Recei ed No embe 15, 2021, accep ed Decembe 1, 2021, da e o publica ion Decembe 3, 2021,
da e o cu en e sion Decembe 20, 2021.
Digi al Objec Iden i ie 10.1109/ACCESS.2021.3132656
Machine Lea ning Based Fall De ec o
Wi h a Senso ized Tip
ASIER BRULL MESANZA 1, ILARIA D’ASCANIO2, ASIER ZUBIZARRETA 1,
LUCA PALMERINI 2, LORENZO CHIARI 2, AND ITZIAR CABANES 1
1Depa men o Au oma ic Con ol and Sys ems Enginee ing, Facul y o Enginee ing o Bilbao, Uni e si y o he Basque Coun y (UPV/EHU), 48013 Bilbao,
Spain
2Depa men o Elec ical, Elec onic, and In o ma ion Enginee ing ‘‘Guglielmo Ma coni,’’ Uni e si y o Bologna, 40126 Bologna, I aly
Co esponding au ho : Asie B ull Mesanza (asie [email p o ec ed])
This wo k was suppo ed in pa by he Uni e si y o he Basque Coun y (UPV/EHU) unde G an PIF18/067, P ojec GIU19/045, P ojec
DPI 2007-82694-R, and P ojec PID2020-112667RB-I00 unded by MCIN/AEI/10.13039/501100011033.
This wo k in ol ed human subjec s o animals in i s esea ch. App o al o all e hical and expe imen al p ocedu es and p o ocols was
g an ed by he Alma Ma e S udio um Uni e si à di Bologna E hics Commi ee unde App o al No. 149960.
ABSTRACT Fall de ec ion has become an a ea o in e es in ecen yea s, as quick esponse o hese
e en s is c i ical o educe he mo bidi y and mo ali y a e. In o de o ensu e p ope all de ec ion, se e al
echnologies ha e been de eloped, including ision sys em, en i onmen al de ec ion sys ems, and wea able
senso based sys ems. Howe e , in elde ly o impai ed people, i has been shown ha he implemen a ion
o senso s in Assis i e De ices o Walking, such as c u ches o canes, can also be a p omising al e na i e.
In his wo k, a Suppo Vec o Machine (SVM) based Fall De ec ion sys em is p oposed, which uses he
da a p o ided by a Senso ized Tip which can be a ached o di e en Assis i e De ices o Walking (ADW).
Unlike o he app oaches, he de eloped one is able o di e en ia e he all o he ADW om he all o he
use . Fo ha pu pose, he de eloped Fall De ec o uses wo modules connec ed in se ies. The i s one de ec s
all alls, while he second di e en ia es be ween use and ADW alls. The p oposed app oach is alida ed in
a se o expe imen al es s ca ied ou by heal hy olun ee s ha ha e simula ed di e en alls. In addi ion,
a compa a i e analysis is ca ied ou by compa ing he pe o mance o he Senso ized Tip based Fall De ec o
and a s a e-o - he-a comme cial accele ome e sys em. Resul s demons a e ha he p oposed app oach
p o ides high Fall De ec ion Ra ios (o e 90%), simila o highe o wea able-senso based app oaches.
INDEX TERMS Machine lea ning, suppo ec o machine, andom o es , all de ec ion, wea able senso s,
ins umen ed c u ch, moni o ing.
I. INTRODUCTION
Recen s udies, including ele an ones om he Wo ld
Heal h O ganiza ion (WHO) [1], [2], s a e ha mo e han 28%
o he popula ion o e 64 yea s su e s a leas one all pe
yea . In elde ly o physically impai ed people alls can ha e
a g ea impac on hei heal h and daily li e [3], [4]. In ac ,
alls cause physical inju ies in 6% o he cases [5], [6], om
which 14% can be se ious inju ies [7]. Mo eo e , he ea
o alls in elde ly people has an impo an impac in hei
social li e, as 15% educe hei social ac i i y ou side hei
home [6].
S udies ha e emphasised ha quick ac ion in he e en o
a all is c i ical, especially in people who li e alone, since he
longe i akes o eac o he e en , he highe he mo bidi y
The associa e edi o coo dina ing he e iew o his manusc ip and
app o ing i o publica ion was Khin Wee Lai .
o mo ali y a e is [8], [9]. Hence, he de elopmen o no el
app oaches o de ec alls and educe he eac ion ime is
c i ical o minimize he impac o hese si ua ions.
In he li e a u e, h ee main sensing sys ems ha e been
p oposed o de ec alls, which a e di e en ia ed conside ing
he na u e o he cap u ed signals [10]–[12]: 1) ision sys-
ems; 2) en i onmen al de ec ion sys ems; and 3) wea able
senso -based sys ems.
Vision sys ems [13]–[21], p ocess images o one o se e al
came as o de ec alls. An ad an age o hese sys ems is
ha hey can also p o ide an image o he allen pe son,
which helps e alua ing he se e i y o he all. Howe e , as he
sys em is designed o be s a ic, hey p esen limi ed ange
o cap u e, ypically cons ained o a speci ic oom, being
unable o de ec alls ou side his a ea. Mo eo e , ha ing a
cons an ly ac i e home ision-based sys em can cause p i-
acy p oblems [15].
164106 This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 License. Fo mo e in o ma ion, see h ps://c ea i ecommons.o g/licenses/by/4.0/ VOLUME 9, 2021
A. B ull Mesanza e al.: Machine Lea ning Based Fall De ec o Wi h a Senso ized Tip
En i onmen al de ec ion sys ems a e based on he de ec-
ion o he a ia ion o en i onmen al signals such as adio
signals [22]–[24], sound signals [25], [26] o g ound ib a-
ions [27] o de ec alls. These app oaches p esen less p i-
acy conce ns, bu hei applicabili y is also limi ed o a
speci ic cap u e ange. In addi ion, in home en i onmen s,
di e en ac i i ies can cause in e e ence wi h he moni o ing
sys ems.
Wea able senso s a e small senso s ha can be placed
almos anywhe e in he body. Thanks o hei small size
and weigh , hey can be ca ied ou by he pe son o be
moni o ed, inc easing hei cap u e ange signi ican ly. Mos
o he app oaches o de ec alls based on wea able senso s a e
based on he use o ine ial senso s. Among hese, mul iple
solu ions can be ound using accele ome e s in he li e a-
u e [19], [28]–[39]. Some wo ks also p opose he use o
IMUs (Ine ial Measu emen Uni ) which combine he o me
wi h gy oscopes and magne ome e s and allow o es ima e
he 3D o ien a ion o he de ice in a global e e ence sys-
em [40]–[43]. A pa icula subse o hese app oaches use
he in e nal IMUs o cu en sma phones [44]–[47]. O he
app oaches p opose he use o ba ome e s [48] o e en he
mic ophone o sma phones [25].
In ecen yea s, wea able senso s ha e become one o he
main app oaches o Fall De ec ion. Howe e , i is o be no ed
ha hei placemen wi h espec o he body is a c i ical issue
when p ocessing he cap u ed da a, as he ecei ed signals
will a y depending on his ela i e posi ion. Mos o he
wo ks p opose o place he senso s on he wais [29], [31],
[34], [35], [48]. Ne e heless, o he s p opose hei use on
he w is [19], [28], [33], he oo [32] o he back [41].
De ining he op imal placemen o senso s has been he ocus
o di e en s udies [30], [42], e alua ing hei placemen in
he ankles, ches and wais [30], and adding o hese he head,
w is s and high [42]. The a o emen ioned s udies conclude
ha he bes posi ion o pe o m Fall De ec ion is a he wais ,
al hough op imal esul s a e also achie ed wi h he senso
elemen loca ed on he ches [30], [38].
Al hough in he las yea s he size o wea able senso s
has educed, in elde ly o impai ed people, he a achmen
o he senso o he body can cause ejec ion by he use .
In hese cases, se e al wo ks ha e p oposed o in oduce
senso s in o Assis i e De ices o Walking (ADW) such as
c u ches [49] o canes [50]–[52] in o de o de ec alls.
The p oposed de ices use ine ial senso s [52] which can be
combined wi h o ce senso s [49], [50], o GPS and hea
a e senso s [51]. These de ices allow minimal discom o
o he use , bu also equi e a p ope algo i hm o de ec he
all.
Fall De ec ion is ca ied ou by in e p e ing he da a p o-
ided by he senso s in eg a ed in he a o emen ioned de ices.
Two main app oaches exis o his pu pose. The i s p o-
cesses di ec ly he aw da a o he senso s [28], [33], equi ing
algo i hms ha ypically imply highe compu a ional cos .
The second conside s a p e-p ocessing s ep, in which a se
o ea u es a e ex ac ed om he aw da a, educing he
dimensionali y o he p oblem [32], [34], [37], [42], [46],
[53], [54].
The implemen a ion o he Fall De ec ion algo i hm is ypi-
cally add essed by he design o a machine lea ning echnique
based classi ie [55]. Among he di e en app oaches, A i-
icial Neu al Ne wo ks (ANN) based on MLP (Mul i-Laye
Pe cep on) [31], [33], [34], [41], [42], [53], Con olu ional
Neu al Ne wo ks (CNN) [17], [19], [24], o Deep Lea n-
ing app oaches [22], [28], [35], [46] can be ound. O he
classi ica ion app oaches based on SVM (Suppo Vec o
Machine) [26], [30], [32], [36], [37], [41], [42], [46]–[48],
o K-NN (K-Nea es Neighbo ) [14], [15], [37] ha e been
also p oposed. The a o emen ioned solu ions p o ide a high
a e o Fall de ec ion when applied o di e en de ices. How-
e e , in he case o Fall De ec o s de eloped o ADW, he
p oposed app oaches ha e no been designed o di e en ia e
be ween he use alling wi h he ADW, and he ADW alling
wi hou he use .
In summa y, i can be concluded ha due o he impo ance
o quick ac ion in he e en o alls, hei de ec ion using
moni o ing de ices has aised as a ele an esea ch line in
ecen yea s. In he case o impai ed o elde ly people he
use o senso ized ADW has been p oposed as an app op ia e
app oach. Howe e , mos wo ks p oposed in his a ea do no
conside hese de ices. Mo eo e , he p oposed ADW-based
Fall De ec o s a e p one o alse posi i es, as hey a e no able
o disce n when he ADW has allen wi h he use o wi hou
i .
Hence, in his wo k, a no el Fall De ec ion app oach is p o-
posed o people ha equi e ADW. The p oposed app oach
is based on a Senso ized Tip which can be a ached o a
s anda d c u ch o cane, and aims o gi e some insigh in o
he p e iously ci ed issues, wi h ou ele an con ibu ions:
1) The app oach is ocused on people ha equi e ADW;
2) A comp ehensi e se o ea u es o de ec alls is p o-
posed and op imized using a Fea u e Selec ion me hodology;
3) Falls o he ADW wi hou he use ( alse posi i es) a e con-
side ed; 4) A compa a i e analysis is ca ied ou conside ing
ou di e en scena ios: using only da a om he Senso ized
Tip, om he wea able senso s, om all accele ome e da a;
o using all da a.
The es o he wo k is s uc u ed as ollows. Sec ion II
de ails bo h he de eloped Senso ized Tip and he wea -
able senso s used o he de elopmen o he Fall De ec o s.
Sec ion III p esen s an o e iew o he p oposed wo-s ep
Machine Lea ning-based Fall De ec ion app oach. Sec ion IV
de ails he expe imen s execu ed o gene a e he da ase s o
de elop he Fall De ec o . Sec ion Vexplains he me hodol-
ogy used o de ine he all De ec ion algo i hms. Sec ion VI
shows he esul s o he compa a i e analysis ca ied ou o
e alua e he app oach. Finally, he mos impo an ideas a e
summa ized in Sec ion VII.
II. FALL MONITORING SYSTEMS
In his wo k, he use o he Senso ized Tip de eloped in [56]
is p oposed o moni o he use ha equi es an Assis i e
VOLUME 9, 2021 164107
A. B ull Mesanza e al.: Machine Lea ning Based Fall De ec o Wi h a Senso ized Tip
FIGURE 1. Senso ized Tip used o all de ec ion (on a c u ch and o ) and
local axes o he Senso ized Tip.
De ice o Walking. As i can be seen in Figu e 1, he
Senso ized Tip can be easily a ached o a c u ch o cane,
p o iding da a o bo h he use ’s mo ion and he o ce
exe ed.
The Senso ized Tip is made o a ligh weigh aluminum
s uc u e, which con ains a se o senso s: an Ine ial Mea-
su emen Uni wi h 9 deg ees o eedom MTi-3 om XSens,
which p o ides in o ma ion on he 3D mo ion o he Sen-
so ized Tip (linea accele a ion, angula eloci y and mag-
ne ic ield in he local xyz axes); a BMP280 ba ome e om
Bosch ha can p o ide es ima ion on he ela i e heigh o
he Senso ized Tip; and a C9C o ce senso om HBM ha
p o ides he axial o ce applied on he ADW. In addi ion, he
MTi-3 p o ides an es ima ion o he global o ien a ion o he
de ice on a global XYZ coo dina e sys em, which allows o
es ima e i s angle o inclina ion (α) wi h espec o he g ound
by,
α=π/2−acos(p ojZz/|p ojZz|) (1)
whe e p ojZzis he p ojec ion o he local zaxis (Figu e 1)
in he global Zaxis (no mal o he g ound) and |p ojZz|is i s
module.
I is o be no ed ha he da a om he magne ome e and
he BMP280 ba ome e will no be used in his wo k.
In o de o e alua e he a o emen ioned de ice as a all
moni o ing sys em, in his wo k, he Fall De ec o s will
be also be de eloped o a wea able senso sys em. The
FIGURE 2. GENEAc i accele ome e senso s posi ions in he body.
GENEAc i comme cial 3-axis accele ome e s, manu ac-
u ed by Ac i insigh s, ha e been selec ed o his pu pose.
In pa icula he GENEAc i de ices we e loca ed on he non-
dominan w is , on he ches , on he lowe back, and in he
pocke co esponding o he dominan side (see Figu e 2).
The GENEAc i wea able senso s on he w is and in he
pocke a e used o simula e a sma wa ch and a sma phone
espec i ely. This way, he di e en senso da a p o ided by
he Senso ized Tip and he Wea able sys em can be e alua ed
o analyze hei e ec i eness o de ec alls.
III. OVERVIEW
This pape p esen s a no el Fall De ec o app oach based on
he da a p o ided by a Senso ized Tip a ached o an Assis-
i e De ice o Walking (ADW). The p oposed app oach is
composed by wo modules connec ed in se ies, as de ailed in
Figu e 3. The i s module (ADW Fall De ec o ) is ocused
on de ec ing he all o he ADW; while he second (Use &
ADW Fall De ec o ) uses he all da a o e alua e i he use
has allen wi h he ADW, o only he ADW has allen. This
la e module is designed o a oid alse posi i es due o ADW
acciden al alls.
Fo he de elopmen o he ADW Fall De ec o module,
wo expe imen ally ob ained da ase s will be used o gene -
a e he aining se : a da ase composed by use alls and a
da ase ha includes di e en physical ac i i ies ca ied ou
by he use . Fo he Use & ADW Fall De ec o module,
whe e he goal is o de e mine i he use has allen wi h
he ADW, he use all da ase will be combined wi h a
se o expe imen s in which only he ADW has allen o
164108 VOLUME 9, 2021
A. B ull Mesanza e al.: Machine Lea ning Based Fall De ec o Wi h a Senso ized Tip
FIGURE 3. Two-module me hodology ollowed o alls de ec ion.
gene a e he aining se . The p o ocol ha was de ined o
ob ain he di e en da ase s will be de ailed in Sec ion IV.
This wo-module app oach has been designed in o de o
de elop wo di e en Machine-Lea ning based de ec o s. The
p esen ed wo-module app oach has also a educed compu a-
ional cos . In ac , i a all o he ADW is no de ec ed by he
i s module, he second module is no applied.
The aining da ase s a e p ocessed o gene a e a se o
ea u es o cha ac e ize each all, and a ea u e e alua ion
p ocedu e is implemen ed o de ec he mos ele an ones o
design each Machine Lea ning-based module. In pa icula
a Suppo Vec o Machine (SVM) app oach will be used o
implemen he algo i hm o each module. The p ocedu e will
be de ailed in Sec ion V.
Finally, in o de o e alua e he p oposed app oach, his
will be compa ed wi h he pe o mance o a Fall De ec o ha
uses di e en se s o senso da a: wi h GENEAc i wea able
senso da a, all possible accele ome e s (Senso ized Tip in e -
nal accele ome e and ou GENEAc i accele ome e s) and
all da a senso s. I is o be no ed ha o hese pa icula cases,
only he i s module (see Figu e 3) will be implemen ed,
as he senso s a e placed also in he use . Resul s will be
analyzed in Sec ion VI.
IV. EXPERIMENTAL PROTOCOL AND DATASET
GENERATION
In o de o de elop a Machine Lea ning-based algo i hm,
a p ope da abase is o be gene a ed. This equi es he de -
ini ion and execu ion o a p o ocol con aining a se o alls
and physical ac i i ies while using an ADW. In his sec ion,
he de ini ion o he expe imen s is de ailed.
The simula ions we e ca ied ou by 12 heal hy olun-
ee s (4 women and 8 men, anging be ween 25-40 yea s,
3 le -handed and he es igh -handed) in a con olled
en i onmen . In o de o pe o m he alling simula ions,
a ma ess was used, o a oid possible inju ies o he ol-
un ee s. The olun ee s wo e he GENEAc i accele ome e s
du ing he expe imen s, and he Senso ized Tip was a ached
o he c u ch (Figu e 2). The p o ocol was app o ed by he
E hics Commi ee a Uni e si y o Bologna and all pa ici-
pan s p o ided in o med w i en consen .
Th ee da ase s ha e been c ea ed in o de o gene a e he
aining se o each module (ADW Fall De ec o and Use
& ADW Fall De ec o ): a) Use Fall da ase , which included
da a om people alling while using a c u ch; b) Use
Physical Ac i i ies da ase , which included da a om peo-
ple pe o ming di e en physical ac i i ies using he c u ch;
c) ADW Fall da ase , in which he c u ch was le s anding
s ill a di e en posi ions, and hen o ced o all wi hou he
use . As p e iously de ailed, da ase s 1 and 2 will be used o
ain he ADW Fall De ec o module, while da ase s 1 and 3
a e used o ain he Use & ADW Fall De ec ion mod-
ule. Nex , he expe imen s included in each da ase a e
de ailed:
A. USER FALL DATASET
In o de o simula e as close as possible eal alls, ideos
associa ed o alls o people alling while using ADW om
he Da ab a y da abase [57], [58] we e analyzed. F om his
analysis, 16 scena ios we e conside ed, in pa icula he p o-
ocol de ined includes 8 s a ic alls om an up igh posi ion
(1-8) and 8 dynamic alls om walking (9-16):
1) While s anding s ill, y o ake a s ep and ip o e he
ADW and all o wa ds.
2) Fall o wa ds.
3) Fall backwa ds simula ing a ain .
4) Fall backwa ds.
VOLUME 9, 2021 164109
A. B ull Mesanza e al.: Machine Lea ning Based Fall De ec o Wi h a Senso ized Tip
FIGURE 4. G aphical ep esen a ion o some o he simula ed alls du ing he walking: a) Cases 9, 10, 11, 12 and 13. b) Cases 14,
15. c) Case 16.
5) Ro a e 90◦ o he igh and all on he igh side.
6) Fall on he igh side.
7) Ro a e 90◦ o he le and all on he le side.
8) Fall on he le side.
9) Walk owa ds he ma ess, ip o e he ADW and all
o wa ds (see Figu e 4a).
10) Walk owa ds he ma ess, simula e a ip o e an
objec and all o wa ds (see Figu e 4a).
11) Walk owa ds he ma ess, simula e a ip o e an
objec and all on he le side (see Figu e 4a).
12) Walk owa ds he ma ess, simula e a ip o e an
objec and all on he igh side (see Figu e 4a).
13) Walk owa ds he ma ess, simula e a ip o e an
objec and all backwa ds (see Figu e 4a).
14) Loss o balance, y o eco e i by walking a ew
me e s and all o wa ds (see Figu e 4b).
15) Loss o balance, y o eco e i by walking a ew
me e s and all backwa ds (see Figu e 4b).
16) Walk and slide o end up alling backwa ds (see
Figu e 4c).
B. USER PHYSICAL ACTIVITY DATASET
In o de o comple e he da abase wi h no- all ac i i ies,
a o al o 7 di e en physical ac i i ies using ADW ha e been
simula ed:
1) Walking a a no mal pace: a ci cui (see Figu e 5) has
been de ined in which he olun ee has o walk s aigh
in se e al di ec ions and make u ns.
2) Walking quickly: he same ci cui (see Figu e 5) pe -
o med p e iously is epea ed, bu in his case walking
app oxima ely 30% as e .
3) S anding s ill: s ay s ill in place o 30 seconds.
4) Going up and down s ai s: going up and down s ai s
epea edly.
5) Ge up and si in a chai : ge up and si down epea edly
o 30 seconds.
164110 VOLUME 9, 2021
A. B ull Mesanza e al.: Machine Lea ning Based Fall De ec o Wi h a Senso ized Tip
FIGURE 5. G aphical ep esen a ion o some o he physical ac i i ies walking ci cui .
6) Pick up an objec om he loo and s and up epea edly
o 30 seconds.
7) Loss o balance wi hou alling (nea all), epea ed
4 imes.
C. ADW FALL DATASET
Finally, a se ies o es s has been ca ied ou in which he
ADW alls wi hou he use :
1) C u ch placed in di e en s a ic posi ions on he loo
o while leaning on a si e.
2) D opping he c u ch while s anding s ill, o while walk-
ing. 80 c u ch alls will be pe o med.
The da ase consis s o 192 use alls (using ADW),
108 minu es o physical ac i i ies (using ADW), 5 minu es
o di e en s a ic ADW posi ions and 80 ADW alls.
V. DESIGN METHODOLOGY
Once he da ase s ha e been gene a ed, he wo algo i hms
p oposed in Figu e 3will be designed. The i s will be a ADW
Fall De ec ion module, which will be designed o de ec a all;
while he second will de e mine i he use is in ol ed in he
all (o only he ADW). As he sys em is designed so ha he
i s module ou pu is used in he second one, each algo i hm
will equi e di e en inpu da a, as i will be explained nex .
A. ADW FALL DETECTOR DESIGN
The pu pose o he ADW Fall De ec ion module is o de ec
when a all happens while using he ADW. In his sec ion,
he me hodology used o design he Machine Lea ning-based
de ec o will be de ailed (see Figu e 7). This me hodol-
ogy is based on well-es ablished me hodologies ones in he
li e a u e [39].
1) DATA SEGMENTATION AND SET GENERATION FOR
TRAINING
The da a used o design he ADW Fall De ec o module is
ex ac ed om he Use Fall da ase and he Use Physical
Ac i i y da ase p e iously de ailed. The ime sequences cap-
u ed in hese da ase s a e i s p ocessed using a segmen a-
ion p ocess, allowing o ex ac a se o ea u es om each
segmen o window.
Fo his pu pose, he da a is di ided in o ixed-size sliding
windows. The window size has been se o 100 samples
(2 seconds), as in he expe imen s his alue allows o cap u e
he all (see Figu e 6a). In addi ion, he beginning o each
window will be shi ed by 20 samples (0.4 seconds) om
he beginning o he p e ious one (see Figu e 6b) o limi he
compu a ional cos .
Once segmen a ion has been ca ied ou , each window
will be conside ed a sample o he design o he ADW
Fall De ec o . Fo his pu pose, each window is labelled o
de ine i i co esponds o a all o no . The e en o a all
will be conside ed i a window con ains mo e han 50% o
i s da a samples associa ed o a all (see Figu e 6a). No e
ha he physical ac i i y ela ed samples a e no agged as
alls.
In o de o de elop he Machine Lea ning-based ADW
Fall De ec o , he a o emen ioned se is di ided in o wo: a
aining da ase , which will be used o de elop he ADW Fall
De ec o , and a es da ase , which will be used o alida e i s
gene aliza ion capabili ies. The aining da ase is composed
he simula ions ca ied ou by 8 subjec s, while he emaining
da a (4 subjec s) a e used o es ing. In addi ion, in o de o
balance he numbe o all/no alls samples, an adjus ed se
is gene a ed, as de ailed in Table 1. This adjus men has been
made, in he case o alls, by elimina ing hose windows ha
do no ha e 50% o he window in he all pe iod. In he case
o physical ac i i ies, his adjus men has been made ying
o main ain a simila numbe o samples wi h espec o he
alls.
As de ined in Sec ion II, each se associa ed o he ADW
Fall De ec o will con ain he segmen ed windows ela ed o
he da a cap u ed om he Senso ized Tip: 3-axis accele om-
e e , 3-axis gy oscope, o ce senso and es ima ed inclina ion
(α) wi h espec o he g ound (see Table 2).
VOLUME 9, 2021 164111
A. B ull Mesanza e al.: Machine Lea ning Based Fall De ec o Wi h a Senso ized Tip
FIGURE 6. G aph o he ADW inclina ion angle: a) ADW all ime and s a es. b) Fixed-size window di ision wi h space be ween window s a s.
TABLE 1. Dis ibu ion o he da ase and adjus men o he numbe o
da a.
2) POTENTIAL FEATURES SET GENERATION
The use o segmen a ion allows o ob ain a da ase composed
by disc e e da a uni s, one o each window, om which a
se o ea u es can be easily ex ac ed. These ea u es (such
as mean, a iance, ...) allow o educe he dimensionali y
o he da a, gene a ing nume ic alues ha can be easily
p ocessed by Machine Lea ning-based app oaches. In his
sec ion, a me hodology o selec he mos app op ia e ea u es
o design a Machine Lea ning (ML)-based Fall De ec o is
de ailed (see Figu e 7).
In he li e a u e, he e a e di e en app oaches o de ine he
se o po en ial ea u es. A ypical app oach is o use s a is ical
ope a o s o cha ac e ize he da a om he window. In his
wo k, he ollowing s a is ical ea u es will be ex ac ed:
•Mean (MEAN).
•S anda d De ia ion (STD).
•Va iance (VAR).
•Ku osis (KUR).
•In e cua ile Range (IR).
•A ea Unde he Signal (AUS).
•Maximum alue o he window (MAX).
•Minimum alue o he window (MIN).
These s a is ical ope a o s a e applied o he p e iously
de ined aining da ase , composed by he segmen ed win-
dows o samples ela ed o each o he signals p o ided by he
moni o ing de ice (Senso ized Tip’s o ce senso , Senso ized
Tip’s gy oscope (x,y,z), Senso ized Tip’s accele ome e
(x,y,z) and Senso ized Tip’s inclina ion angle (α)). The
combina ion o hese ope a o s on each senso signal de ice
gene a es a ea u e. In he case o he Senso ized Tip, a o al
o 64 ea u es can be de ined pe each sample.
3) ADW FALL DETECTOR TRAINING
Al hough all possible ea u es can be used o ain he
ML-based Fall De ec o , due o he high dimension o he
inpu da a, i is ad isable o pe o m an analysis o de ec
he mos ele an ea u es. This will allow o educe he
compu a ional cos o he app oach, i implemen ed in eal-
ime.
In he li e a u e, di e en app oaches a e p oposed o de e -
mine he ela i e impo ance o a ea u e o a classi ica ion
p oblem, such as Random Fo es (RF) [59] and Relie [60].
In his wo k, he Random Fo es app oach has been selec ed,
as i p o ided be e esul s. This app oach consis s o he
gene a ion o a la ge se o decision ees o classi ica ion
pu poses, also known as a o es . The ees a e gene a ed
by using a andom se o samples and ea u es, so ha in
he aining p ocess di e en ea u es can be es ed and hei
ela i e impo ance e alua ed.
Hence, once he aining da ase is p ocessed by he Ran-
dom Fo es and he ea u es ha e been o de ed conside ing
hei ela i e impo ance o he Fall De ec ion p ocess, a se
o Suppo Vec o Machines (SVMs) will be ained, consid-
e ing di e en subse s o ea u es. The goal is o de e mine
he minimum numbe o ea u es o achie e an app op ia e
Fall De ec ion pe o mance.
To achie e his goal, i s he mos ele an ea u es will be
used o ain he Fall De ec o SVM, hen he numbe o ea-
u es will be g adually inc eased. Each SVM is ained using
Ma lab’s S a is ic and Machine Lea ning oolbox, whe e he
SVM hype pa ame e s a e op imized by he use o a K-Fold
c oss alida ion app oach wi h K=10. Once ained, he es
se is used o e alua e he Fall De ec ion pe o mance o each
SVM. Resul s will be de ailed in Sec ion VI.
164112 VOLUME 9, 2021
A. B ull Mesanza e al.: Machine Lea ning Based Fall De ec o Wi h a Senso ized Tip
FIGURE 7. Me hodology ollowed o he design o he Fall De ec o modules based on Machine Lea ning.
B. USER AND ADW FALL DETECTOR
This algo i hm is execu ed only when he ADW Fall De ec o
module has de ec ed a all. In his scena io, Use & ADW Fall
De ec o module analyzes he all da a in o de o de e mine
i he use has allen wi h he ADW, o only he ADW has
allen. This is one o he no el con ibu ions o he p esen
wo k.
The me hodology used o he de elopmen o his module
is simila o he p e ious one (see Figu e 7). Howe e , he
inpu da ase s di e , as da a om he ADW Fall and Use
Fall da ase s a e used o ain and es his algo i hm.
1) DATA SEGMENTATION AND SET GENERATION FOR
TRAINING
This module uses he da a p o ided by he ADW Fall De ec-
ion module. Hence, he sample de ec ed as a all by he la e
ADW Fall De ec o will be p ocessed in his algo i hm. Based
on his p emise, a se o all samples is gene a ed om he
Use Fall da ase , in which he de ec ed cen al all window
is only conside ed o his module. In addi ion, he alls
associa ed o ADW all da ase will also be included, agged
as nega i e use alls. A se composed by a o al o 192 use
alls and 80 ADW alls (wi hou he use ) is gene a ed, om
which 128 use alls and 50 ADW alls a e used o aining
(Table 1).
2) USER AND ADW FALL DETECTOR TRAINING
Once he da ase s a e gene a ed, he s a is ical ope a o s p e-
iously de ailed a e applied on he senso signals o he
Senso ized Tip o ex ac he 64 ea u es associa ed o each
sample. These a e hen p ocessed h ough a Random-Fo es
app oach, ob aining he ela i e impo ance o each ea u e.
A se o SVMs is ained using he same app oach as he ADW
Fall De ec o module.
VI. RESULTS AND COMPARATIVE ANALYSIS
This sec ion ocuses on e alua ing he p oposed wo-module
Fall De ec ion app oach using he da a p o ided by he
Senso ized Tip. Fo ha pu pose, a compa a i e analysis is
ca ied ou by conside ing also he da a p o ided by he
wea able senso GENEAc i de ailed in Sec ion II.
In he analysis ou cases a e compa ed: 1) The p oposed
app oach based on he Senso ized Tip Da a; 2) The use o he
GENEAc i e wea able senso da a; 3) The use o all possible
accele ome e s (Senso ized Tip in e nal accele ome e and
ou GENEAc i accele ome e s); 4) The use o all da a
senso s (Table 2).
In o de o pe o m he compa ison, he p ocedu e o design
he ADW Fall De ec o module has been applied o all a o e-
men ioned cases: ea u e gene a ion, Random Fo es -based
ela i e impo ance de ec ion and SVM aining. No e ha he
Use & ADW Fall De ec ion module has been only used o
he Senso ized Tip case, as no senso is placed on he use .
Hence, his case will be analyzed in a sepa a e subsec ion.
A. ADW FALL DETECTOR MODULE EVALUATION
1) FEATURE RELEVANCE ANALYSIS
Following he ea u e ex ac ion p ocedu e, a se o 64 ea-
u es pe sample a e gene a ed o he da ase based on
he Senso ized Tip’s senso da a; 96 o he case o he
GENEAc i wea able senso s; 120 i all accele ome e da a
is conside ed; and 160 i all senso da a is conside ed.
As he numbe o ea u es is impo an , in o de o
educe he dimensionali y o he p oblem, a Random-Fo es
app oach is used o each case o de ec he mos ele an
ea u es, de ailed in Sec ion VI. These a e de ailed in he Fall
De ec o columns o Table 3.
As i can be seen, when he angle o inclina ion o he ADW
is conside ed in he se o da a, he ea u es ex ac ed om his
signal a e among he mos ele an , he maximum angle being
he mos impo an , as i e lec s la ge a ia ions due o alls.
In he case o GENEAc i senso s, he 3 mos impo an
ea u es a e de i ed om he senso on he lowe back o
he use (in pa icula i s X axis, e ical), which is he one
which su e s he mos a ia ion when he use alls. No e ha
ea u es om senso s loca ed on he ches a e also among he
10 mos ele an ea u es.
I only accele ome e da a is used om bo h he Sen-
so ized Tip and he GENEAc i senso s, he mos ele an
VOLUME 9, 2021 164113
A. B ull Mesanza e al.: Machine Lea ning Based Fall De ec o Wi h a Senso ized Tip
TABLE 2. Analyzed Cases conside ing senso inpu .
TABLE 3. Weigh o he ea u es p o ided by he RF in he di e en case s udies. (α=ADW inclina ion angle, accel =accele ome e , gy o =gy oscope),
o he ADW Fall De ec o and he Use & ADW Fall De ec o .
ea u es include bo h he Senso ized Tip’s accele ome e and
he GENEAc i senso on he lowe back, as in he p e ious
case.
These ends seem o be con i med i all senso s a e
used, being he Tip inclina ion angle, he accele a ion o
he back senso and ip senso among he mos ele an
ones.
2) PERFORMANCE ANALYSIS
As de ined in Sec ion V, once he ela i e impo ance o
ea u es has been de e mined, a se o SVMs is ained wi h an
inc easing numbe o ea u es, aking in o accoun he mos
ele an ea u es.
Table 4shows he pe o mance esul s o he Fall De ec o
associa ed o each numbe o he i s nmos ele an ea u es
( i s column). In gene al, all he analyzed cases p o ide
F-sco e o e 0.96, which alida e he use o he Senso ized
Tip. In addi ion, i can be seen ha he numbe o ea u es
used is no especially ele an , since e y good esul s a e
achie ed o all scena ios. Howe e , he e a e sligh di e -
ences be ween he app oaches ha can be analyzed.
Using only he senso s included in he Senso ized Tip,
esul s o his case a e good o any numbe o ea u es.
Conside ing he 2 mos ele an ea u es (αmaximum and
Tip gy oscope minimum in Z axis, Table 3) p o ide he
bes esul s: a p ecision o 0.986, a speci ici y o 0.988,
164114 VOLUME 9, 2021
