
achie ed h ough he use o AAN con ol s a egies in eg a ed in o wea able assis i e de ices
[29]–[31], [33], [37]. AAN con ol s a egies a e ocused on dynamically adjus ing he le el o assis ance
based on he use 's eal- ime pe o mance and locomo ion needs, ensu ing ha he wea able assis i e
de ice assis s mo emen only when and as much as equi ed. The goal is o encou age he use 's
ac i e pa icipa ion and muscle engagemen while co ec ing he join pa e n du ing he
execu ion o he walking mo ion [29]–[31], [33], [37].

Chap e 6
98
This Ph.D. hesis p og esses he cu en li e a u e by p oposing he de elopmen o h ee AAN
con ol s a egies, as de ailed below:
1. Locomo ion Mode-d i en T ajec o y Con ol: a con ol esponsible o mapping he
use ’s locomo ion in en ion and gene a ing e e ence ajec o ies acco ding o he
desi ed/in ended LM;
2. Assis -As-Needed Elec omyog aphy-based Con ol: a con ol combining posi ion
and o que con olle s o adjus he o hosis assis ance acco ding o he use 's muscula
needs and join pa e n;
3. Human-in- he-Loop Con ol: a con ol designed o op imize he use me abolic cos by
modi ying con ol a iables, namely, o que peak magni udes.
6.2 LOCOMOTION MODE-DRIVEN TRAJECTORY CONTROL
6.2.1 CRITICAL ANALYSIS OF RELATED WORK
O en, wea able assis i e de ices ha e been equipped wi h non-in usi e senso con igu a ions and
sophis ica ed algo i hms o decode LMs, wi h he aim o ailo ing he obo 's assis ance o help use s
pe o m hei daily mobili y asks.
F om he li e a u e, only i e s udies ha e adap ed he assis ance o he wea able assis i e de ice
a ending o he use ’s decoded LM [40], [42], [43], [116], [117]. In he s udies [40], [42], a wo-le el
hie a chical con ol design was ollowed. In he s udy [40], a cons an o que o 10 N.m was p o ided by
an ankle o hosis (low-le el) when he SD ask was ecognized (high-le el). Fo ha , FSR senso s
embedded in he o hosis oo we e used o de ec he gai cycle phase. Then, he LM was decoded, and
he cons an o que was p o ided. The o que was deli e ed du ing (i) he i s 50% o he gai cycle o
suppo he ankle join mo ion when he leg wi hou assis ance was in he swing phase; (ii) he las 20%
o he gai cycle o gua an ee ha he oo was pe o ming plan a lexion be o e ouching he nex s ai
s ep. In he s udy [42], a wo-le el hie a chical con ol was designed o p o ide a phase-locked assis i e
o que (low-le el) when an LM (among he Si , S , LGW, SA, and SD asks and ansi ions be ween hem)
was ecognized (high-le el). A e ecognizing he LW, SA, and SD asks, a gai phase es ima ion algo i hm
was employed such ha he mos sui able phase-locked assis i e o que was p o ided. The p o ided
o que was adop ed om public da abases wi h join ajec o ies eco ded du ing locomo ion- ela ed
ac i i ies [133], [135]–[137]. Du ing he Si and S asks, no o que was p o ided. None heless, o que
Chap e 6
99
pa e ns modeled acco ding o he body mass o each pa icipan we e deli e ed du ing Si -S o S -Si
ansi ions. These ansi ions we e de ec ed based on he hip join angles, using a h eshold algo i hm.
On he o he hand, he s udies [43], [116], [117] ollowed a h ee-le el hie a chical con ol
a chi ec u e. In he s udy [117], he high-le el p esen ed a gai phase es ima ion algo i hm using IMU
signals o segmen he inpu da a in o s ides. Based on hese s ides, he LM (LGW, SA, and SD) was
decoded h ough a CNN. A Pa ame e Op imal I e a i e Lea ning Con ol me hod was implemen ed in he
mid-le el. This me hod was designed o a so lowe limb exoskele on o p o ide hip and knee assis ance
acco ding o he ension o Bowden cables. The ension o hese cables was modeled acco ding o he
hip and knee join momen s when pe o ming he LW, SA, and SD asks. The low-le el was esponsible
o d i ing he exoskele on acco ding o he ension o he Bowden cables.
The high-le el de eloped in he s udy [43] was di e en om he one de eloped in he s udy [117].
Fi s ly, he LM was dis inguished be ween s a ic and dynamic asks a he high-le el. Then, in he case o
a dynamic ask, he LM was classi ied (among LW, SA, and SD asks), ollowed by a gai phase es ima ion
algo i hm o dis inguish he s ance om he swing phase. Du ing he s ance phase, h ee o he subphases
(heel-s ike, heel-o , and oe-o ) we e iden i ied a he high-le el. In addi ion, he mid-le el de eloped in
he s udy [43] was hyb id. I used a ze o- o que mode du ing he S ask and he swing phase o LW, SA,
and SD asks in an a emp o educe he cons ain s associa ed wi h he ac ua o ’s ic ions. Mo eo e ,
du ing he i s and second subphases ( om he heel s ike o heel-o e en s and om heel-o o oe-o
e en s, espec i ely) o he SA ask, he knee should pe o m he ex ension and lexion mo emen s,
espec i ely. Based on hese assump ions, he knee exoskele on p o ided a closed-loop o que con ol o
ex end and lex he knee join , assis ing he use du ing he i s and he second subphases o he SA
ask. Fo he s ance phase o he LW and SD asks, he muscles a ound he knee join p o ide nega i e
o que o help he knee o mo e and a oid excessi e knee lexion, suppo ing he human body. Fo his
eason, an open-loop damping con ol was adop ed o p o ide esis ance a he knee join o suppo he
body and abso b he shock. The low-le el was esponsible o gene a ing assis i e commands as a
esponse o he compa ison be ween he desi ed o que (gene a ed by he mid-le el) and he measu ed
o que (measu ed by an embedded load cell).
In he s udy [116], ano he h ee-le el hie a chical con ol was adop ed o assis he S , LGW, SA,
and SD asks, and ansi ions be ween hem. The high-le el was esponsible o decoding LMs based on
a h eshold algo i hm ( o dis inguish be ween s a ic and dynamic asks) and a Fuzzy-logic-based algo i hm
( o dis inguish be ween LGW, SA, and SD asks). The mid-le el con olle employs adap i e oscilla o s o
Chap e 6
100
ack he desi ed o que cu e, while he low-le el con olle d i es he mo o s o main ain he o ce
eedback loop.
Addi ionally, p o iding e icien assis ance acco ding o he decoded use ’s LM implies
an accu a e and imely iden i ica ion o he use ’s in en ions. This is o u mos impo ance since
he highe he an icipa ion ime in iden i ying he LM, he mo e ime emains o swi ch he con ol o assis
he use s acco ding o hei needs imely. This means ha , ideally, he upcoming LM should be p edic ed
be o e i s occu ence, in o de o adap he assis ance o he wea able assis i e de ice acco ding o he
LM decoded. Conside ing he i e s udies ha adap ed he assis ance acco ding o he decoded LM, only
he s udy [43] demons a ed he abili y o p edic and assis speci ic LMs. The s udy [43] success ully
p edic ed he ansi ion om (i) SD-LGW a 0.70 ± 58.7 ms, (ii) LGW-SA a 5.40 ± 74.6 ms, (iii) SA-LGW
a 46.7 ± 46.6 ms, and (i ) LGW-SD a 78.5 ± 25.0 ms. Ne e heless, he s udy [43] equi ed o line
aining o c ea e a use -speci ic decoding model o each pa icipan p io o eal- ime expe imen s, and
he assis ance was only p o ided du ing he s ance phase.
Despi e ecen ad ances, i is s ill unclea how o con ol he de ice o p ope ly and
imely assis use s acco ding o hei locomo ion in en ions. The a ailable s udies a e limi ed by
(i) only ac i ely assis ing du ing speci ic phases o he gai cycle (s ance/swing o be ween gai cycle
e en s) [40] [43] o LMs [42], [117]; (ii) he adap a ion he con ol acco ding o he decoded LM occu s
a e he c i ical ins an (i.e., a e he use en e s on he new LM) [40], [42], [43], [116], [117]; (iii)
es ing he con ol pe o mance in a single scena io, which is commonly used o ain he LM decoding
ool; and (i ) only add ess he walking speeds ypically adop ed by able-bodied use s. This esea ch
ad ances he cu en s a e-o - he-a by p oposing a h ee-le el assis i e con ol s a egy o imely
adap he assis ance o he Sma Os sys em acco ding o he decoded LM (among S , LGW,
SA, and SD), while ac i ely assis ing h oughou he en i e gai cycle. Fo ha , he p oposed con ol
s a egy in eg a es he DL ool de eloped in Chap e 5 in o he high-le el o he p oposed con ol s a egy.
The obus ness o he p oposed LM-d i en ajec o y con ol was assessed (i) in di e en scena ios
(indoo and ou doo ); and (ii) in a ange o speeds, all wi hin he p e e ed alues o s oke pa ien s.
6.2.2 METHODOLOGY
A. Hie a chical Con ol A chi ec u e
Once he op imal usion be ween senso da a and DL algo i hms has been ound o decoding
LMs (Chap e 5), a hie a chical h ee-le el con ol s a egy was de eloped o adap he assis ance o
Chap e 6
101
he Sma Os sys em acco ding o he ou pu o he LM decoding ool (i.e., a classi ica ion be ween
S , LGW, SA, o SD). The p oposed con ol s a egy is p esen ed in Figu e 6.1.
Figu e 6.1 – P oposed hie a chical con ol a chi ec u e.
LM
is he p edic ed LM. ϴ
e .h
and ϴ
e .o
a e he
human and o hosis ankle join e e ence angles, espec i ely.
e
ϴ
is he o hosis posi ion e o .
u
is he
P opo ional-In eg al-De i a i e (PID) command.
SS
is a speed scaling block achie ed by (2).
PID
s ands
o he PID con olle .
The high-le el uns he LM decoding ool. As depic ed in Figu e 6.1, new inpu da a om he
Ine ialLab sys em a i es a he LM decoding ool (
LM Decoding Tool
block) a 100 Hz. The e o e,
he DL model p o ides a new classi ica ion e e y 10 ms. This classi ica ion is used o upda e a 3-
sample bu e ha con ains he wo p e ious classi ica ions. The ou pu o he 3-sample bu e was
used oge he wi h he gai cycle phase o de e mine he human e e ence ajec o y o he ankle
join angle (ϴ
e .h
) h ough he T ajec o y Se ing block. This block is esponsible o upda ing he ϴ
e .h
acco ding o he decoded LM. The se ϴ
e .h
will assis he use in pe o ming he decoded LM, i.e., a
ajec o y ha suppo s he use du ing he S , LGW, SA, o SD asks. The ajec o ies (depic ed in
Figu e 6.1) we e ob ained om he da ase collec ed o de eloping he LM decoding ool (Chap e
5).
Nex , he
Speed Scaling
block ha ope a es a 100 Hz, ecei es he ϴ
e .h
gene a ed by he
T ajec o y Se ing
block and he gai speed. This occu s a he mid-le el o he a chi ec u e. The ϴ
e .h
LM
Decoding
Tool
Senso Da a
Type
T ajec o y
Se ing SS
O hosis
uman
PID
u
meas o
e
e o
Mid le el ( )Low le el ( k )
igh le el ( )
LM
e h
ai speed
LM
LM
is
LM
is L
LM
is
LM
is
e h
e h
e h
e h
ajec o y e ing
Gai Cycle ( )
Ankle Angle (deg)
Gai Cycle ( )
Ankle Angle (deg)
Gai Cycle ( )
Ankle Angle (deg)
Gai Cycle ( )
Ankle Angle (deg)
Chap e 6
108
(ii) depend on Hill- ype models o es ima e he use ’s join o que, which may esul in complex, use -
dependen calib a ion, and ime-consuming me hods [28], [35], [36], [38], [145]; (iii) a e no adap able
since hey ely on ixed assis ance a ios o de ine he desi ed join o que ajec o ies [28], [38], [145];
and (i ) we e no de eloped o assis he walking mo ion, bu a he ocus on isola ed lexion and ex ension
mo emen s o he lowe limb join s [28], [35]–[38].
This Ph.D. hesis ackles he abo e-men ioned limi a ions by p oposing an AAN EMG-based con ol
s a egy o au oma ically assis he walking mo ion h oughou he en i e gai cycle. The sys em
implemen s an ou e loop o que con ol and an inne loop posi ion con ol, which adap s he
le el o assis ance o be p o ided o he use in eal- ime and au oma ically. To do his, he
sys em p esen s wo DL eg esso s: (i) one eg esso o de ine use -o ien ed o que e e ence ajec o ies
based on he use 's body heigh , body mass, and gai speed [146], wi hou he need o using ixed
assis ance a ios; and (ii) ano he eg esso o de e mine he use 's engagemen (es ima ing he use ’s
join o que, in eal- ime, h ough a DL eg esso ed by a usion o kinema ic, EMG, an h opome ic, and
demog aphic da a [65]). The con ol s a egy was designed, implemen ed, and alida ed in o he Sma Os
sys em o assis he ankle join while conside ing he use 's engagemen .
6.3.2 METHODOLOGY
A. Hie a chical Con ol A chi ec u e
The AAN EMG-based con ol p oposed in his esea ch ollows a hie a chical a chi ec u e
o ganized in o high-, mid-, and low-le els, as depic ed in Figu e 6.4. I combines a o que ou e loop
and a posi ion inne loop con ol. In his app oach, he ou e loop in oduces a "so ening"
e ec in he human- obo in e ac ion, while he inne loop ensu es join s i ness. This
helps supp ess undesi able dis u bances, elimina ing he necessi y o compensa ion based on
an icipa o y con ol models [27].
The high-le el con olle , implemen ed a 100 Hz, is esponsible o (i) de ining he human
ankle join e e ence angle (ϴ
e .h
) and o que ajec o ies (Ԏ
e .h
); (ii) es ima ing he human ankle join
o que ajec o y (Ԏ
es .h
), in eal- ime, h ough he
Human-To que Es ima ion
(HTE) block; and (iii)
de e mine he a ia ion o he Ԏ
e .h
and Ԏ
es .h
(
∆
Ԏ
e .h
and
∆
Ԏ
es .h
, espec i ely). The ϴ
e .h
is gene a ed by
a eg ession model p oposed in [147], conside ing he use ’s body heigh and gai speed. The
me hodology o de ining Ԏ
e .h
is de ined in sub-sec ion
B. Es ima ion o Ankle Join Re e ence To que
T ajec o ies
. In he p oposed con ol s a egy, he
Ԏ
e .h
is de ined acco ding o he ϴ
e .h
and i s

Chap e 6
109
de i a i es (angula eloci y and angula accele a ion), gai speed, an h opome ic (body heigh and
mass, shank, and oo leng hs), and demog aphic (gende and age) da a. The HTE block is explained
in de ail in sub-sec ion
C. Es ima ion o Real- ime Ankle Join To que T ajec o ies
. The HTE block
es ima es he Ԏ
es .h
in less han 2 ms, conside ing he EMG signals om he
ibialis an e io
and
gas ocnemius la e alis
muscles (no malized by he co esponding MVCs), hip join kinema ics,
walking speed, an h opome ic (body heigh and mass, shank, and oo leng hs), and demog aphic
(gende , age) da a.
Figu e 6.4 – AAN EMG-based o que con ol, combining a o que ou e loop wi h a posi ion inne loop.
Ԏ
e .h
and Ԏ
es .h
a e he human ankle join e e ence and es ima ed o ques, espec i ely.
∆
Ԏ
e .h
and
∆
Ԏ
es .h
a e he a ia ions o he Ԏ
e .h
and Ԏ
es .h
, espec i ely.
∆
Ԏ
e .o
and
∆
Ԏ
es .o
a e he
∆
Ԏ
e .h
and
∆
Ԏ
es .h
in e pola ed
o he low-le el equency, espec i ely.
e
Ԏ
.o
is he o hosis o que e o be ween
∆
Ԏ
e .h
and
∆
Ԏ
es .h
. ϴ
e .h
and
ϴ
e .o
a e he human and o hosis ankle join e e ence angles, espec i ely.
e
ϴԎ
.o
is he o hosis o que-
de i ed posi ion e o . ϴ
meas.o
is he o hosis ankle join measu ed angle.
e
ϴ
.o
is he o hosis posi ion e o .
u
is he PID command.
EMGTA
and
EMGGAL
a e he EMG signals measu ed om he
ibialis an e io
and
gas ocnemius la e alis
muscles, espec i ely. ϴ
meas.h
is he human hip join measu ed angle.
GS
is he gai
speed.
BH
and
BM
a e he human body heigh and mass, espec i ely.
SL
and
FL
a e he shank and oo
leng hs, espec i ely.
K
is a ixed ac o equal o 1.33 N.m/ adians.
SS
is a speed scaling block achie ed
by Equa ion 6.1.
PID
is a P opo ional-In eg al-De i a i e (PID) con olle .
HTE
is a
Human-To que
Es ima ion
(HTE) block.
In he mid-le el con olle , which ope a es a 100 Hz, he human ankle join e e ence angles
(ϴ
e .h
)
, ∆
Ԏ
e .h
, and
∆
Ԏ
es .h
a e in e pola ed (achie ing he o hosis ankle join e e ence angles (ϴ
e .o
)
,
∆
Ԏ
e .o
, and
∆
Ԏ
es .o
, espec i ely) o he low-le el con ol equency h ough he
Speed Scaling
(SS)
block (see Equa ion 6.1). The
∆
Ԏ
e .o
and
∆
Ԏ
es .o
a e compa ed, gene a ing he o hosis o que e o
(
e
Ԏ
.o
). The
e
Ԏ
.o
is con e ed in o an o hosis o que-de i ed posi ion e o (
e
ϴԎ
.o
) h ough a ixed scaling
es h
e
o
K
Exo
uman
e h e
o
PID
M M L
meas h
u
e
o
TE
meas o
Mid le el ( )
igh le el ( )
e o
M L L ende ge
es h
e h
e h
SS
SS
es o
e o
Chap e 6
110
alue
K
se a 1.33 N.m/ adians. This alue was empi ically ound o he used AO and i ep esen s
he o a ional s i ness.
A he low-le el con olle (wo king a 1 kHz), i was empi ically e i ied ha was equi ed o
apply a eed o wa d posi ion con ol o p e en he ankle o hosis om mo ing o posi ions ha could
cause inju y o he use and could damage i sel . This eed o wa d posi ion con ol was also p oposed
in a s udy [27] o imp o e ajec o y acking. To his end, a eed o wa d posi ion con ol was applied
o he inne loop, adding he ϴ
e .o
ha he o hosis should mo e o while aking in o accoun he Ԏ
e .h
o he ou e loop. The
e
ϴԎ
.o
, ϴ
e .o
, and he o hosis ankle join measu ed angle (ϴ
meas.o
) a e compa ed,
gene a ing he o hosis posi ion e o (
e
ϴ.o), ed o he PID con olle . The PID con olle was
implemen ed using p opo ional, in eg al, and de i a i e gains o 95, 1.5, and 1.5, espec i ely.
Conside ing he
e
ϴ, he PID con olle compu es a PID command (
u
) ha is limi ed o maximum and
minimum alues o 2500 and -2500, h ough a sa u a o . This command is in e p e ed by he ankle
o hosis, gene a ing he co esponding mo o o que.
Wi h his con ol a chi ec u e, i is in ended ha i he use pe o ms (i) an ankle join angle
and o que ajec o ies close o he e e ences, he assis ance o he ankle o hosis should
be minimal; (ii) an ankle join angle and/o o que ajec o ies di e en om he e e ences,
he ankle o hosis should p o ide assis ance in o de o compensa e o hese
di e ences.
B. Es ima ion o Ankle Join Re e ence To que T ajec o ies
Since he p oposed AAN EMG-based con ol s a egy mus ollow a use -cen e ed design, i
was equi ed ha he AO p o ides assis ance o ien ed o he use 's needs. This may be achie ed by
de ining a con ol e e ence a iable, i.e., ankle join o que, adap ed o each use . In his connec ion,
eg esso -based DL models we e ained o es ima e ankle join e e ence o que ajec o ies (Ԏ
e .h
),
acco ding o he use ’s body heigh , body mass, and gai speed. In his Ph.D. hesis, he ocus is
gi en o CNN. Fu he de ails o es ima ing use -o ien ed o que e e ence ajec o ies wi h he LSTM
model can be ound in he s udy [146].
To ain he DL models, a da a collec ion was pe o med. The s udy in ol ed hi een able-
bodied adul pa icipan s (6 males and 7 emales) wi h an a e age age o 24.2 ± 1.85 yea s, an
a e age body mass o 65.2 ± 10.3 kg, and an a e age body heigh o 168 ± 12.0 cm. Balanced
gende dis ibu ion was ackled conside ing possible biomechanical gende di e ences [148]. Be o e
Chap e 6
111
unde going he expe imen s, each pa icipan ga e w i en and in o med consen acco ding o he
e hical conduc o he Uni e si y o Minho Commi ee (CEICVS 006/2020).
As depic ed in Figu e 6.5, pa icipan s we e ins umen ed wi h a o al o 12 pai s o e o-
e lec i e ma ke s, which combined wi h a wel e-came a mo ion-cap u e sys em (Oqus; Qualisys—
Mo ion Cap u e Sys em, Gö ebo g, Sweden) and i e o ce pla o ms embedded on he loo
(FP4060; Be ec, Ohio, OH, USA) we e u ilized o collec he lowe limb join angles and o ques (a
he ankle, knee, and hip join s). Mo eo e , i was s udied how he use o EMG signals a ec s he
ankle join o que es ima ion. Fo ha , an 8-channel su ace EMG sys em [T igno A an i (Delsys
Inco po a ed, Na ick, USA)] was employed o collec EMG da a om he
ibialis an e io
,
gas ocnemius la e alis
,
biceps emo is
, and
as us la e alis
muscles.
Each subjec was ins uc ed o pe o m en o wa d walking ials, walking sequen ially a
se en con olled walking speeds (1.0, 1.5, 2.0, 2.5, 3.0, 3.5, and 4.0 km/h). These speeds we e
con olled by a me onome and we e chosen o co e he ypical walking speeds o pos -s oke
pa ien s and able-bodied use s. The pa icipan s we e asked o look ahead and walk na u ally
acco ding o he bea s o he me onome. Fu he de ails o he p o ocol a e p esen ed in [65], [121].
Figu e 6.5 – EMG con igu a ion and ma ke -se adop ed: (A)
as us la e alis
; (B)
ibialis an e io
; (C)
biceps emo is
; (D)
gas ocnemius la e alis
; (1) An e io supe io iliac spine; (2) ochan e ; (3) high; (4)
la e al knee; (5) medial knee; (6) shank; (7) la e al ankle; (8) medial ankle; (9) oo me a a sal 5; (10)
oo me a a sal 1; (11) oo me a a sal 2; (12) pos e io supe io iliac spine.
Ou o he hi een pa icipan s, one was andomly selec ed o es he ained model. A
LOSOCV me hod among he emaining wel e pa icipan s was implemen ed o e alua e he model
Chap e 6
112
gene aliza ion and op imize he model hype pa ame e s. Fo he CNN, he pa ame e s explo ed we e
(i) he ke nel size (2 × 2, 5 × 5, and 10 × 10); (b) he numbe o il e s pe con olu ional laye (8,
16, 32, 64); and (c) he numbe o con olu ional laye s (1, 2, 3). Fu he , an empi ical analysis was
applied o selec (i) he no maliza ion me hod (conside ing he max-min, z-sco e, and obus
no maliza ion me hods); (ii) he op imal ba ch size; and (iii) he d opou pe cen age. In addi ion, bo h
neu al ne wo ks’ weigh s and biases we e upda ed acco ding o an adap i e momen es ima ion
op imiza ion algo i hm (ADAM) conside ing he mean squa e e o (MSE) ([36,51]).
The eg ession model pe o mance was e alua ed du ing he LOSOCV and model es ing
p ocedu es. Th ee me ics we e de e mined, namely,
,
R2
, and he compu a ional load (in
milliseconds). The p edic ions pe o med by he ained model we e also e alua ed using he Bland–
Al man Plo [151]. All p edic ions we e pe o med in a Hewle -Packa d compu e wi h an In el®
Co e™ i7-4710MQ CPU @ 2.50 GHz p ocesso and a Random Access Memo y wi h 16.0 GB.
C. Es ima ion o Real- ime Ankle Join To que T ajec o ies
Conside ing ha he Sma Os sys em is no able o measu e he human ankle join o que, i
was equi ed o explo e a me hodology o es ima e use -o ien ed ankle join o que ajec o ies (Ԏ
meas.h
)
in eal- ime, in a simila way as a o que senso . In his connec ion, his Ph.D. hesis p oposes he
HTE block (Figu e 6.4), which ep esen s an EMG-based o que es ima ion DL eg esso . Fo ha , a
CNN model was explo ed conside ing he ollowing da a usion: gai speed, kinema ic (hip
kinema ics), EMG (
ibialis an e io
and
gas ocnemius la e alis
muscles), an h opome ic (body
heigh and mass, anging om 1.50 o 1.90 m and 50.0 o 90.0 kg, espec i ely, and shank and
oo leng hs) and demog aphic da a (age, gende ). Fu he de ails can be ound in he s udy [65].
The CNN model was ained wi h a walking da ase collec ed om se en een heal hy
pa icipan s (9 emales and 8 males) wi h a mean body heigh o 168.0 ± 10.31 cm, a mean body
mass o 70.11 ± 14.26 kg, and a mean age o 28.05 ± 3.66 yea s.
Pa icipan s we e ins umen ed wi h (i) an 8-channel su ace EMG sys em [T igno A an i
(Delsys Inco po a ed, Na ick, USA)] o collec EMG da a om he
ibialis an e io
,
gas ocnemius
la e alis
,
biceps emo is
, and
as us la e alis
muscles; (ii) a o al o 12 pai s o e o- e lec i e
ma ke s, in which combined wi h a wel e-came a mo ion-cap u e sys em (Oqus; Qualisys—Mo ion
Cap u e Sys em, Gö ebo g, Sweden) and Fo ce Pla e-Ins umen ed T eadmill (Side-by-Side T eadmill
– AMTI, MA, USA) we e used o acqui e he lowe limb join angles and o ques (a he ankle, knee,
Chap e 6
113
and hip join s). A e , each pa icipan was ins uc ed o walk o 4 minu es on he ins umen ed
eadmill, pe o ming 2 minu es a 1.5 km/h and hen 2 minu es a 2 km/h.
An empi ical analysis was conduc ed o selec (i) he ke nel size (2, 10, 20, 40, 60); (ii) he
numbe o con olu ional laye s (1, 2, 3, 4); (iii) he numbe o il e s pe con olu ional laye (8, 16,
32, 64, 128); (i ) he sequence leng h (40, 50, 60, 80, 100, 120, 150); ( ) he ba ch size; and ( i)
he d opou a e. The no maliza ion me hod was also s udied among max-min, z-sco e, and obus
no maliza ion me hods. Fu he mo e, he ADAM algo i hm based on he MSE was used o upda e
he weigh s and biases o he CNN. This empi ical analysis was conduc ed h ough he LOSOCV
p ocedu e. O he se en een pa icipan s, one ( emale 27 yea s old and a body mass and heigh o
73.4 kg and 1.63 m, espec i ely) was andomly selec ed o es he model. Thus, he model was
ained wi h da a om 16 subjec s.
Th ee e alua ion me ics we e employed o e alua e he o line CNN’s pe o mance, namely,
he RMSE, he No malized Mean Squa e E o (NMSE), and he
. These me ics we e compu ed
be ween he p edic ed and he eal (g ound u h) ankle join o que ajec o ies du ing he LOSOCV
and he model es ing p ocedu es (bo h pe o med in a Hewle -Packa d compu e wi h an In el®
Co e™ i7-4710MQ CPU @ 2.50 GHz p ocesso and a Random Access Memo y wi h 16.0 GB).
Fu he , an expe imen al p o ocol was pe o med o e alua e he eal- ime pe o mance, i.e.,
he p edic ion ime o he CNN in o Sma Os CCU. This p o ocol in ol ed one able-bodied male
pa icipan (27 yea s old) wi h a body heigh o 1.70 m and a body mass o 81.2 kg. In he beginning,
he pa icipan 's gende , age, body mass, heigh , leg leng h, and oo leng h we e measu ed and
in oduced in he mobile g aphical applica ion o he Sma Os sys em. Then, he pa icipan
pe o med wo MVCs o each muscle o no malize he EMG da a. A e ha , he pa icipan was
ins uc ed o pe o m a 5-seconds s anding calib a ion ial o calib a ing he Ine ialLab sys em. A
las , he pa icipan walked on he ins umen ed eadmill o 5 minu es a 1.5 km/h, while he ankle
join o que ajec o ies we e es ima ed in eal- ime.
C. Expe imen al Valida ion o AAN EMG-based Con ol
To alida e he comple e AAN EMG-based con ol p oposed in Figu e 6.4, i e heal hy
pa icipan s (3 males and 2 emales wi h an a e age age, body heigh , and body mass o 27.8 ±
3.1, 170.4 ± 8.2, and 67.8 ± 11.4, espec i ely) wi hou e idence o mo o diso de s we e in ol ed.
All pa icipan s p o ided in o med consen acco ding o he Uni e si y o Minho E hics Commi ee

Chap e 6
114
(CEICVS 006/2020). I is impo an o no e ha hese pa icipan s we e no in ol ed in aining he
eg esso s desc ibed in sub-sec ions
B
and
C
.
The p o ocol s a ed by collec ing he pa icipan ’s gende , age, body heigh and mass, and
oo and shank leng hs. Then, pa icipan s we e equipped wi h (i) wo EMG senso s placed on he
igh
ibialis an e io
and he igh
gas ocnemius la e alis
muscles, ollowing he SENIAM
ecommenda ions [105]; and (ii) 2 IMUs om he Ine ialLab sys em posi ioned in he pel is and in
he igh high, o collec hip join angles. All hese da a a e used in he HTE block, as depic ed in
Figu e 6.4. Once ins umen ed wi h he senso sys ems, he pa icipan pe o med wo MVCs pe
muscle. A e ha , he pa icipan was ins umen ed wi h he Sma Os sys em (Figu e 6.6).
Figu e 6.6 – Male pa icipan equipped wi h he Ine ialLab, T igno A an i, and Sma Os sys ems.
Once ins umen ed, pa icipan s unde wen a pe iod o amilia iza ion wi h he de ice ollowing
a ajec o y- acking posi ion con ol. A e amilia iza ion, he da a collec ion s a ed wi h pa icipan s
walking in wo di e en sessions. One session consis ed o pa icipan s being assis ed by he AO
ollowing a ajec o y- acking posi ion con ol. The o he session consis ed o pa icipan s
being assis ed by he AO ollowing an AAN EMG-based con ol. Bo h sessions we e pe o med a
h ee speeds (1.2, 1.4, and 1.6 km/h). Howe e , he pa icipan s did no know which con ol s a egy
was con olling he AO in o de o e alua e he use s’ pe cep ion o which s a egy was mo e o less
adap ed o hei eal- ime mo ion. In his connec ion, a esea che andomly selec ed he assis i e
con ol s a egy. Subsequen ly, pa icipan s pe o med h ee asks. In he i s , hey we e asked o
emain in a s anding posi ion wi h hei igh leg aised ( he leg ins umen ed wi h he Sma Os
sys em), wi h no mo emen , o 20 seconds. Secondly, pa icipan s we e asked o walk o 1 minu e
on he eadmill, which was ini ially se a a speed o 1.2 km/h. This ask is e e ed o as he
uncondi ioned ask
, in which pa icipan s we e asked o ac i ely pe o m he walking mo ion. In he
IMU senso s (Ine ialLab)
EMG senso s (Delsys)
Ankle O hosis Backpack
Chap e 6
115
hi d es , he pa icipan s we e asked o lean on he sideba s o he eadmill, so ha hey we e
almos suspended (only hei ee ouched he eadmill ma , bu wi hou suppo ing hei body) o
1 minu e. A his poin , pa icipan s we e asked o simula e a pa hological condi ion by ying no o
ac i a e he
ibialis an e io
and
gas ocnemius la e alis
muscles o pe o m he gai . In o he wo ds,
pa icipan s had o le he AO do mos o he walking mo ion. Thus, his ask is e e ed o as he
condi ioned ask
, in which pa icipan s we e asked o passi ely pe o m he walking mo ion. Gi en
his condi ion, i would be expec ed ha he e would be a dec ease in he ac i a ion o he
a o emen ioned muscles and, consequen ly, he AO would inc ease i s assis ance when guided by
an AAN EMG-based con ol. When guided by a ajec o y- acking posi ion con ol, i would be
expec ed ha he assis ance p o ided by he AO du ing bo h
condi ioned
and
uncondi ioned
asks
would be iden ical, since his con ol s a egy does no conside he use ’s pa icipa ion. A e
comple ing hese h ee asks, pa icipan s pe o med he same con inuous asks a he same gai
speed (1.2 km/h) o he second assis i e con ol s a egy.
A e pe o ming he wo sessions (i.e., walking unde he ajec o y- acking posi ion con ol
and AAN EMG-based con ol in andom o de ), pa icipan s we e gi en a ou -ques ion open
ques ionnai e wi h he ollowing ques ions:
1. Among he asks pe o med, which condi ion p o ided mo e a iable assis ance?
2. Among he asks pe o med, which condi ion p o ided mo e adequa e assis ance?
3. Conside ing he ask o walking suspended om he sideba s, in which condi ion did you
eel ha he o hosis p o ided mo e assis ance?
4. Conside ing he possibili y o a igue, which condi ion would you choose o p olonged use
o he AO?
A e answe ing he ques ions, he p o ocol was epea ed o he emaining speeds, i.e., 1.4
and 1.6 km/h.
6.3.3 RESULTS AND DISCUSSION
A. Es ima ion o Ankle Join Re e ence To que T ajec o ies
In his Ph.D. hesis, i is p esen ed a p oo -o -concep o DL eg esso s’ applicabili y o achie e
an accu a e and gene alized me hod o es ima ing use -o ien ed ankle o que ajec o ies o he
en i e gai cycle, conside ing gai speeds anging om 1.0 o 4.0 km/h and o indi iduals wi h body
Chap e 6
116
heigh and mass a ying om 1.51 o 1.83, and om 52.0 o 83.7 kg, espec i ely. This enables
he es ima ion o ankle join e e ence o que ajec o ies o a widesp ead popula ion, including
Sma Os’ use s.
Table 6.3 p esen s he bes esul s achie ed o he DL-based eg esso algo i hm, o bo h
alida ion and es condi ions, when conside ing/no conside ing he EMG signals as inpu . Pu suing
he hypo hesis ha he ankle join o que ajec o ies can be accu a ely es ima ed wi hou using EMG
signals, a s a is ical analysis was conduc ed o e alua e possible signi ican di e ences. The
Shapi o–Wilk no mali y es showed ha all da a we e pa ame ic, and he assump ions o
homoscedas ici y and he exis ence o ou lie s we e accomplished. Thus, a wo- ailed and pai ed
-
es was conduc ed wi h a le el o con idence o 95%. The esul s indica e ha he e we e no
signi ican di e ences be ween bo h app oaches, wi h a
p
- alue > 0.05 ha ing been achie ed o all
me ics (
:
p
- alue = 0.17, and
R2
:
p
- alue = 0.28). Since he
p
- alues we e g ea e han 0.05 o all
me ics, i is concluded ha he e we e no signi ican di e ences be ween he wo CNN models. This
sugges s ha he CNN model can es ima e ankle join e e ence o que ajec o ies
accu a ely e en wi hou using EMG signals.
The bes pa ame e s ound o he CNN we e he ollowing: (i) ke nel size: 2x2; (ii) numbe o
con olu ional laye s: 2; (iii) numbe o il e s: 8 and 16 in he i s and second con olu ional laye s,
espec i ely; (i ) no maliza ion me hod: obus ; ( ) ba ch size: 20; and ( i) d opou : 25%.
Table 6.3 – DL models pe o mance when EMG signals we e/we e no included as inpu s
Inpu s
Deep
Lea ning
Model
R2
P edic ion
Time
(ms/sample)
LOSOCV
Tes
LOSOCV
Tes
Body heigh , mass, gai speed, and
ankle join angles
CNN
0.89 ±
0.03
0.92
0.91 ±
0.03
0.94
0.51
Body heigh , mass, gai speed, ankle
join angles, and EMG om
ibialis an e io
and
gas ocnemius la e alis
0.90 ±
0.04
0.89
0.92 ±
0.04
0.92
0.78
Gi en i s p ominen capaci y o es ima e he ankle join o que ajec o ies, he p edic ions
pe o med by CNN we e compa ed o he eal ankle join o que ajec o ies o he es da ase using
he Bland–Al man Plo . The esul s depic ed in Figu e 6.7 – a) show ha he p edic ions made by
CNN a e close o he eal ankle join o que ajec o ies since he majo i y o he measu es a e
Chap e 6
117
wi hin he limi s o ag eemen , he bias is close o 0 N.m, and he limi s o ag eemen a e small.
Fu he mo e, Figu e 6.7 – b) illus a es he a e age and s anda d de ia ion alues o he eal and
CNN-based p edic ed ankle join o que o he es da ase . These esul s show ha he CNN
p oduced e e ence ankle join o que ajec o ies closely simila o he eal ones.
Figu e 6.7 – (a) Bland–Al man plo esul s; (b) A e age (dashed lines) and s anda d de ia ion (SD) ( illed
a ea) o he eal and CNN-based p edic ed ankle join o que o he es da ase .
The achie ed esul s we e encou aging when conside ing he indings p esen ed in p e ious
s udies [152]–[156]. The s udy [152] eached an
R2
o 0.48 o p edic ing he peak ankle do si lexion
o que using linea and quad a ic equa ions. The s udy [153] es ima ed ankle, knee, and hip angles
and o ques o he s ance phase using eed o wa d and LSTM neu al ne wo ks, e ealing a
o
0.98. In he s udy [154], peak ankle, knee, and hip join angles and o ques we e es ima ed using
linea , quad a ic second-o de , and quad a ic hi d-o de equa ions. The highes alue o
R2
o peak
ankle join o que es ima ion was 0.93. O he s udies [155], [156] ha e epo ed ha a Mul ilaye
Pe cep on can es ima e knee join o que ajec o ies wi h an
o 0.97 by combining EMG signals
wi h kinema ic senso s.
Despi e p esen ing di e en da ase s and s udy condi ions, he CNN ool p oposed by his
esea ch p esen s an es ima ion pe o mance in line wi h he s udies [152]–[156], by p edic ing
ankle join o que ajec o ies o he en i e gai cycle wi h an
R2
o 0.91 ± 0.03 (
o 0.95 ± 0.18).
The ool de eloped in his hesis ad ances by es ima ing o que o he comple e gai cycle, which is
ad an ageous o e he es ima ion o speci ic gai e en s [152]–[156] because he o que cu e is
cha ac e ized no only in magni ude bu also empo ally [157]. Fu he , he p oposed ool
p edic s ankle join o que ajec o ies wi hou he need o complex da a acquisi ions
ela ed o EMG [155], [156].
Chap e 6
124
Table 6.7 – Va ia ion o he EMG signals om
ibialis an e io
and
gas ocnemius la e alis
, ange o mo ion (ROM) o he hip join , and human ankle join o que
be ween
uncondi ioned
and
condi ioned
asks du ing AAN EMG-based con ol s a egy (a nega i e and a posi i e sign means a educ ion and an inc ease,
espec i ely, when compa ing he a iable measu ed a he
condi ioned
ask o he
uncondi ioned
ask)
ID
EMG om
ibialis an e io
EMG om
gas ocnemius
la e alis
Hip ROM
Human o que
1.6 km/h
1.4 km/h
1.2 km/h
1.6 km/h
1.4 km/h
1.2 km/h
1.6 km/h
1.4
km/h
1.2 km/h
1.6
km/h
1.4
km/h
1.2
km/h
1
-29.7%
-32.0%
-33.5%
-70.5%
-64.7%
-57.3%
-22.7%
-24.7%
-16,1%
-14.3%
-15.0%
-14.4%
2
-63.1%
-64.5%
-63.0%
-67.1%
-62.6%
-59.8%
-4.9%
-18.6%
-1,0%
-14.0%
-11.0%
-5.8%
3
-55.7%
-63.4%
-32.0%
-22.2%
-9.4%
-11.1%
-22.0%
-17.8%
-22,8%
-16.3%
-22.7%
-20.1%
4
-1.8%
-2.0%
-16.0%
-48.9%
-18.9%
-26.8%
-48.8%
-44.3%
-38,0%
-7.0%
-9.3%
-1.2%
5
-1.7%
-9.3%
-14.2%
-37.3%
-64.5%
-59.8%
-22.8%
-21.3%
-17,6%
-3.4%
-6.7%
-8.9%
A e age ±
STD/Speed
-30.4% ±
25.9%
-34.2% ±
26.2%
-31.7% ±
17.5%
-49.2% ±
18.2%
-44.0% ±
24.6%
-42.9% ±
20.3%
-24.2% ±
14.1%
-25.3%
± 9.8%
-19,1% ±
11,9%
-11.0%
± 4.9%
-12.9%
± 5.6%
-10.1%
± 6.6%
To al
a e age ±
STD
-32.1% ± 23.2%
-45.4% ± 21.0%
-22.9% ± 11.9%
-11.3% ± 5.7%

Chap e 6
125
Table 6.8 – Va ia ion o he do si lexion and plan a lexion mo o o ques be ween
uncondi ioned
and
condi ioned
asks (a nega i e and a posi i e sign means a
educ ion and an inc ease, espec i ely, when compa ing he a iable measu ed a he
condi ioned
ask o he
uncondi ioned
ask). Maximum do si lexion, and
plan a lexion angles measu ed a he
condi ioned
ask
ID
Do si lexion Mo o To que
Plan a Flexion Mo o To que
Maximum Do si lexion Angle
(º)
Maximum Plan a Flexion
Angle (º)
1.6 km/h
1.4 km/h
1.2 km/h
1.6 km/h
1.4 km/h
1.2 km/h
1.6
km/h
1.4
km/h
1.2
km/h
1.6
km/h
1.4
km/h
1.2
km/h
1
18.4%
10.4%
8.1%
11.6%
10.9%
10.5%
11.2
11.8
12.6
-6.3
-8.1
-7.6
2
6.1%
6.9%
6.9%
11.4%
12.3%
17.6%
10.4
11.9
12.9
-7.4
-7.1
-7.4
3
16.9%
1.8%
1.4%
8.7%
1.2%
3.8%
12.7
13.8
14.6
-6.9
-7.9
-7.9
4
12.9%
26.8%
14.5%
11.0%
21.6%
5.9%
11.4
13.8
13.4
-8.8
-7.9
-8.7
5
2.4%
2.8%
12.9%
6.4%
1.5%
10.6%
11.9
13.2
14.7
-7.2
-7.2
-8.6
A e age ±
STD
11.3% ±
6.2%
9.7% ±
9.1%
8.8% ±
4.7%
9.8% ±
2.0%
9.5% ±
7.6%
9.7% ±
4.8%
11.5 ±
0.7
12.9 ±
0.9
13.6 ±
0.9
-7.3 ±
0.8
-7.7 ±
0.4
-8.0 ±
0.5
To al a e age
± STD
9.9% ± 6.6%
9.7% ± 4.8%
12.7 ± 0.83
-7.7 ± 0.59
Chap e 6
126
In addi ion, he con ibu ions o he p oposed AAN EMG-based con ol s a egy we e compa ed
o he con ibu ion o he ajec o y- acking posi ion con ol. Fo ha , he EMG signals o he
ibialis
an e io
and
gas ocnemius la e alis
, he mo o o que, and he human ankle join angle ajec o ies
we e analyzed conside ing he
uncondi ioned
and
condi ioned
asks. Table 6.9 p esen s he achie ed
esul s.
Table 6.9 – Va ia ion o he EMG signals om
ibialis an e io
and
gas ocnemius la e alis
, and do si lexion
and plan a lexion mo o o ques be ween
uncondi ioned
and
condi ioned
asks (a nega i e and a posi i e
sign means a educ ion and an inc ease, espec i ely, when compa ing he a iable measu ed a he
condi ioned
ask o he
uncondi ioned
ask). Maximum do si lexion, and plan a lexion angles measu ed
a he
condi ioned
ask
A e age ± STD
AAN
Posi ion
EMG om
ibialis an e io
-32.1% ± 23.2%
-29.2% ± 20.9%
EMG om
gas ocnemius la e alis
-45.4% ± 21.0%
-42.8% ± 19.4%
Do si lexion Mo o To que
9.7% ± 6.4%
0.8% ± 2.6%
Plan a Flexion Mo o To que
9.7% ± 4.8%
-0.9% ± 4.1%
Maximum Do si lexion Angle (º)
12.7 ± 0.8
10.5 ± 0.1
Maximum Plan a Flexion Angle (º)
-7.7 ± 0.6
-5.8 ± 1.0
Table 6.9 shows ha , du ing
condi ioned
asks o bo h con ol s a egies (AAN EMG-
based and ajec o y- acking posi ion con ols), he e was a educ ion in he muscle
ac i a ion o
ibialis an e io
and
gas ocnemius la e alis
muscles compa ed o he
uncondi ioned
asks. The dec ease in ac i a ion o he
ibialis an e io
muscle du ing he AAN EMG-
based con ol s a egy (32.1% ± 23.2%) was simila o he dec ease in ac i a ion o he same muscle
du ing he applica ion o he ajec o y- acking posi ion con ol (29.2% ± 21.0%). An iden ical esul
was ob ained o he dec ease in ac i a ion o he
gas ocnemius la e alis
muscle, wi h dec eases o
45.4% ± 21.0% and 42.8% ± 19.4% du ing he applica ion o he AAN EMG-based and he ajec o y-
acking posi ion con ol s a egy, espec i ely. This means ha he quasi-passi e walking
Chap e 6
127
mo ion pe o med by pa icipan s a di e en gai speeds was simila be ween he wo
con ol s a egies, as he e was a simila educ ion in muscle ac i a ion in bo h.
None heless, despi e a simila educ ion in muscle ac i a ion, he abili y o he
AO o assis pa icipan s di e ed be ween he wo con ol s a egies. While du ing he
AAN EMG-based con ol s a egy he AO inc eased i s do si lexion mo o o que con ibu ion (9.7% ±
6.4%) o compensa e o he lowe ac i a ion o he do si lexo muscle (
ibialis an e io
), du ing he
ajec o y- acking posi ion con ol he inc ease in he o hosis con ibu ion was e y small (0.8% ±
2.6%). The same beha io was obse ed o he plan a lexion mo o o que con ibu ion, in which
(i) he AAN EMG-based con ol inc eased he mo o o que o he AO (9.7% ± 4.8%) o compensa e
o he lowe ac i a ion o he plan a lexo s (
gas ocnemius la e alis
); (ii) he ajec o y con ol e en
dec eased he mo o o que o he AO (-0.9% ± 4.1%) in he p esence o lowe plan a lexo muscle
ac i i y. This esul is in line wi h wha would be expec ed since ajec o y- acking posi ion con ols
do no ake in o accoun he eal- ime muscula needs o he pa icipan [27], [28]. Thus, hese
s a egies ypically do no adap o he use 's pa icipa ion. Con e sely, he p oposed AAN EMG-
based con ol s a egy p ecep s he use ’s muscula pa icipa ion and acco dingly adap s he AO
assis ance. Thus, an inc ease in mo o o que is expec ed as a esul o a dec ease in he EMG
signals o he
ibialis an e io
and
gas ocnemius la e alis
.
As a esul , he abili y o each he maximum do si lexion angle (12.0º) was close
wi h he AAN EMG-based con ol (12.7 ± 0.8º) han wi h he ajec o y- acking
posi ion con ol (10.5 ± 0.1º). This di e ence in join angles was also seen in he plan a lexion
angle, whe e he AAN EMG-based con ol achie ed a alue (-7.7 ± 0.6º) close o he
maximum plan a lexion angle (-11.0º) han he ajec o y- acking posi ion con ol (-
5.8 ± 1.0º). These esul s a e also in line wi h wha would be expec ed since he AAN EMG-based
con ol no only in okes he use ’s pa icipa ion bu also a emp s o co ec he use 's walking mo ion
and p o ide assis ance when and as much as needed.
In addi ion, when compa ing he e e ence and measu ed ankle join angles, he use o he
AAN EMG-based con ol achie ed an a e age RMSE and delay o 4.94º ± 0.24º and 182.0
± 15.4 ms, espec i ely. This is e lec ed in a educ ion in RMSE o 8.0% ± 4.2% and a
educ ion in he delay o 20.7% ± 5.9% compa ed o he ajec o y- acking posi ion con ol.
A las , he eplies o each pa icipan o he applied ques ionnai e we e analyzed. Acco ding
o he p o ided eplies (depic ed in Figu e 6.11), (i) 100% o he pa icipan s ecognized ha he
AAN EMG-based con ol p o ided mo e a iable assis ance; (ii) 91.7% o he pa icipan s
Chap e 6
128
ecognized ha he AAN EMG-based Con ol p o ided mo e adequa e assis ance; (iii) 91.7%
o he pa icipan s el ha he AAN EMG-based Con ol p o ided mo e assis ance; and (i )
100% o he pa icipan s p e e ed he AAN EMG-based Con ol o he p olonged use o he
AO conside ing he exis ence o a igue. These esul s a e p omising in he sense ha mos o
he pa icipan s who es ed he p oposed AAN EMG-based con ol s a egy ecognized i s main
ad an ages, i.e., he abili y o p o ide assis ance ailo ed o he use 's needs a he ime and in he
amoun equi ed (when and as much as needed).
Figu e 6.11 – Rep esen a ion o he answe s o he open ques ionnai e.
6.4 HUMAN-IN-THE-LOOP CONTROL
6.4.1 CRITICAL ANALYSIS OF RELATED WORK
In addi ion o changes in he physiological, kinema ic, and spa io empo al pa ame e s pos -s oke
pa ien s ypically s a o ace a igue h ee mon hs a e s oke [72], [160], [161]. Be ween
25% and 85% o pos -s oke pa ien s epo a igue, i.e., hey ha e an inc eased need o es o a lack o
ene gy wi h high equency (e e y day o nea ly e e y day). These a igue episodes may in luence
he expec ed ehabili a ion ou comes since hey a ec he pa ien s’ ac i e pa icipa ion
du ing ehabili a ion sessions. Consequen ly, he abili y o accomplish daily ac i i ies is
comp omised, nega i ely impac ing pa ien s’ quali y o li e [160].
Chap e 6
129
Acco ding o s udies [37], [134], pa ien s' ac i e pa icipa ion should be encou aged o
achie e neu omuscula eco e y du ing ehabili a ion aining. Thus, AAN HITL con ol s a egies ha e
been p oposed o wea able assis i e de ices. Wi h his con ol scheme, con ol pa ame e s a e adap ed
based on physiological signals ob ained om he use [31], [162], [163]. Wi h he AAN HITL con ol
s a egy, i is expec ed ha a igue can be con olled while s ill allowing he pa ien o
ac i ely pa icipa e in he he apy.
In AAN HITL con ol s a egies, he physiological pa ame e commonly op imized is ene gy
expendi u e. Howe e , he s anda d app oach o es ima ing his signal elies on indi ec calo ime y
h ough a espi ome e , which equi es expensi e and non-po able equipmen , is ime-consuming,
p oduces noisy es ima es, and is imp ac ical o eal-wo ld applica ions [164]. A his le el, di e en AI
algo i hms ha e been employed o es ima e ene gy expendi u e using da a collec ed om
wea able senso s (EMG, IMUs, and hea a e senso s) [52], [165]–[168]. owe e , o he au ho ’s bes
knowledge, only he s udy [168] has in eg a ed he ene gy expendi u e AI algo i hm in o a knee
exoskele on o pe o m a HITL con ol scheme.
In HITL con ol schemes, he educ ion o ene gy expendi u e commonly occu s by
adjus ing he o que p o iles o he wea able assis i e de ice [31], [50], [67], [168]–[170].
Howe e , mos o he a ailable s udies (i) depend on indi ec calo ime y o es ima e ene gy expendi u e,
which limi s he wea abili y o he echnology [31], [50], [170]; (ii) ake oo long o ind he op imal con ol
pa ame e s, ypically anging om 24 o 72 minu es [31], [50], [67], [169]; and (iii) use wea able
assis i e de ices wi h an elec ic cable-d i en ansmission [31], [50], [67], [169], [170]. The ac ha
wea able assis i e de ices use an elec ic cable-d i en ansmission means ha an ankle o que pa e n
simila o he na u al human o que p o ile can be used. Howe e , acco ding o Chap e 2, elec ic
mo o -based con olle s wi h a gea -based ansmission co espond o he mos used ac ua o s
employed in wea able assis i e de ices o pos -s oke ehabili a ion. As al eady exposed o he knee join
in he s udy [168], o elec ic mo o -based con olle s wi h a gea -based ansmission, he ankle o que
pa e n o be in e p e ed by he wea able assis i e de ice mus be di e en om he na u al human o que
pa e n. O he wise, he gai pa e n will no be me because he ins an s o do si lexion and plan a lexion
ha should be pe o med will no be co ec ly execu ed by he wea able assis i e de ice. The e o e, when
designing a ajec o y- acking o que con olle o an elec ic mo o -based con olle wi h a gea -based
ansmission, he o que p o ile o be used as he e e ence ajec o y mus be manually adap ed o he
in ended applica ion.

Chap e 6
130
The AAN HITL con ol s a egy p oposed he e is designed and in eg a ed in o he Sma Os sys em
o add ess he limi a ions o exis ing s a egies in he li e a u e. The p oposed con ol employs a eg esso
p oposed in he s udy [168] o es ima e ene gy expendi u e based on da a om ou IMUs (elimina ing
he need o a espi ome e ). The inno a ion s ems om (i) he de elopmen o a HITL con ol ha allows
he ankle o que ajec o y o be op imized acco ding o he use ’s ene gy expendi u e; and
(ii) he p oposal and implemen a ion o a o que con ol o an elec ic mo o -based con olle wi h
a gea -based ansmission.
6.4.2 METHODOLOGY
A. Hie a chical Con ol A chi ec u e
The AAN HITL con ol p oposed in his esea ch ollows a hie a chical a chi ec u e o ganized
in o high-, mid-, and low-le els, as depic ed in Figu e 6.12.
Figu e 6.12 – AAN HITL con ol.
BMI
: body mass index de ined h ough he mobile g aphical applica ion
o he Sma Os sys em.
Ine ial Da a
: 3D accele a ion da a om he ches , igh w is , igh wais , and le
ankle eco ded wi h he Ine ialLab sys em.
EE
: ene gy expendi u e es ima ed e e y 10 seconds.
IT
and
TI
: cumula i e in e ac ion o que, and he in eg al o he e e ence ankle o que, espec i ely.
CP
: con ol
pa ame e s (magni ude o he plan a lexion and do si lexion o ques a he push-o ins an and mid-
swing e en , espec i ely). Ԏ
e .h
and Ԏ
meas.o
: human ankle join e e ence o que and he measu e Sma Os’
mo o o que, espec i ely.
e
Ԏ: o que e o .
u
: PID command.
SS
: speed scaling block (Equa ion 6.1).
PID
: P opo ional-In eg al-De i a i e con olle .
The high-le el con olle is esponsible o (i) ecei ing and p ocessing he 3D accele a ion
da a om 4 IMUs (Ine ialLab sys em) posi ioned a he ches , igh w is , igh wais , and le ankle,
and he body mass index (
BMI
); and (ii) es ima ing he use ’s ene gy expendi u e (
EE
) e e y 10
seconds. De ails abou da a p ocessing s eps and ene gy expendi u e es ima ion can be ound in
B.
Ene gy Expendi u e Es ima ion
.
Ene gy
Expendi u e
Es ima ion
Cubic Spline
In e pola o SS
O hosis
uman
PID
u
meas o
e
e o
Mid le el ( )Low le el ( k )
igh le el ( )
e h
CMA ES
C
ne ial
a a
M
Chap e 6
131
Once es ima ed, he ene gy expendi u e alue is sen o he mid-le el, whe e i is used oge he
wi h he cumula i e in e ac ion o que (
IT
) and he in eg al o he e e ence ankle o que (
TI
), as
inpu s o he HITL op imize (
CMA-ES
). These wo a iables (
IT
and
TI
) a e compu ed o each gai
cycle and a e ese a he beginning o he gai cycle. The
CMA-ES
block p esen s a cos unc ion
(de ailed in
D. CMA-ES op imize
) ha ends o be minimized h ough he a ia ion o wo con ol
pa ame e s (
CP,
ep esen ing he
CMA-ES
ou pu s), namely, he magni ude o he plan a lexion
and do si lexion o ques a he push-o ins an and mid-swing e en , espec i ely. These con ol
pa ame e s ac as inpu in a cubic spline in e pola o (de ailed in
C. Ankle To que P o ile
), gene a ing
he human ankle join e e ence o que (Ԏ
e .h
). The Ԏ
e .h
is in e pola ed o he Ԏ
e .o
, h ough he
Speed
Scaling
(
SS
)
block, acco ding o Equa ion 6.1.
The Ԏ
e .o
is hen sen o he low-le el and i is compa ed wi h he Sma Os’ mo o o que
(Ԏ
meas.o
), gene a ing a o que e o (
e
Ԏ). The
e
Ԏ eeds he PID con olle , using p opo ional, in eg al,
and de i a i e gains o 300, 7.5, and 2.5, espec i ely. Conside ing he
e
Ԏ, he PID con olle
compu es a PID command (
u
) ha is limi ed o maximum and minimum alues o 2500 and -2500,
h ough a sa u a o . This command is in e p e ed by he ankle o hosis, gene a ing he co esponding
mo o o que.
Wi h his con ol a chi ec u e, i is expec ed he adap ion o he con ol pa ame e s o
educe he use 's ene gy expendi u e while s ill p o iding ankle suppo .
B. Ene gy Expendi u e Es ima ion
In his Ph.D. hesis, he ene gy expendi u e was es ima ed in eal- ime using a machine
lea ning eg ession model, namely an Exponen ial Gaussian P ocess Reg ession (EGPR), p esen ed
in he s udy [171]. This eg ession model was chosen since i ou pe o med o he AI models,
including boos ed decision ees, bagged decision ees, suppo ec o machines, and CNNs [171].
The EGPR model was ained and alida ed o line using da a om a publicly a ailable 10-pa icipan
da ase [172]. The aining in ol ed LOSOCV wi h a da ase including accele a ion measu emen s
om he ches , igh w is , le wais , and igh ankle, as well as he pa icipan s' body mass index.
The ained eg ession model was implemen ed in o he Sma Os high-le el o es ima e he
ene gy expendi u e in eal- ime, e e y 10 seconds. This ime window was chosen o ensu e ha a
leas wo espi a o y cycles we e cap u ed, which is conside ed su icien o accu a e ins an aneous
me abolic cos es ima ion [171]. Mo eo e , despi e aining he EGPR wi h pa icipan s' body mass
index and 3D accele a ion da a om he ches , igh w is , le wais , and igh ankle, eal- ime es s
Chap e 6
132
we e conduc ed wi h he ollowing da a: pa icipan s' body mass index and accele a ion da a om
he ches , igh w is , igh wais , and le ankle. This al e a ion was done since (i) he Sma Os
sys em does allow an IMU senso o be posi ioned a he igh ankle join ; and (ii) no di e ences
we e ound in he ene gy expendi u e es ima ion when using he IMU senso s a he igh /le wais
and ankle.
To es ima e ene gy expendi u e, he ollowing da a p ocessing was applied. A e ecei ing he
3D accele a ion signals om he ou speci ied body loca ions, hese signals we e il e ed wi h a 4 h-
o de Bu e wo h low-pass il e a 20 Hz. This equency was chosen acco ding o s udies [165]–
[167], since i p o ides a balance be ween signal a enua ion and p ese a ion o ele an
in o ma ion. Then, he il e ed da a we e eo ganized in o 10-second windows, as sugges ed in he
s udy [171]. A las , he mean absolu e de ia ion was compu ed o each signal since i co esponds
o a good ea u e o dis inguish di e en physical ac i i y in ensi ies ([173]), being hen no malized
by he z-sco e me hod. Once p ocessed, he da a we e in e p e ed by he EGPR, and he ene gy
expendi u e was es ima ed.
C. Ankle To que P o ile
The ankle o que p o ile ep esen s he a ia ion o he o que magni ude and sign in ime,
du ing a gai cycle. Figu e 6.13 depic s he na u al human ankle join angle and o que, du ing le el-
g ound walking a se e al gai speeds anging om 1.0 o 4.0 km/h. Figu e 6.13 shows ha i he
na u al human ankle o que p o ile is used as a e e ence in a ajec o y- acking o que con ol
applied o elec ic mo o -based ac ua o s wi h a gea -based ansmission, he de ice will pe o m (i)
small do si lexion immedia ely a e he heel-s ike e en ; (ii) a con inuous plan a lexion un il he
end o he s ance phase; and (iii) a o que a ound 0 N.m h oughou he swing phase. In his way,
he kinema ic mo emen o he ankle join would no be simila o he ajec o y ypically ollowed by
his join du ing he gai cycle. In addi ion, he a ea o he cu e co esponding o he do si lexion
mo ion mus be iden ical o he a ea o he cu e co esponding o he plan a lexion mo ion.
O he wise, he posi ion o he ankle a he heel-s ike e en will be di e en o each gai cycle. I is
he e o e necessa y o c ea e a o que ajec o y speci ic o elec ic mo o -based ac ua o s wi h a
gea -based ansmission.
Chap e 6
133
Figu e 6.13 – Human ankle o que ( op iew) and ankle angle (bo om iew). Adap ed om he s udy
[121].
In an a emp o ind he bes o que p o ile o use, an empi ical es was pe o med wi h a
heal hy pa icipan walking o one minu e on a eadmill, assis ed by he Sma Os sys em ope a ing
wi h ajec o y- acking posi ion con ol, a a gai speed o 1.5 km/h. The a e age mo o o que cu e
gene a ed by he Sma Os sys em was hen eco ded o each gai cycle.
Figu e 6.14 depic s he egis e ed mo o o que p o ile. Acco ding o Figu e 6.14, he o que
pa e n was simila o he ankle angle pa e n shown in Figu e 6.13 (bo om iew), as he do si lexion
and plan a lexion phases coincided in ime. On he o he hand, he sum o he do si lexion and
plan a lexion a eas a he end o each gai cycle was app oxima ely 0 N.m, indica ing ha he ankle
posi ion was app oxima ely he same a he beginning o each gai cycle. Thus, he o que p o ile
shown in Figu e 6.14 seems o be a good candida e o be used as a e e ence in ajec o y- acking
o que con ols.
.
.
.
nkle angle
ai Cycle
A e age ial ( . km h)
A e age ial ( . km h)
A e age ial ( . km h)
A e age ial ( . km h)
A e age ial ( . km h)
A e age ial ( . km h)
A e age ial ( . km h)
Collec ed ials
Plan a Flexion
Do si lexion
Plan a Flexion
Do si lexion
nkle o que m kg