a review on fall prediction and prevention system for ...fall prediction systems typically estimate...

Review ArticleA Review on Fall Prediction and Prevention System forPersonal Devices Evaluation and Experimental Results

Masoud Hemmatpour Renato Ferrero BartolomeoMontrucchio andMaurizio Rebaudengo

Dipartimento di Automatica e Informatica Politecnico di Torino Italy

Correspondence should be addressed to Masoud Hemmatpour masoudhemmatpourpolitoit

Received 27 July 2018 Revised 1 April 2019 Accepted 20 May 2019 Published 1 July 2019

Academic Editor Antonio Piccinno

Copyright copy 2019 Masoud Hemmatpour et al This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited

Injuries due to unintentional falls cause high social cost in which several systems have been developed to reduce them Recently twotrends can be recognized Firstly the market is dominated by fall detection systems which activate an alarm after a fall occurrencebut the focus ismoving towards predicting and preventing a fall as it is themost promising approach to avoid a fall injury Secondlypersonal devices such as smartphones are being exploited for implementing fall systems because they are commonly carried bythe user most of the day This paper reviews various fall prediction and prevention systems with a particular interest to the onesthat can rely on the sensors embedded in a smartphone ie accelerometer and gyroscope Kinematic features obtained from thedata collected from accelerometer and gyroscope have been evaluated in combination with different machine learning algorithmsAn experimental analysis compares the evaluated approaches by evaluating their accuracy and ability to predict and prevent a fallResults show that tilt features in combination with a decision tree algorithm present the best performance

1 Introduction

Health centers have to deal with a large number of patientsdue to unintentional falls resulting in huge cost on thesociety For example the average hospital cost for fall injuryis over $ 30000 [1] Thus there is a critical need for thedevelopment of cost-effective systems to reduce the injuriesof a fall and to give faster assistancewhen a fall occurs Severalrisk factors for falling can be identified and specific interven-tions can be designed in order to reduce injuries To this endseveral systems were developed and are now available Mostof these systems concern fall detection [2ndash8] and they onlynotify userrsquos acquaintances after a fall occurrence Howeverthere are systems with the goal of predicting and preventing afall called fall prediction and prevention systems (FPPSs) [9ndash13] Such systems track and report data fromwearable sensorswithout engaging the users in the monitoring process FPPSsinclude sensors to collect data and software applications toprocess them first data are collected from sensors then thecollected data are analyzed to extract an appropriate featureset Afterwards a machine learning algorithm is applied on

the obtained data Since smartphones nowadays are broadlyused as personal digital assistance and they are equippedwith precise sensors and communication component theyare used commonly in FPPSs as amonitoring device to collectdata

Fall prediction systems typically estimate real-time orfuture fall risk These systems are helpful in reducingthe financial and health consequences of a fall Sinceboth prediction and prevention systems evaluate a fallrisk sometimes prediction and prevention terms are inter-leaved These systems are essential to check the feasibilityof performing recovery mechanisms before a fall occur-rence

Real-time fall prediction system aims to identify anabnormal gait pattern in order to estimate the probabilityof a real-time fall occurrence [14ndash16] In real-time fall riskprediction data are collected from sensors and are analyzedto compute the appropriate feature set Then the risk of apossible fall is evaluated through classification algorithmsReal-time systems continuously assess the fall risk while theuser is doing hisher daily activity When an abnormal gait is

HindawiAdvances in Human-Computer InteractionVolume 2019 Article ID 9610567 12 pageshttpsdoiorg10115520199610567

2 Advances in Human-Computer Interaction

FALL PREDICTION AND PREVENTION

Physiology

Sensor

ECG

EMG

Blood Pressure

GSR(EDA)

Neuromuscular noise

Heartbeat

Skin electrical characteristic

Sensor

Accelerometer

Gyroscope

User Profile

Velocity

Tilt

Acceleration

Motion sensor

Piezoelectricsensor

Phase Transition

Stability and Symmetry

Kinematics

Fall Factor

Foot State

Fall Factor

Figure 1 Fall prediction and prevention taxonomy

detected then the user is alerted [14ndash16] or an external aidsuch as a walker or robot is exploited to prevent a probablefall [17 18]

Future fall prediction is estimated through some clinicalassessment tests Probable future falls are prevented throughimproving gait and mobility by some exercises [17 18] Thesetests often involve questionnaires or functional assessmentsof posture gait cognition and other risk factors Theseclinical tests are subjective and qualitative and typically usethreshold assessment scores to categorize people as fallers andnonfallers Typically these tests are timed up and go (TUG)[19] Berg Balance Scale (BBS) [20] sit to stand (STS) [21]and one leg stand (OLS) [22] to evaluate balance and lowerlimb strength

The design of a fall prediction and prevention systemfaces several significant challenges They need to be accuratereliable robust and cost-effective [1] In this paper a fallprediction and prevention system is described in three partsfall factors (ie fall symptoms) their features and machinelearning algorithms This paper investigates every mentionedstage and experimentally evaluates the various approachesThis paper does not present systems using sensors such ascamera and sound since they are prone to violate individualrsquosprivacy comparing to kinematic wearable sensors In thispaper accelerometer (for measuring the acceleration) andgyroscope (for measuring the angular rate around one ormore axes of the space) are considered for evaluation Thesesensors are chosen since they are easily accessible and do notdisturb the privacy of the userThemain contributions of thispaper are the following

(i) A comprehensive discussion on fall prediction andprevention systems

(ii) Preparing a dataset with realistic parameters to simu-late abnormal gait

(iii) Finding the most representative fall factors(iv) Evaluation of the fall factors based on the extracted

feature on commonly used machine learning algo-rithms

This paper is organized as follows In Section 2 existingFPPSs are classified according to the fall factorsThen in Sec-tions 3 and 4 feature extraction techniques and machinelearning algorithms are described Evaluation criteria areillustrated in Section 5 Afterwards experimental results areshown in Section 6 Finally conclusions are presented inSection 7

2 Classification of Fall Factors

Fall prediction and prevention is a multifaceted problemthat can be broadly categorized into two different domainsphysiology and kinematics as can be seen in Figure 1

Physiological solutions consider intrinsic fall factors ieparameters which mostly originate from the body Thesesolutions entail an in-depth medical evaluation of the riskfactors and exploit sensors related to body monitoring

(i) Electrocardiogram (ECG) sensorTypically ECG is used to assess the electrical andmuscular operations of the heart but ECG sensing can

Advances in Human-Computer Interaction 3

also determine abnormalities that might lead to a fall[23 24]

(ii) Electromyography (EMG) sensorEMG is a technique for evaluating the electricalpotential produced by muscle cells EMG in combi-nation with other medical sensors is used in FPPSs[23]

(iii) Blood pressure sensorBlood pressure sensing is a physiological sign that canbe investigated to determine abnormalities that maylead to a fall [23]

(iv) Galvanic Skin Response (GSR) sensorGSR is a method for measuring the electrical charac-teristics of the skin GSR in combination with ECGand EMG are used to predict and prevent falls [23]

Fall factors in physiological analysis are as follows

(i) Neuromuscular noiseThe increased neuronal noise associated with agingincreases gait variability and consequently fall risk[25]

(ii) HeartbeatAn irregular heartbeat increases the risk of a fall [2326]

(iii) Skin electrical characteristicSince the sweat glands are controlled by the sympa-thetic nervous system which controls also emotionsa variation of the skin electrical characteristic coulddemonstrate a state of stress which indicates the riskof a fall [23]

Unlike physiological solutions kinematics-based FPPSsconsider userrsquos posture or gait variables These solutionsusually exploit movement sensors to investigate the extrinsicparameters of fall ie characteristics of the movement of thebody

(i) AccelerometerAn accelerometer is a device that measures accelera-tion ie the rate of change of the velocity of an object

(ii) GyroscopeA gyroscope gives the angular rate around one ormore axes of the space Angular measurement aroundlateral longitudinal and vertical plane are referredto as pitch roll and yaw respectively Typically inFPPSs the gyroscope is used in combination with anaccelerometer

(iii) Motion sensorA motion sensor detects the movement of an objectin the environment

(iv) Piezoelectric sensorA piezoelectric sensor measures the variations inpressure and force using the piezoelectric effect andconverts them into an electrical charge

Kinematic-based FPPSs focus on future or real-time falloccurrence Future fall solutions evaluate a user to estimatehisher fall risk if the fall risk is high a probable future fall canbe prevented through some exercises [27] In contrast real-time fall solutions avoid a fall while the user is doing hisherdaily activity by alerting the user [14ndash16] or using an externalaid such as a walker or robot [17 18]

As extrinsic fall factors are among the most commoncauses of fall this study surveys kinematic-based FPPSsconsidering in particular data acquired with gyroscope andaccelerometer sensorsThemain kinds of factors in kinematicanalysis which can increase the risk of fall in FPPSs areexplained in the following

21 User Profile User profile can affect the fall risk For exam-ple the risk of a fall for elderly people is higher thanyoung people and the risk of a fall is higher in people whoexperienced a previous fall Fall risk can be assessed through aweighted generic formula that combines all these factors [28]

22 Velocity People with increased fall risk tend to walkslowly As such the actual fall risk can be quantified accordingto gait speed [29] Gait speed is estimated by measuringduration and length of userrsquos steps The step duration iscalculated as the time between two consecutive foot contactsThe step length is calculated as the sum of the displacementduring the swing phase and the stance phase

23 Acceleration Changing of body movement in a prefallstate causes alternation in the acceleration so by processingthe Acceleration Time Series (ATS) a fall event can bepredicted As Figure 2 illustrates human motion during thetime period 119878 can be presented with 119899 smaller periods 119879119904[30] Period 119879119904 itself consists of 119898 short periods 119879 with 119898acceleration samples Basically ATS is characterized by aseries of elements 119888119894 over time where each element describesthe feature of the movement during period 11987911990424 Tilt Tilt is inclination from horizontal or vertical lineWhen a user significantly tilts in a direction it shows anabnormal posture which can lead to a fall So the user tiltcan be a factor to assess the risk of a fall Table 1 illustrates thenotations to measure userrsquos tilt [14ndash16]

Some traditional standard balance tests such as sit tostand (STS) uses trunk tilt to evaluate the risk of a fall Thetrunk tilt is calculated based on the angles between the sensorand the horizontal line of the ground [31]

25 Postural Transition Duration Posture specifies the posi-tion of the body Any activity begins with a posture and endswith another posture Postural transition duration specifiesthe duration of a transition from a posture to another oneBalance control and stability of the body during postural tran-sitions are key factors for avoiding falls Postural transitionduration can be an indicator of fall because it is significantlycorrelated with the fall risk [31] Higher transition durationmeans lower muscle strength and consequently higher fallriskThe duration of the postural transition can be computedby means of the accelerometer and gyroscope by measuring


Time (ms)

S

４Ｍ lowast Ｈ４Ｍ = Ｇ lowast ４

＝1 ＝Ｃ＝Ｈ

Acce

lera

tion

(ms

2)

Figure 2 ATS accelerometer

Table 1 Results of different approaches with decision tree and support vector machine classifications

Measures Tilt Speed AccelerationDT SVM DT SVM DT SVM

Accuracy 8388 657 7211 6153 8142 5666Error Rate 1611 342 2788 3846 1857 4333Sensitivity 088 074 087 066 078 026Generality 021 042 043 043 015 012Precision 081 063 067 060 083 068Recall 088 074 087 066 078 026ROC Area 086 065 069 061 082 057

the depression on the vertical axis acceleration signal andthe positive and negative angular rotations of the horizontalaxisThe depression of the acceleration signal comes from themovement of the body and the angular rotation is due to theforward and backward leans of the trunk during transition

26 Foot State The foot state indicates the posture of thefoot during the gait cycle Since foot state can specify thebalance of the user investigating the foot state can help todefine the prefall state and estimate the fall risk Some FPPSsfocus on features related to foot state such as foot clearance(ie the distance of the foot and ground during walking)and foot age (ie relation of the age with the foot pressure)[32 33]

27 Stability and Symmetry Stability means the resistance ofstanding against a position change Symmetry is the balanceof the pressure on two feet Stability and symmetry affect thefunctionality of the user gait a gait with weak stability andsymmetry has a higher fall risk According to the stability andsymmetry of gait an assessment model can be used to predictthe fall risk [34]

3 Feature Extraction

After acquiring signals from sensors a feature extractiontechnique should be applied to extract appropriate informa-tion Since data collected from sensors contain undesiredinformation filtering techniques are essential A filtering


Mean Mean

Energy

SMA

SVM

Derivative

HjorthParametersPeak-to-Peak

Energy

Minimum Foot Clearance

HjorthParameters

Mean

StandardDeviation Standard

Deviation

Maximum

Fall Factor

Velocity Acceleration Tilt Phase Transition Foot State StabilitySymmetryGeneric

Mean

StandardDeviation

Step length

Sole pressure

Single support time

Double support time

Age

Gender

Previous fall

Figure 3 Fall factors approaches

technique eliminates some frequencies from the originalsignal to attenuate the backgroundnoise and to remove unde-sired frequencies [29 35] Frequently used filters in FPPSsare high-pass filters which eliminate frequencies lower thanthe cutoff frequency and low-pass filters which pass onlyfrequencies lower than a certain threshold frequency Afterfiltering the collected data appropriate features should beselected Since analyzing a high number of features requires alarge amount of memory finding the optimal feature set canimprove the performance of the system The main featuresextracted from each fall factor are listed in Figure 3 anddescribed in the following

31 User Profile Falls are the result of a combination of fac-tors involving age sex mobility daily activity cognition andprevious fall Thus fall risk can be expressed as a simplefunction of user profile features with appropriate weights[28]

119865119886119897119897 119877119894119904119896 = 013 (119868119886) + 015 (119868119904) + 014 (119868119898)+ 01 (119868119886119889119897) + 018 (119868119888) + 033 (119868119891)

(1)

where

(i) 119868119886 is the age index the risk of falls in the elderly isassumed increasing with age [36]

(ii) 119868119904 is the sex index female gender is associated withgreater risks of fall [37]

(iii) 119868119898 is the mobility index mobility implies the abilityto move from place to place which can be indicator ofa fall [38 39]

(iv) 119868119886119889119897 is the index derived from the activities of dailyliving (ADL) fall risk and a personrsquos perception ofcapabilities within a particular domain of activitieshave strong independent correlation with ADL [40]

(v) 119868119888 is the cognition index impaired cognition anddementia independently predict falls [41]

(vi) 119868119891 is the previous fall index a history of previous fallshas been recognized as being a significant risk factorfor future falls [41]

Theweights in the above formula and the choice of indicesare made by statistical result of the earlier study [28] Theweights differ for male and female in the above formulaweights are calculated for female gender

32 Velocity As mentioned in Section 22 low gait speedincreases the risk of a fall The average speed of the gait canbe measured to estimate the fall risk [29]

33 Acceleration Frequently used features of the accelerationare described in the following In the formulas 119860(119905) refersto the acceleration and 119860119909(119905) 119860119910(119905) and 119860119911(119905) indicate thecomponents of the acceleration in the 3 axes Moreover119860119909119894 119860119910119894 and 119860119911119894 are the 119894-th acceleration samples in the 3axes

The Signal Magnitude Area (SMA) can be used as afeature of the acceleration signal to classify the activities ofthe user [35] SMA is computed as follows

119878119872119860 = 1119879 (int1198790

1003816100381610038161003816119860119909 (119905)1003816100381610038161003816 119889119905 + int1198790

10038161003816100381610038161003816119860119910 (119905)10038161003816100381610038161003816 119889119905

+ int1198790

1003816100381610038161003816119860119911 (119905)1003816100381610038161003816 119889119905)(2)

where 119879 is the length of measurement timeThe Signal Magnitude Vector (SMV) is one of the com-

mon measures to calculate the resultant of the signal

119878119872119881 = 1119899119899sum119894=1

radic1198602119909119894 + 1198602119910119894 + 1198602119911119894 (3)


SMV demonstrates the degree of the movement intensity andit is an essential metric in FPPSs [15 16 30 35]

Moreover the derivative (1198601015840(119905)) of the acceleration indi-cates the vibration of the movement and can be used as anacceleration feature [35]

Hjorth parameters are statistical features of the signal intimedomain [42]They are based on the variance of the signalV119886119903(119860(119905))

(i) 119867119895119900119903119905ℎ 119886119888119905119894V119894119905119910 = V119886119903(119860(119905)) it can indicate thesignal power

(ii) 119867119895119900119903119905ℎ 119898119900119887119894119897119894119905119910 = radicV119886119903(1198601015840(119905))V119886119903(119860(119905)) it can bean indicator of the smoothness of the signal curve

(iii) 119867j119900119903119905ℎ 119888119900119898119901119897119890119909119894119905119910 =119898119900119887119894119897119894119905119910(1198601015840(119905))119898119900119887119894119897119894119905119910(119860(119905)) it can effectivelymeasure irregularities in the frequency domain

The Hjorth parameters are mostly used to analyze theelectroencephalography signals but they are also utilized toanalyze accelerometer and gyroscope signals in FPPSs [15]

Peak is the absolute maximum value of the signal over theperiod of time and peak-to-peak is the difference betweenthe minimum and the maximum value of the signal over theperiod of time The peak-to-peak acceleration amplitude andthe peak-to-peak acceleration derivative are two features usedin FPPSs [29]

The energy of the acceleration signal describes theamount of physical activity in the vertical and horizontaldirections It can determine the strength of the contact withthe floor so it can be used to recognize abnormal walkingpattern such as stumbling [14 16] The energy of the signalcan be computed as

119864119909 = intinfinminusinfin

|119860 (119905)|2 119889119905 (4)

As described in Section 23 ATS can be characterizedby a series of features 119888119894 Feature 119888119894 can be determined bycalculating the resultant acceleration 997888997888rarr119860119865

997888997888rarr119860119865 = radic997888997888rarr|119909|2 + 997888997888rarr100381610038161003816100381611991010038161003816100381610038162 +997888997888rarr|119911|2 (5)

The resultant acceleration (997888997888rarr119860119865) varies within a small range B= [1198871 1198872] around 119892 where 119892 is the gravity force 1198871 lt 119892 lt 1198872Therefore if the resultant acceleration exceedsB an abnormalwalking is probable

34 Tilt Trunk tilt has an important role in the maintenanceof the posture The average and standard deviation of trunktilt are measured during the sit to stand phase of STS testto assess the risk of a fall [31] Moreover energy and Hjorthparameters of the tilt vector are used as an indicator of theabnormal motion [14ndash16]

35 Postural Transition Duration The average and standarddeviation of the duration of the postural transition can beused to estimate the userrsquos fall risk [27 31]

36 Foot State Step length is a feature of the foot in a gaitcycle that can be a good indicator of the fall risk Since ahigh step length decreases the stability of the user the fallrisk increases as the length of the step grows Single supporttime is the time when only one limb is on the ground in a gaitcycle Double support time is the time spent when both feetare on the ground in a gait cycleThe foot age is an index of thegait which shows how old is the gait condition of the subjectThrough the foot age the falling risk can be quantifiedThe foot age is computed through the four gait features(step length step center of sole pressure (CSP) distance ofsingle supporting period and time of double support period)[32]

The Minimum Foot Clearance (MFC) is another footstate feature that indicates the vertical distance between thelowest point of the foot of the swing leg and the walkingsurface during the swing phase of the gait cycle The footclearance is an important gait parameter that is related to therisk of falling The low foot clearance for a step during thewalking increases the probability of fall because of hitting toan obstacleThe foot clearance is extensively studied to detecttrips and falls [33 43 44]

4 Machine Learning Algorithm

Features extracted from the input signals are processed by amachine learning algorithm in order to classify the abnormalbehavior and the normal daily activity Exploited machinelearning algorithms in FPPSs are described in following

(1) Threshold-based Algorithm utilizes a threshold toclassify the feature set of the user gait After extracting thedesired features from the input signals these features arecompared with predefined thresholds Since the thresholdshave an important effect on the performance of the algo-rithm the biggest challenge of a threshold-based algorithmis determining the thresholds Moreover the performance ofthe algorithm depends on the number of features which needto be analyzed However complexity and power consumptionof this type of algorithms are low so it can be adequate fordevices with limited resources Some examples of differentFPPSs that exploit a threshold-based algorithm are describedin the following

Features of the acceleration signal of human upper trunk[30] in a short time interval before the fall are denoted as120582 After obtaining the ATS of the user P(ATS |120582) states theprobability of a fall occurrence during the user motion Twothresholds P1 and P2 are specified to predict and detect a fallAs Figure 4 illustrates the output of P(ATS |120582) is an input tothe algorithm Then based on predefined thresholds if P ishigher than P1 the fall risk is notified if P is higher than P2a possible fall is noticed

SMV SMA peak-to-peak and derivative of accelerationsignals are computed as feature set of user gait [35] After-wards thresholds are determined to define a near fall state

The gait status can be classified based on mean andstandard deviation of stability and symmetry [34]

(i) If index le mean+std then the gait status is normalthe subject walks normally and there is not fall risk


P (ATS|)

PltP1

PltP2Fall predicted

No fallYES

NO

YES

Fall detected

NO

Figure 4 Threshold-based algorithm

(ii) If mean+std lt index le 3lowastmean then the gait status isattentive the subject needs to care when walking

(iii) If index gt 3lowastmean then the gait status is dangerousthe subject should present a risk to fall

(2) Decision Tree (DT) is a directed tree with a rootnode without incoming edges and all other nodes knownas decision nodes with one incoming edge and possibleoutcoming edges a leaf node is a node without outcomingedges At the training stage each internal node splits theinstance space into two or more parts After that every pathfrom the root node to a leaf node forms a decision rule todetermine which class a new instance belongs to [45] Eachinternal node represents a test on an attribute or on a subsetof attributes and each edge is labeled with a specific value orrange of values of the input attributes DT is a fast algorithmbut the computation cost on the tree grows as the size of thetree increases

Figure 5 illustrates how a decision tree algorithm can beused in the classification of normal and abnormal walking[14ndash16] Firstly accelerometer and gyroscope signals arecollected then a general tilt vector is computed Afterwardsappropriate features (eg energy Hjorth parameters) arecalculated from tilt vector Then DT is used to determine theabnormal walking

(3) Support Vector Machine (SVM) finds the best hyper-plane with the maximum margin to separate two classesSVM can be defined as linear or nonlinear according to thekind of hyperplane function SVM is a prevailing classifica-tion model for gait pattern recognizing [46ndash48] SVM can beused to determine the threshold value to classify the user gait[30]

(4) Fuzzy Logic defines amembership function in order toassign to objects a grade ofmembership ranging between zeroand one For example if119883 is a class of objects with a genericelement denoted by 119909 a fuzzy set 119860 in 119883 is characterized bya membership function 119891119860(119909) The value of 119891119860(119909) representsthe ldquograde of membershiprdquo of 119909 in 119860 which is a real number

in the interval [0 1] The nearer the value of 119891119860(119909) to unitythe higher the grade of membership of 119909 is in 119860 [49]

Based on the relationship between fall risk and age thefuzzy logic is used to prevent a fall using the sole pressuresensor to estimate the age [32] Firstly the fuzzy membershipfunction for young age 120583119884 middle age 120583119872 and elderly age120583119864 are calculated based on the four extracted features (steplength step center of sole pressure width distance of singlesupporting period and time of double support period) of theuser gait Then a fuzzy logic is used to estimate the foot age

5 Evaluation Criteria

Evaluation criteria of a machine learning algorithm aredescribed in the following In all presented formulas Pand N represent the total number of positive and negativeinstances A positive and negative instance can be definedas an abnormalnormal walk True Positive (TP) and TrueNegative (TN) are defined as correct identification of a trueclassification of positive and negative instance respectivelyFalse Positive (FP) and False Negative (FN) misidentifypositive and negative instances respectively

(1) Specificity or True Negative Rate (TNR) measuresthe rate of negative instances that are correctly identified asnegative

119878119901119890119888119894119891119894119888119894119905119910 = 119879119873119879119873 + 119865119875 (6)

Moreover 119866119890119899119890119903119886119897119894119905119910 is computed as 1 - Specificity(2) Sensitivity or True Positive Rate (TPR) measures therate of positive instances that are correctly identified aspositive

119878119890119899119904119894119905119894V119894119905119910 = 119879119875119879119875 + 119865119873 (7)

(3) Accuracy of an algorithm computes the number ofsamples correctly classified

119860119888119888119906119903119886119888119910 = 119879119875 + 119879119873119875 + 119873 (8)

(4) Error rate is the number of wrong classifications

119864119903119903119900119903 119877119886119905119890 = 119865119875 + 119865119873119875 + 119873 (9)

(5) Precision is the percentage of the samples correctlyclassified as true

119875119903119890119888119894119904119894119900119899 = 119879119875119879119875 + 119865119875 (10)

(6) Recall is the percentage of truly classified positivesamples

119877119890119888119886119897119897 = 119879119875119879119875 + 119865119873 (11)

It should be noted that one criterion alone may not besufficient to evaluate the algorithm


COLLECT DATA

TRAINING

ABNORMAL

CLASSIFY WITH DECISION TREEUPDATE TRAINING SET

YES NO

YES

WARN USER

NO

EXTRACT FEATURE

Figure 5 Fall prediction and prevention algorithm based on decision tree

00

01

02

03

04

05

06

07

08

09

10

01 02 03 04 05 06 07 08 09 10

more erroneous

less erroneous

A

CB

Generality

Sens

itivi

ty D

Figure 6 ROC plot of classifiers

The Receiver Operating Characteristic (ROC) curve canbe used to compare different algorithms The ROC curve is agraphical plot that illustrates the performance of a classifier[50] Generality and sensitivity are plotted on 119909 and 119910 axesof the ROC plot respectively The best classifier is locatedat the top left corner of the ROC graph which represents100 sensitivity and 100 specificity The diagonal line fromthe left bottom to the top right corner divides the ROCspace into two parts The space above the diagonal representsclassification with few errors while the space below theline shows more erroneous results For example Figure 6compares four possible algorithms A(0109) B(01022)C(09015) and D(0806) 119860 has the best prediction amongthe four instances The further the result is from the diagonal

in the above space the better the accuracy is 119862 is the worstamong the four instances because it is below and far fromthe diagonal line 119861 is a good classifier but not as muchas 119860 because it is above but not far from the diagonalline Moreover since 119863 is closer to diagonal line it is moreacceptable than 119862

To show how ROC curve is plotted the output of twoclassifiers is illustrated in Figure 7 The 119909-axis shows theprobability that the user gait is abnormal and the 119910-axisrepresents the number of instances with the same probabilityFor instance point (08 5) means that the gait of 5 users ispredicted as abnormal with probability 08 and the gait of allusers is abnormal because it is located along the abnormaldistribution To prepare the ROC curve firstly a randomvariable 119883 is defined and a threshold (119879) is set Everythingabove the threshold (119883 gt 119879) is classified as abnormal andbelow the threshold (119883 lt 119879) as normal Then sensitivityand generality of this classification with threshold 119879 arecomputed To generate the ROC curve the sensitivity versusthe generality for all possible thresholds should be plottedFigure 7(a) shows a classifier with its associated ROC curveSince the distributions of normal and abnormal cases barelyoverlap the corresponding ROC curve of the classifier is closeto the upper left corner of the plot Figure 7(b) shows anotherclassifier where the distribution of normal and abnormalcases overlap almost completely so the ROC curve of theclassifier is close to diagonal line

6 Experimental Results

In this section the state-of-the-art FPPSs with accelerometerand gyroscope have been implemented and then comparedaccording to the criteria described in Section 5

First objective of the experiment is empirical comparisonof fall factors to find the most representative one among


Generality

1

0 Probability

AbnormalNormal

11 008

5

Sens

itivi

ty

ob

serv

atio

ns

(a)

Generality

1

ProbabilityProbability

AbnormalNormal

10 1

Sens

itivi

ty

ob

serv

atio

ns

(b)Figure 7 ROC of a good classifier (a) and poor classifier (b)

acceleration tilt and velocityThe second objective is evaluat-ing different machine learning algorithm based on presentedfeatures for each fall factor in the presented dataset It shouldbe noted that the goal of the experiment is not finding anoptimal feature set for each fall factor

A fall occurs due to passive causes like weakness balancedeficit gait deficit visual deficit and mobility limitation Thefollowing are the most frequently used methods to simulatean abnormal gait which can lead to a fall

(i) Walking with straightened knee [14ndash16](ii) Walking with leg length discrepancy [14ndash16](iii) Walking on a rough surface [51](iv) Walking through obstacles [35]

In the experiments an abnormal walk is modeled asirregular gaits obtained by walking through obstacles whichcan cause stepping tripping and stumbling Thus a flatarea with different types of obstacles is prepared Obstaclesincluded empty boxes (height 37 cm length 20 cm andwidth 17 cm) and plastic bottles (height 20 cm diameter 6cm) which are placed 60 cm far from each other

Since the real falls cannot be experimented due to therisk of injury only forward fall is simulated with protection

However the applied method in this paper can easily begeneralized to other type of falls (ie backward lateral) Usersare asked to walk through obstacles without looking themfor 10 seconds The users in the experiments are 19 menwith weight in the range of 65-110 kg and height in therange of 160-185 cm and 3 women with weight in the rangeof 50-60 kg and height in the range of 157-165 cm Allusers are without gait disturbances Furthermore users arein the range of 18-35 years old Data is collected throughMATLAB R2015b WEKA tool version 3613 (WEKA isan open source data mining tool that can be downloadedfrom httpwwwcswaikatoacnz mlweka) has been usedto classify the obtained data

An iPhone 4S is adopted in the experiments equippedwith the STMicro STM33DH 3-axis accelerometer and theSTIMicro AGDI 3-axis gyroscope Commonly adopted sam-pling frequencies range from some dozens to hundred ofHertz such that they are constant and higher than gait cyclefrequency In the experiment the frequency is fixed to 10HzSince the body Center of Pressure (COP) reveals severalinformation of user gait the smartphone is placed on thelower back of trunk near the real Center Of Mass (COM)position assuming that this position moves parallel to theCOP and the same acceleration and positions will be mea-sured [52]


In the following velocity acceleration and tilt fall factorswith different combinations of machine learning algorithmshave been evaluated

SMA SVM maximum derivative Hjorth parameterspeak-to-peak and energy features of acceleration are con-sidered in the experiments Moreover mean standard devi-ation energy and Hjorth parameters of tilt and mean ofvelocity are considered as the features in the experiment Theperformance of different features with a particular machinelearning algorithm was already presented in previous studiesThe novelty of this paper is the comparison of different fallfactors with presented features on different machine learningalgorithm

Table 1 shows the result of the experiments when thedecision tree and support vector machine are selected toclassify the obtained data As reported in the last line ofTable 1 a higher ROC area corresponds to a better accuracyThe comparison of the results fromdifferent fall factors showsthat the tilt always has better accuracy among the other fallfactors Combination of tilt with decision tree gives 839 ofaccuracy and with support vector machines gives 657 ofaccuracy So the preciseness of tilt factor shows that it can beadopted as a deserved representative of fall factors in FPPSsimplemented with personal monitoring device

There are several factors that can affect the decision tochoose a machine learning algorithm In literature there areseveral studies which compare the performance of differentmachine learning algorithms [53ndash55] The best machinelearning algorithm cannot be universally identified becausemachine learning algorithms are task-dependent In additionthe best machine learning algorithms for a particular taskdepends on several factorsThe feature set is the primer factorwhich impacts on the performance In addition also datasetcharacteristic such as number of samples type and kind ofdata and skewed data can impact on the performance In askewed dataset almost all samples fall in one particular classrather than in the other classes

The comparison of the different machine learning algo-rithms with the presented setting in this paper shows that DThas better performance than SVM in all the combinationsAlthough DT has better performance comparing to the SVMit cannot be generalized to all experiments and datasets Thereason is based on the no free lunch theorems that indicatesthere is not superiority for any machine learning algorithmover the others so the best classifier for a particular task istask-dependent [56 57] However it should be noted thatDT requires more memory space when the size of the treegrows Moreover as Figure 8 shows the tilt fall factor withDT algorithm has the best performance to detect abnormalwalks and speed has the lower performance Since in theexperiment patients tilt in a direction it is not surprising thattilt is the most representative fall factor

7 Conclusion

This paper analyzed different aspects of fall prediction andprevention systems It provides a comprehensive overviewof various fall factors and corresponding features Moreover

0 02 04 06 08 10

010203040506070809

1

Generality

Sens

itivi

ty

Tilt-DTAcceleration-DTSpeed-DT

Tilt-SVMAcceleration-SVMSpeed-SVM

Figure 8 ROC on different approaches of 22 userswith 110 samples

different machine learning algorithms in fall predictionand prevention systems have been reviewed Furthermoremultiple combinations of features and fall prediction andprevention algorithms have been experimentally evaluatedto find an optimal solution Based on the presented resultstilt features in combination with the decision tree algorithmpresent the best performance among the other permutationsof fall factors and fall prediction and prevention algorithms

Future work may include

(i) generalizing the dataset to older adults or patientswith neurological disorders

(ii) adopting new machine learning algorithms in thecomparison list

(iii) comparing systems across other performance metricssuch as time accuracy with respect to the size ofdataset number of features memory consumptionand power consumption

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this article

Acknowledgments

This work was partially supported by the grant ldquoBando SmartCities and Communitiesrdquo OPLON project (OPportunitiesfor active and healthy LONgevity) funded by the ItalianMinistry for University

References

[1] R Rajagopalan I Litvan and T-P Jung ldquoFall predictionand prevention systems Recent trends challenges and futureresearch directionsrdquo Sensors vol 17 no 11 2017

[2] V R L Shen H-Y Lai and A-F Lai ldquoThe implementation of asmartphone-based fall detection systemusing a high-level fuzzyPetri netrdquo Applied Soft Computing vol 26 pp 390ndash400 2015


[3] Y Li K C Ho and M Popescu ldquoA microphone array systemfor automatic fall detectionrdquo IEEE Transactions on BiomedicalEngineering vol 59 no 5 pp 1291ndash1301 2012

[4] Y Zigel D Litvak and I Gannot ldquoA method for automaticfall detection of elderly people using floor vibrations andsoundProof of concept on human mimicking doll fallsrdquo IEEETransactions on Biomedical Engineering vol 56 no 12 pp2858ndash2867 2009

[5] Z-P Bian J Hou L-P Chau andNMagnenat-Thalmann ldquoFalldetection based on body part tracking using a depth camerardquoIEEE Journal of Biomedical and Health Informatics vol 19 no2 pp 430ndash439 2015

[6] YWang KWu and LMNi ldquoWiFall device-free fall detectionby wireless networksrdquo IEEE Transactions on Mobile Computingvol 16 no 2 pp 581ndash594 2017

[7] H Wang D Zhang Y Wang J Ma Y Wang and S LildquoRT-Fall a real-time and contactless fall detection systemwith commodity WiFi devicesrdquo IEEE Transactions on MobileComputing vol 16 no 2 pp 511ndash526 2017

[8] K Zhao K Jia and P Liu ldquoFall detection algorithm based onhuman posture recognitionrdquo in Proceedings of the InternationalConference on Intelligent Information Hiding and MultimediaSignal Processing pp 119ndash126 2016

[9] K H Frith A N Hunter S S Coffey and Z Khan ldquoAlongitudinal fall prevention study for older adultsrdquoThe Journalfor Nurse Practitioners vol 15 no 4 pp 295ndash300e1 2019

[10] K Khanuja J Joki G Bachmann and S Cuccurullo ldquoGaitand balance in the aging population Fall prevention usinginnovation and technologyrdquoMaturitas vol 110 pp 51ndash56 2018

[11] M Hemmatpour R Ferrero F Gandino B Montrucchioand M Rebaudengo ldquoNonlinear predictive threshold modelfor real-time abnormal gait detectionrdquo Journal of HealthcareEngineering vol 2018 Article ID 4750104 9 pages 2018

[12] M Hemmatpour R Ferrero B Montrucchio and M Rebau-dengo ldquoA neural network model based on co-occurrencematrix for fall predictionrdquo in Proceedings of the InternationalConference on Wireless Mobile Communication and Healthcarepp 241ndash248 2016

[13] M Hemmatpour M Karimshoushtari R Ferrero B Montruc-chioM Rebaudengo andCNovara ldquoPolynomial classificationmodel for real-time fall prediction systemrdquo in Proceedings of theComputer Software and Applications Conference vol 1 pp 973ndash978 July 2017

[14] A J A Majumder I Zerin S I Ahamed and R O Smith ldquoAmulti-sensor approach for fall risk prediction and prevention inelderlyrdquoACM SIGAPP Applied Computing Review vol 14 no 1pp 41ndash52 2016

[15] A J Majumder I Zerin M Uddin S I Ahamed and R OSmith ldquoSmartprediction A real-time smartphone-based fallrisk prediction and prevention systemrdquo in Proceedings of theResearch in Adaptive and Convergent Systems pp 434ndash439ACM October 2013

[16] A J A Majumder F Rahman I Zerin W Ebel Jr and S IAhamed ldquoiPrevention Towards a novel real-time smartphone-based fall prevention systemrdquo in Proceedings of the 28th AnnualACM Symposium on Applied Computing pp 513ndash518 ACMMarch 2013

[17] P Di J Huang K Sekiyama and T Fukuda ldquoA novel fallprevention scheme for intelligent cane robot by using a motordriven universal jointrdquo in Proceedings of the InternationalSymposium on Micro-NanoMechatronics and Human Science(MHS rsquo11) pp 391ndash396 IEEE 2011

[18] Y Hirata A Muraki and K Kosuge ldquoMotion control of intel-ligent passive-type walker for fall-prevention function basedon estimation of user staterdquo in Proceedings of the 2006 IEEEInternational Conference on Robotics and Automation (ICRArsquo06) pp 3498ndash3503 May 2006

[19] A Shumway-Cook S Brauer and M Woollacott ldquoPredictingthe probability for falls in community-dwelling older adultsusing the timed up amp go testrdquo Physical Therapy in Sport vol 80no 9 pp 896ndash903 2000

[20] SWMuir K Berg B Chesworth andM Speechley ldquoUse of theberg balance scale for predicting multiple falls in community-dwelling elderly people a prospective studyrdquo Physical Therapyin Sport vol 88 no 4 pp 449ndash459 2008

[21] B Najafi K Aminian F Loew Y Blanc and P Robert ldquoMea-surement of stand-sit and sit-stand transitions using aminiaturegyroscope and its application in fall risk evaluation in theelderlyrdquo IEEE Transactions on Biomedical Engineering vol 49no 8 pp 843ndash851 2002

[22] B J Vellas S J Wayne L Romero R N Baumgartner L ZRubenstein and P J Garry ldquoOne-leg balance is an importantpredictor of injurious falls in older personsrdquo Journal of theAmerican Geriatrics Society vol 45 no 6 pp 735ndash738 1997

[23] E T Horta I C Lopes J J Rodrigues and S Misra ldquoReal timefalls prevention and detection with biofeedback monitoringsolution for mobile environmentsrdquo in Proceedings of the IEEE15th International Conference on e-Health Networking Applica-tions and Services (Healthcom rsquo13) pp 594ndash600 IEEE LisbonPortugal October 2013

[24] L Z Rubenstein ldquoFalls in older people epidemiology riskfactors and strategies for preventionrdquo Age and Ageing vol 35supplement 2 pp ii37ndashii41 2006

[25] P E Roos and J B Dingwell ldquoInfluence of simulated neuro-muscular noise on movement variability and fall risk in a 3Ddynamic walking modelrdquo Journal of Biomechanics vol 43 no15 pp 2929ndash2935 2010

[26] Md Shahiduzzaman ldquoFall detection by accelerometer andheart rate variability measurementrdquoGlobal Journal of ComputerScience and Technology vol 15 no 3 2016

[27] B N Ferreira V Guimaraes and H S Ferreira ldquoSmartphonebased fall prevention exercisesrdquo in Proceedings of the 15thInternational Conference on e-Health Networking Applicationsamp Services (Healthcom 11990513) pp 643ndash647 IEEE 2013

[28] K Cameron K Hughes and K Doughty ldquoReducing fallincidence in community elders by telecare using predictivesystemsrdquo in Proceedings of the 19th Annual International Con-ference of the IEEE Engineering in Medicine and Biology SocietyrsquoMagnificent Milestones and Emerging Opportunities in MedicalEngineeringrsquo vol 3 pp 1036ndash1039 IEEE Chicago ILUSA 1997

[29] V Guimaraes P M Teixeira M P Monteiro and D EliasldquoPhone based fall risk predictionrdquo in Wireless Mobile Commu-nication and Healthcare pp 135ndash142 Springer 2012

[30] L Tong Q Song Y Ge and M Liu ldquoHMM-based human falldetection and predictionmethod using tri-axial accelerometerrdquoIEEE Sensors Journal vol 13 no 5 pp 1849ndash1856 2013

[31] V Guimaraes D Ribeiro and L Rosado ldquoA smartphone-basedfall risk assessment tool measuring one leg standing sit tostand and falls efficacy scalerdquo in Proceedings of the IEEE 15thInternational Conference on e-Health Networking Applicationsand Services (Healthcom rsquo13) pp 529ndash533 IEEE October 2013

[32] T Takeda Y Sakai K Kuramoto S Kobashi T Ishikawa andY Hata ldquoFoot age estimation for fall-prevention using sole


pressure by fuzzy logicrdquo in Proceedings of the IEEE InternationalConference on Systems Man and Cybernetics (SMC rsquo11) pp 769ndash774 Anchorage AK USA October 2011

[33] R Begg R Best L DellrsquoOro and S Taylor ldquoMinimum footclearance during walking strategies for the minimisation oftrip-related fallsrdquoGait ampPosture vol 25 no 2 pp 191ndash198 2007

[34] S Jiang B Zhang and D Wei ldquoThe elderly fall risk assessmentandprediction based on gait analysisrdquo inProceedings of the IEEE11th International Conference on Computer and InformationTechnology (CIT rsquo11) pp 176ndash180 IEEE Paphos Cyprus August2011

[35] AWeiss I ShimkinNGiladi and JMHausdorff ldquoAutomateddetection of near falls algorithm development and preliminaryresultsrdquo BMC Research Notes vol 3 no 1 p 362 2010

[36] M C Hornbrook V J Stevens D JWingfield J F Hollis M RGreenlick andMG Ory ldquoPreventing falls among community-dwelling older persons results from a randomized trialrdquo TheGerontologist vol 34 no 1 pp 16ndash23 1994

[37] M E Tinetti J Doucette E Claus and R Marottoli ldquoRiskfactors for serious injury during falls by older persons in thecommunityrdquo Journal of the American Geriatrics Society vol 43no 11 pp 1214ndash1221 1995

[38] B J Goldfarb and S R Simon ldquoGait patterns in patients withamyotrophic lateral sclerosisrdquoArchives of Physical Medicine andRehabilitation vol 65 no 2 pp 61ndash65 1984

[39] R W Bohannon ldquoCorrelation of lower limb strengths andother variables with standing performance in stroke patientsrdquoPhysiotherapy Canada vol 41 no 4 pp 198ndash202 1989

[40] M E Tinetti C FDe Leon J T Doucette andD I Baker ldquoFearof falling and fall-related efficacy in relationship to functioningamong community-living eldersrdquo Journal of Gerontology vol49 no 3 pp M140ndashM147 1994

[41] J Teno D P Kiel and V Mor ldquoMultiple stumbles a risk factorfor falls in community-dwelling elderly a prospective studyrdquoJournal of the American Geriatrics Society vol 38 no 12 pp1321ndash1325 1990

[42] B Hjorth ldquoEEG analysis based on time domain propertiesrdquoElectroencephalography and Clinical Neurophysiology vol 29no 3 pp 306ndash310 1970

[43] F Dadashi B Mariani S Rochat C J Bula B Santos-Eggi-mann and K Aminian ldquoGait and foot clearance parametersobtained using shoe-worn inertial sensors in a large-populationsample of older adultsrdquo Sensors vol 14 no 1 pp 443ndash457 2014

[44] D T Lai S B Taylor and R K Begg ldquoPrediction of footclearance parameters as a precursor to forecasting the risk oftripping and fallingrdquo Human Movement Science vol 31 no 2pp 271ndash283 2012

[45] WDai andW Ji ldquoAmapreduce implementation of c45 decisiontree algorithmrdquo Journal of DatabaseTheory andApplication vol7 no 1 pp 49ndash60 2014

[46] R K Begg M Palaniswami and B Owen ldquoSupport vectormachines for automated gait classificationrdquo IEEE Transactionson Biomedical Engineering vol 52 no 5 pp 828ndash838 2005

[47] J Yoo D Hwang and M S Nixon ldquoGender classificationin human gait using support vector machinerdquo in AdvancedConcepts for Intelligent Vision Systems pp 138ndash145 Springer2005

[48] D Lai P Levinger R Begg W Gilleard and M PalaniswamildquoAutomatic recognition of gait patterns exhibiting patellofemo-ral pain syndrome using a support vector machine approachrdquoIEEE Transactions on Information Technology in Biomedicinevol 13 no 5 pp 810ndash817 2009

[49] L A Zadeh ldquoFuzzy sets fuzzy logic and fuzzy systems selectedpapersrdquoWorld Scientific vol 6 1996

[50] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006

[51] M J D Otis and B A J Menelas ldquoToward an augmented shoefor preventing falls related to physical conditions of the soilrdquoin Proceedings of the IEEE International Conference on SystemsMan and Cybernetics (SMC rsquo12) pp 3281ndash3285 IEEE October2012

[52] V Guimaraes D Ribeiro and L Rosado ldquoA smartphone-basedfall risk assessment tool measuring one leg standing sit tostand and falls efficacy scalerdquo in Proceedings of the 15th IEEEInternational Conference on e-Health Networking Applicationsand Services (Healthcom rsquo13) pp 529ndash533 Portugal October2013

[53] R Caruana and A Niculescu-Mizil ldquoAn empirical comparisonof supervised learning algorithmsrdquo in Proceedings of the 23rdInternational Conference on Machine Learning (ICML rsquo06) pp161ndash168 June 2006

[54] N Zerrouki F Harrou A Houacine and Y Sun ldquoFall detectionusing supervised machine learning algorithms A comparativestudyrdquo in Proceedings of the 8th International Conference onModelling Identification and Control (ICMIC rsquo16) pp 665ndash670November 2016

[55] O Aziz M Musngi E J Park G Mori and S N Robi-novitch ldquoA comparison of accuracy of fall detection algorithms(threshold-based vs machine learning) using waist-mountedtri-axial accelerometer signals from a comprehensive set offalls and non-fall trialsrdquo Medical amp Biological Engineering ampComputing vol 55 no 1 pp 45ndash55 2017

[56] D H Wolpert ldquoThe lack of a priori distinctions betweenlearning algorithmsrdquoNeural Computation vol 8 no 7 pp 1341ndash1390 1996

[57] DHWolpert andWGMacready ldquoNo free lunch theorems foroptimizationrdquo IEEE Transactions on Evolutionary Computationvol 1 no 1 pp 67ndash82 1997

Computer Games Technology

International Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Journal ofEngineeringVolume 2018

Advances in

FuzzySystems


Volume 2018


ReconfigurableComputing



Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

thinspArtificial Intelligence

Hindawiwwwhindawicom Volumethinsp2018


Civil EngineeringAdvances in


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia


Biomedical Imaging



Engineering Mathematics


RoboticsJournal of



Computational Intelligence and Neuroscience


Mathematical Problems in Engineering

Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018


Human-ComputerInteraction

Advances in


Scientic Programming

Submit your manuscripts atwwwhindawicom


FALL PREDICTION AND PREVENTION

Physiology

Sensor

ECG

EMG

Blood Pressure

GSR(EDA)

Neuromuscular noise

Heartbeat

Skin electrical characteristic

Sensor

Accelerometer

Gyroscope

User Profile

Velocity

Tilt

Acceleration

Motion sensor

Piezoelectricsensor

Phase Transition

Stability and Symmetry

Kinematics

Fall Factor

Foot State

Fall Factor

Figure 1 Fall prediction and prevention taxonomy

detected then the user is alerted [14ndash16] or an external aidsuch as a walker or robot is exploited to prevent a probablefall [17 18]

Future fall prediction is estimated through some clinicalassessment tests Probable future falls are prevented throughimproving gait and mobility by some exercises [17 18] Thesetests often involve questionnaires or functional assessmentsof posture gait cognition and other risk factors Theseclinical tests are subjective and qualitative and typically usethreshold assessment scores to categorize people as fallers andnonfallers Typically these tests are timed up and go (TUG)[19] Berg Balance Scale (BBS) [20] sit to stand (STS) [21]and one leg stand (OLS) [22] to evaluate balance and lowerlimb strength

The design of a fall prediction and prevention systemfaces several significant challenges They need to be accuratereliable robust and cost-effective [1] In this paper a fallprediction and prevention system is described in three partsfall factors (ie fall symptoms) their features and machinelearning algorithms This paper investigates every mentionedstage and experimentally evaluates the various approachesThis paper does not present systems using sensors such ascamera and sound since they are prone to violate individualrsquosprivacy comparing to kinematic wearable sensors In thispaper accelerometer (for measuring the acceleration) andgyroscope (for measuring the angular rate around one ormore axes of the space) are considered for evaluation Thesesensors are chosen since they are easily accessible and do notdisturb the privacy of the userThemain contributions of thispaper are the following

(i) A comprehensive discussion on fall prediction andprevention systems

(ii) Preparing a dataset with realistic parameters to simu-late abnormal gait

(iii) Finding the most representative fall factors(iv) Evaluation of the fall factors based on the extracted

feature on commonly used machine learning algo-rithms

This paper is organized as follows In Section 2 existingFPPSs are classified according to the fall factorsThen in Sec-tions 3 and 4 feature extraction techniques and machinelearning algorithms are described Evaluation criteria areillustrated in Section 5 Afterwards experimental results areshown in Section 6 Finally conclusions are presented inSection 7

2 Classification of Fall Factors

Fall prediction and prevention is a multifaceted problemthat can be broadly categorized into two different domainsphysiology and kinematics as can be seen in Figure 1

Physiological solutions consider intrinsic fall factors ieparameters which mostly originate from the body Thesesolutions entail an in-depth medical evaluation of the riskfactors and exploit sensors related to body monitoring

(i) Electrocardiogram (ECG) sensorTypically ECG is used to assess the electrical andmuscular operations of the heart but ECG sensing can























Time (ms)

S


＝1 ＝Ｃ＝Ｈ

Acce

lera

tion

(ms

2)











Mean Mean

Energy

SMA

SVM

Derivative


Energy


HjorthParameters

Mean


Deviation

Maximum

Fall Factor


Mean

StandardDeviation

Step length

Sole pressure

Single support time

Double support time

Age

Gender

Previous fall




119865119886119897119897 119877119894119904119896 = 013 (119868119886) + 015 (119868119904) + 014 (119868119898)+ 01 (119868119886119889119897) + 018 (119868119888) + 033 (119868119891)

(1)

where











119878119872119860 = 1119879 (int1198790

1003816100381610038161003816119860119909 (119905)1003816100381610038161003816 119889119905 + int1198790

10038161003816100381610038161003816119860119910 (119905)10038161003816100381610038161003816 119889119905

+ int1198790

1003816100381610038161003816119860119911 (119905)1003816100381610038161003816 119889119905)(2)



119878119872119881 = 1119899119899sum119894=1

radic1198602119909119894 + 1198602119910119894 + 1198602119911119894 (3)












|119860 (119905)|2 119889119905 (4)
















P (ATS|)

PltP1

PltP2Fall predicted

No fallYES

NO

YES

Fall detected

NO













119878119901119890119888119894119891119894119888119894119905119910 = 119879119873119879119873 + 119865119875 (6)


119878119890119899119904119894119905119894V119894119905119910 = 119879119875119879119875 + 119865119873 (7)


119860119888119888119906119903119886119888119910 = 119879119875 + 119879119873119875 + 119873 (8)


119864119903119903119900119903 119877119886119905119890 = 119865119875 + 119865119873119875 + 119873 (9)


119875119903119890119888119894119904119894119900119899 = 119879119875119879119875 + 119865119875 (10)


119877119890119888119886119897119897 = 119879119875119879119875 + 119865119873 (11)



COLLECT DATA

TRAINING

ABNORMAL


YES NO

YES

WARN USER

NO

EXTRACT FEATURE


00

01

02

03

04

05

06

07

08

09

10

01 02 03 04 05 06 07 08 09 10

more erroneous

less erroneous

A

CB

Generality

Sens

itivi

ty D









Generality

1

0 Probability

AbnormalNormal

11 008

5

Sens

itivi

ty

ob

serv

atio

ns

(a)

Generality

1


AbnormalNormal

10 1

Sens

itivi

ty

ob

serv

atio

ns















7 Conclusion


0 02 04 06 08 10

010203040506070809

1

Generality

Sens

itivi

ty











Acknowledgments


References


































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in


























Time (ms)

S


＝1 ＝Ｃ＝Ｈ

Acce

lera

tion

(ms

2)











Mean Mean

Energy

SMA

SVM

Derivative


Energy


HjorthParameters

Mean


Deviation

Maximum

Fall Factor


Mean

StandardDeviation

Step length

Sole pressure

Single support time

Double support time

Age

Gender

Previous fall




119865119886119897119897 119877119894119904119896 = 013 (119868119886) + 015 (119868119904) + 014 (119868119898)+ 01 (119868119886119889119897) + 018 (119868119888) + 033 (119868119891)

(1)

where











119878119872119860 = 1119879 (int1198790

1003816100381610038161003816119860119909 (119905)1003816100381610038161003816 119889119905 + int1198790

10038161003816100381610038161003816119860119910 (119905)10038161003816100381610038161003816 119889119905

+ int1198790

1003816100381610038161003816119860119911 (119905)1003816100381610038161003816 119889119905)(2)



119878119872119881 = 1119899119899sum119894=1

radic1198602119909119894 + 1198602119910119894 + 1198602119911119894 (3)












|119860 (119905)|2 119889119905 (4)
















P (ATS|)

PltP1

PltP2Fall predicted

No fallYES

NO

YES

Fall detected

NO













119878119901119890119888119894119891119894119888119894119905119910 = 119879119873119879119873 + 119865119875 (6)


119878119890119899119904119894119905119894V119894119905119910 = 119879119875119879119875 + 119865119873 (7)


119860119888119888119906119903119886119888119910 = 119879119875 + 119879119873119875 + 119873 (8)


119864119903119903119900119903 119877119886119905119890 = 119865119875 + 119865119873119875 + 119873 (9)


119875119903119890119888119894119904119894119900119899 = 119879119875119879119875 + 119865119875 (10)


119877119890119888119886119897119897 = 119879119875119879119875 + 119865119873 (11)



COLLECT DATA

TRAINING

ABNORMAL


YES NO

YES

WARN USER

NO

EXTRACT FEATURE


00

01

02

03

04

05

06

07

08

09

10

01 02 03 04 05 06 07 08 09 10

more erroneous

less erroneous

A

CB

Generality

Sens

itivi

ty D









Generality

1

0 Probability

AbnormalNormal

11 008

5

Sens

itivi

ty

ob

serv

atio

ns

(a)

Generality

1


AbnormalNormal

10 1

Sens

itivi

ty

ob

serv

atio

ns















7 Conclusion


0 02 04 06 08 10

010203040506070809

1

Generality

Sens

itivi

ty











Acknowledgments


References


































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





Mean Mean

Energy

SMA

SVM

Derivative


Energy


HjorthParameters

Mean


Deviation

Maximum

Fall Factor


Mean

StandardDeviation

Step length

Sole pressure

Single support time

Double support time

Age

Gender

Previous fall




119865119886119897119897 119877119894119904119896 = 013 (119868119886) + 015 (119868119904) + 014 (119868119898)+ 01 (119868119886119889119897) + 018 (119868119888) + 033 (119868119891)

(1)

where











119878119872119860 = 1119879 (int1198790

1003816100381610038161003816119860119909 (119905)1003816100381610038161003816 119889119905 + int1198790

10038161003816100381610038161003816119860119910 (119905)10038161003816100381610038161003816 119889119905

+ int1198790

1003816100381610038161003816119860119911 (119905)1003816100381610038161003816 119889119905)(2)



119878119872119881 = 1119899119899sum119894=1

radic1198602119909119894 + 1198602119910119894 + 1198602119911119894 (3)












|119860 (119905)|2 119889119905 (4)
















P (ATS|)

PltP1

PltP2Fall predicted

No fallYES

NO

YES

Fall detected

NO













119878119901119890119888119894119891119894119888119894119905119910 = 119879119873119879119873 + 119865119875 (6)


119878119890119899119904119894119905119894V119894119905119910 = 119879119875119879119875 + 119865119873 (7)


119860119888119888119906119903119886119888119910 = 119879119875 + 119879119873119875 + 119873 (8)


119864119903119903119900119903 119877119886119905119890 = 119865119875 + 119865119873119875 + 119873 (9)


119875119903119890119888119894119904119894119900119899 = 119879119875119879119875 + 119865119875 (10)


119877119890119888119886119897119897 = 119879119875119879119875 + 119865119873 (11)



COLLECT DATA

TRAINING

ABNORMAL


YES NO

YES

WARN USER

NO

EXTRACT FEATURE


00

01

02

03

04

05

06

07

08

09

10

01 02 03 04 05 06 07 08 09 10

more erroneous

less erroneous

A

CB

Generality

Sens

itivi

ty D









Generality

1

0 Probability

AbnormalNormal

11 008

5

Sens

itivi

ty

ob

serv

atio

ns

(a)

Generality

1


AbnormalNormal

10 1

Sens

itivi

ty

ob

serv

atio

ns















7 Conclusion


0 02 04 06 08 10

010203040506070809

1

Generality

Sens

itivi

ty











Acknowledgments


References


































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in















|119860 (119905)|2 119889119905 (4)
















P (ATS|)

PltP1

PltP2Fall predicted

No fallYES

NO

YES

Fall detected

NO













119878119901119890119888119894119891119894119888119894119905119910 = 119879119873119879119873 + 119865119875 (6)


119878119890119899119904119894119905119894V119894119905119910 = 119879119875119879119875 + 119865119873 (7)


119860119888119888119906119903119886119888119910 = 119879119875 + 119879119873119875 + 119873 (8)


119864119903119903119900119903 119877119886119905119890 = 119865119875 + 119865119873119875 + 119873 (9)


119875119903119890119888119894119904119894119900119899 = 119879119875119879119875 + 119865119875 (10)


119877119890119888119886119897119897 = 119879119875119879119875 + 119865119873 (11)



COLLECT DATA

TRAINING

ABNORMAL


YES NO

YES

WARN USER

NO

EXTRACT FEATURE


00

01

02

03

04

05

06

07

08

09

10

01 02 03 04 05 06 07 08 09 10

more erroneous

less erroneous

A

CB

Generality

Sens

itivi

ty D









Generality

1

0 Probability

AbnormalNormal

11 008

5

Sens

itivi

ty

ob

serv

atio

ns

(a)

Generality

1


AbnormalNormal

10 1

Sens

itivi

ty

ob

serv

atio

ns















7 Conclusion


0 02 04 06 08 10

010203040506070809

1

Generality

Sens

itivi

ty











Acknowledgments


References


































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





P (ATS|)

PltP1

PltP2Fall predicted

No fallYES

NO

YES

Fall detected

NO













119878119901119890119888119894119891119894119888119894119905119910 = 119879119873119879119873 + 119865119875 (6)


119878119890119899119904119894119905119894V119894119905119910 = 119879119875119879119875 + 119865119873 (7)


119860119888119888119906119903119886119888119910 = 119879119875 + 119879119873119875 + 119873 (8)


119864119903119903119900119903 119877119886119905119890 = 119865119875 + 119865119873119875 + 119873 (9)


119875119903119890119888119894119904119894119900119899 = 119879119875119879119875 + 119865119875 (10)


119877119890119888119886119897119897 = 119879119875119879119875 + 119865119873 (11)



COLLECT DATA

TRAINING

ABNORMAL


YES NO

YES

WARN USER

NO

EXTRACT FEATURE


00

01

02

03

04

05

06

07

08

09

10

01 02 03 04 05 06 07 08 09 10

more erroneous

less erroneous

A

CB

Generality

Sens

itivi

ty D









Generality

1

0 Probability

AbnormalNormal

11 008

5

Sens

itivi

ty

ob

serv

atio

ns

(a)

Generality

1


AbnormalNormal

10 1

Sens

itivi

ty

ob

serv

atio

ns















7 Conclusion


0 02 04 06 08 10

010203040506070809

1

Generality

Sens

itivi

ty











Acknowledgments


References


































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





COLLECT DATA

TRAINING

ABNORMAL


YES NO

YES

WARN USER

NO

EXTRACT FEATURE


00

01

02

03

04

05

06

07

08

09

10

01 02 03 04 05 06 07 08 09 10

more erroneous

less erroneous

A

CB

Generality

Sens

itivi

ty D









Generality

1

0 Probability

AbnormalNormal

11 008

5

Sens

itivi

ty

ob

serv

atio

ns

(a)

Generality

1


AbnormalNormal

10 1

Sens

itivi

ty

ob

serv

atio

ns















7 Conclusion


0 02 04 06 08 10

010203040506070809

1

Generality

Sens

itivi

ty











Acknowledgments


References


































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





Generality

1

0 Probability

AbnormalNormal

11 008

5

Sens

itivi

ty

ob

serv

atio

ns

(a)

Generality

1


AbnormalNormal

10 1

Sens

itivi

ty

ob

serv

atio

ns















7 Conclusion


0 02 04 06 08 10

010203040506070809

1

Generality

Sens

itivi

ty











Acknowledgments


References


































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in










7 Conclusion


0 02 04 06 08 10

010203040506070809

1

Generality

Sens

itivi

ty











Acknowledgments


References


































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in



































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in




a review on fall prediction and prevention system for ...fall prediction systems typically estimate...

Documents