journal.pone.0037062 Evaluation of Accelerometer-Based Fall Detection.pdf

of 9
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Evaluation of Accelerometer-Based Fall Detection Algorithms on Real-World Falls Fabio Bagala` 1 *, Clemens Becker 2 , Angelo Cappello 1 , Lorenzo Chiari 1 , Kamiar Aminian 3 , Jeffrey M. Hausdorff 4 , Wiebren Zijlstra 5 , Jochen Klenk 2 1Department of Electronics, Computer Science and Systems, University of Bologna, Bologna, Italy, 2Department of Clinical Gerontology, Robert-Bosch-Hospital, Stuttgart, Germany, 3Laboratory of Movement Analysis and Measurement, Ecol
  Evaluation of Accelerometer-Based Fall DetectionAlgorithms on Real-World Falls Fabio Bagala ` 1 * , Clemens Becker 2 , Angelo Cappello 1 , Lorenzo Chiari 1 , Kamiar Aminian 3 ,Jeffrey M. Hausdorff  4 , Wiebren Zijlstra 5 , Jochen Klenk  2 1 Department of Electronics, Computer Science and Systems, University of Bologna, Bologna, Italy,  2 Department of Clinical Gerontology, Robert-Bosch-Hospital, Stuttgart,Germany,  3 Laboratory of Movement Analysis and Measurement, Ecole Polytechnique Fe´de´rale de Lausanne (EPFL), Lausanne, Switzerland,  4 Tel Aviv Sourasky MedicalCenter, Laboratory for Gait and Neurodynamics, Movement Disorders Unit and Department of Physical Therapy, Sackler Faculty of Medicine, Tel-Aviv University, Tel-Aviv,Israel,  5 Center for Human Movement Sciences, University Medical Center Groningen, Groningen, The Netherlands Abstract Despite extensive preventive efforts, falls continue to be a major source of morbidity and mortality among elderly. Real-timedetection of falls and their urgent communication to a telecare center may enable rapid medical assistance, thus increasingthe sense of security of the elderly and reducing some of the negative consequences of falls. Many different approacheshave been explored to automatically detect a fall using inertial sensors. Although previously published algorithms reporthigh sensitivity (SE) and high specificity (SP), they have usually been tested on simulated falls performed by healthyvolunteers. We recently collected acceleration data during a number of real-world falls among a patient population witha high-fall-risk as part of the SensAction-AAL European project. The aim of the present study is to benchmark theperformance of thirteen published fall-detection algorithms when they are applied to the database of 29 real-world falls. Tothe best of our knowledge, this is the first systematic comparison of fall detection algorithms tested on real-world falls. Wefound that the SP average of the thirteen algorithms, was (mean 6 std) 83.0% 6 30.3% (maximum value=98%). The SE wasconsiderably lower (SE=57.0% 6 27.3%, maximum value=82.8%), much lower than the values obtained on simulated falls.The number of false alarms generated by the algorithms during 1-day monitoring of three representative fallers rangedfrom 3 to 85. The factors that affect the performance of the published algorithms, when they are applied to the real-worldfalls, are also discussed. These findings indicate the importance of testing fall-detection algorithms in real-life conditions inorder to produce more effective automated alarm systems with higher acceptance. Further, the present results support theidea that a large, shared real-world fall database could, potentially, provide an enhanced understanding of the fall processand the information needed to design and evaluate a high-performance fall detector. Citation:  Bagala` F, Becker C, Cappello A, Chiari L, Aminian K, et al. (2012) Evaluation of Accelerometer-Based Fall Detection Algorithms on Real-World Falls. PLoSONE 7(5): e37062. doi:10.1371/journal.pone.0037062 Editor:  Antony Bayer, Cardiff University, United Kingdom Received  October 17, 2011;  Accepted  April 13, 2012;  Published  May 16, 2012 Copyright:    2012 Bagala` et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permitsunrestricted use, distribution, and reproduction in any medium, provided the srcinal author and source are credited. Funding:  The research leading to these results has been partly funded by the European Union under grant agreements n. 045622 (FP6/2002-06: SensAction-AALproject) and n. 288940 (FP7/2007-13: FARSEEING project). The funders had no role in study design, data collection and analysis, decision to publish, or preparationof the manuscript. Competing Interests:  The authors have declared that no competing interests exist.* E-mail: Introduction Despite extensive preventive efforts, falls continue to be a majorsource of morbidity and mortality among older adults. Falls oftenlead to serious injuries such as hip fractures, hospitalization anddeath. Even when no serious injury occurs, the resultant fear of falling and self-imposed restrictions in mobility and function maycontribute to nursing home admission [1] and lead to a loss of personal autonomy that directly affects the quality of life of subjects.Falls among older people remain a very important publichealthcare issue.Real-time detection of falls allows for the immediate commu-nication of these adverse events to a telecare center so that medicalassistance can be supplied quickly. Such assistance is needed topromote the sense of security of older adults, especially among those who are living alone, and to reduce fear of falling and thesubsequent negative impact of falls. Indeed, one of the seriousconsequences of falling is the ‘‘long-lie’’ condition, where a faller isunable to get up and remains on the ground for several hours.‘‘Long-lies’’, and falls, in general, are associated with socialisolation, fear of falling, muscle damage, pneumonia, pressuresores, dehydration and hypothermia [2–5]. Half of the elderlypeople who experience a ‘long-lie’ die within 6 months [6], even if no direct injury from the fall has occurred. The ‘long-lie’ occurs inmore than 20% of elderly people admitted to hospital as a result of a fall [7] and up to 47% of non-injured fallers are unable to get upoff the floor without assistance [8]. Detection of a fall, eitherthrough automatic fall detection or through a personal emergencyresponse system, might reduce the consequences of the ‘long-lie’by reducing the time between the fall and the arrival of medicalattention [9]. If an older person living alone experiences a fall athome, he or she may not be able to get to the phone or press analarm button due to sustained injuries or loss of consciousness [10].Moreover, some elderly people do not activate their personalemergency response systems, even when they have the ability to doso [11]. PLoS ONE | 1 May 2012 | Volume 7 | Issue 5 | e37062  For these reasons, a variety of different methods were developedover the last decade to automatically detect falls. These have beenbased on video-cameras [12–17], acoustic [18–21] or inertialsensors [22–44], and mobile phone technology [45–47].Several of these studies focused on the monitoring of activities of daily living (ADL) and fall detection using wearable sensors.Compared to traditional movement analysis systems, wearablesensors offer advantages in terms of cost, size, weight, powerconsumption, ease of use and, most importantly, portability. Withwearable sensors, data collection is no longer confined toa laboratory environment, thus leading to ubiquitous healthmonitoring.Many different approaches have been explored to solve the falldetection problem using only accelerometers or an inertialmeasurement unit (both gyroscopes and accelerometers) [28– 35,38–40]. The analysis of accelerometer and/or gyroscopeoutputs allows for detecting specific events, such as voluntary(e.g., walking, sitting, lying) or involuntary (e.g., fall) activities of daily living, based on statistical or threshold-based algorithms.The inertial sensor-based fall detection algorithms usuallyprovide: i) a definition of a set of parameters related to theaccelerometer and gyroscopes outputs, used for the characteriza-tion of the movement, ii) impact detection, using a threshold-basedmethod, iii) orientation detection, e.g., using the vertical acceler-ometer output or angular rate measurements, and iv) fall alarm,which occurs when all the test conditions are true.Published algorithms have generally been tested only onsimulated falls. Most authors have used simulations with healthy volunteers [29,30,32–34,38,41,42,44,45] or martial arts students[28] as a surrogate for real-world falls [48]. To the best of ourknowledge, there is a lack of published inertial measurement-basedreal-world fall data of older people measured in a real-worldenvironment. Although the rate of falls is quite high (approximately 30% of persons over 65 years fall at least once per year), it is very difficultto capture real-world fall data. This largely is a result of therelatively short measurement intervals allowed by commerciallyavailable sensors. As an example, to capture 100 real-world falls, itwould be necessary to record approximately 100,000 days of physical activity (300 person years). If the battery lifetime is limitedto 10 days, 10,000 measurement cycles would be needed. Additionally, compliance problems may arise with long measure-ment periods. As far as we know, most international studies havefailed to gather sufficient numbers of fall events. Recently, Kangas et al.  [49] collected acceleration data of 5 real-world falls during a six-month test period in older people.To address the challenges of capturing real-world falls, webegan to collect acceleration data during a number of real-worldfalls as part of a European project (SensAction-AAL) that studieda population with a high-risk of falling. Based on these data,a recent study [50] compared acceleration signals, measured using a tri-axial accelerometer placed on the waist of the subjects, fromsimulated falls and these real-world falls and found largedifferences between them, even though a relatively simple exampleof falling backward to the ground was selected.Several problems are associated with the simulation approachincluding the anticipation of the volunteer that a fall will occur andthe choice of the floor material to reduce the impact of the falls forsafety reasons. These findings underline the importance of gathering real-world fall data for designing accurate algorithms.With the limitations of simulated falls in mind, the aim of thepresent study is to benchmark, for the first time, the performanceof 13 different published algorithms as applied to the database of 29 real-world falls collected during the SensAction-AAL project.In order to compare the performance in the same test conditionsas our real-world fall data, only algorithms based on waist or trunk accelerometer measurements were investigated. Algorithms basedon gyroscopes measurements or on more than one sensor are notconsidered in this paper. Materials and Methods 1. The real-world fall database  Acceleration signals of 32 falls from 15 subjects were collectedduring the SensAction-AAL project and clinical routine assess-ments. 30 falls from 9 subjects (7 women, 2 men, age: 66.4 6 6.2 years, height: 1.63 6 8.68  m , weight: 77.2 6 11.5  kg   ) were recordedwithin a cross-sectional study of patients suffering from progressivesupra-nuclear palsy (PSP) [51] and from an intervention study toinvestigate the feasibility of audio-biofeedback to improve balance[52]. PSP is an atypical Parkinson’s syndrome with a prevalence of 5 per 100,000 [53]. Postural instability and falls are common andare the most disabling features of the disease [54,55]. A 48-hactivity measurement was conducted on 29 subjects as part of theassessment in the cross-sectional study and during days withoutintervention. A fall was defined as ‘‘ an unexpected event in which the  participant comes to rest on the ground, floor, or lower level  ’’ [56]. Patientsor their proxies reported the time, the place and the circumstancesof the falls.Two additional falls were recorded from one subject withina cross-sectional study in community-dwelling older people. All of these falls were recorded during daily physical activity measure-ment using an ambulatory device based on accelerometers(Dynaport H  MiniMod, McRoberts, The Hague, NL). For thesake of the present study, for each fall, we extracted, from the24 hour recording, a 60 second time-window centered around thefall event. The falls were characterized with respect to location,pre-fall phase, fall direction, and impact spot (Table 1). TheMiniMod H , composed of a tri-axial seismic acceleration sensor(LIS3LV02DQ STMicroelectronics, Agrate Brianza, Italy), wasfixed by a belt at the lower back. The orientation of the axes arex=vertical, y=medio-lateral (left/right), and z=anterior-posteri-or (forward/backward). The sensor has a resolution of 12 bit anda sampling frequency of   f  c ~ 100  Hz . The published fall detectionalgorithms were usually based on measurements carried out byaccelerometers with a sampling frequency varying from 50  Hz   to250  Hz   and a range of  6 10  g   or 6 12  g  . We recorded 14 falls witha sensor’s range of   6 6  g  , the remaining 18 falls with a sensor’srange of  6 2  g  . When the acceleration exceeds the threshold 6 2  g  ,the so-called ‘‘clipping effect’’ (or saturation) produces a cut-off of the signal. Since this could affect the results of the analysis, threefalls that show saturation effects are not included in the analysis.Therefore, the total number of falls considered in this study was29. Raw data were stored for off-line analysis on a SD card. 2. The algorithms The algorithms used are summarized here; additional detailscan be found in the literature [28–32]. Table S1 summarizes theparameters, thresholds and the phases of a fall event that areconsidered: beginning of the fall, falling velocity, fall impact andorientation after the fall. The outputs of the tri-axial accelerometerare  A x ( k  ), A  y ( k  ), A z ( k  ) , with  k =1,…,n  where  n  is the number of samples.Chen  et al.  [28] used a tri-axial accelerometer worn on the waistof two martial arts students, who performed some common fallmotions over 10 trials. If the root sum vector (  SV   ) of the threesquared accelerometer outputs exceeds a threshold, it is possiblethat a fall has occurred (IMPACT DETECTION). Additionally, Fall Detection Algorithms on Real-World FallsPLoS ONE | 2 May 2012 | Volume 7 | Issue 5 | e37062  the orientation is calculated over 1 second before the first impactand 2 second after the last impact using the dot product of theacceleration vectors (CHANGE IN ORIENTATION). The anglechange that constitutes a change in orientation can be setarbitrarily based on empirical data, as suggested by the authors.We set this threshold to 20 u  in order to have the best sensitivityand specificity. No results are reported in the paper, but theauthors point out the benefits due to the evaluation of change inorientation.Kangas  et al.  [29] attached a tri-axial accelerometer to the waist,wrist and head of three healthy middle-aged volunteers, whoperformed three standardized types of falls (forward, backward,and lateral) towards a mattress. Examples of activities of dailyliving (ADL) were collected from two healthy subjects, represent-ing dynamic activities (e.g., walking, walking on the stairs, picking up objects from the floor). Four different detection algorithms,Kangas1a to Kangas1d, with increasing complexity were in- vestigated. The thresholds are related to the waist measurement.These four algorithms had in common IMPACT DETECTION  + POSTURE MONITORING. They were based on the detectionof the impact by threshold on the sum vector (  SV   ), the dynamicsum vector (  SV  D  ) related to the high-pass filtered (HPF)accelerometer outputs, the sliding sum vector (  SV  MaxMin  ) andthe vertical acceleration (  X  2  ), respectively, followed by monitoring of the subject’s posture. The posture was detected 2 seconds afterthe impact from the low-pass filtered (LPF) vertical signal, basedon the average acceleration in a 0.4 second time interval, witha signal value of 0.5  g   or lower considered to be a lying posture.Two further algorithms, Kangas2a and Kangas2b, wereconsidered from Kangas et al. [29] based on START OF FALL +  IMPACT DETECTION  +  POSTURE MONITORING.These algorithms detected the start of the fall by monitoring   SV  lower than a threshold of 0.6  g  , followed by the detection of theimpact within a time frame of 1  s   by a threshold value of   SV   or X  2 , followed by posture monitoring.Three further algorithms, Kangas3a to Kangas3c, based onSTART OF FALL  +  VELOCITY  +  IMPACT DETECTION  + POSTURE MONITORING were considered from [29]. Thesealgorithms detected the start of the fall, followed by detection of the velocity  v 0  (calculated by integrating the area of   SV   from thetrough (see Fig. 1), at the beginning of the fall, until the impact,where the signal value is lower than 1  g   ) exceeding the threshold,followed by the detection of the impact within a time frame of 1  s  by a threshold value of   SV   or  X  2 , followed by posture monitoring.The fall detection sensitivity, declared by the authors [29], of thedifferent eight algorithms at the waist varied from 76% to 97%and the specificity was 100%.Bourke  et al.  [30] fixed two tri-axial accelerometers to the trunk (at the sternum) and the thigh. Ten young subjects were involvedin simulated falls onto large crash mats. Ten community-dwelling elderly subjects performed ADL in their own homes (e.g., sit tostand, lying, walking). In these algorithms (Bourke1a andBourke1b), the  SV   of the three signals was evaluated from thesternum and thigh accelerometer outputs and a fall was detectedwhen the  SV   is over the upper (UFT) threshold (3.52  g   ) or underthan the lower (LFT) threshold (0.41  g   ). Declared specificity is100% for the upper threshold and 91.25% for the lower threshold,related to the trunk sensor. In this paper, as suggested by theauthors [30], the thresholds were set according with the fallsdatabase. The UFT and LFT were set at the level of the smallestmagnitude upper fall peak (Bourke1a) and at the level of thebiggest magnitude lower fall peak (Bourke1b), respectively. Basedon the accelerometer data of the 29 falls, we set the two thresholdsto 1.79 g (UFT) and 0.73 g (LFT). Exceeding any individual limitwould indicate a fall.Bourke  et al.  [31] developed a second fall detection system using a tri-axial accelerometer to detect impacts. The algorithm (Bourke2), considered the  SV   of the accelerometer outputs, and monitorposture, assuming a lying posture if the vertical accelerometersignal value is between 2 0.5  g   and 0.5  g  . The sensor was attachedto a custom designed vest. Two teams of 5 elderly subjects testedthe algorithm. Over 833 hours of monitoring, no actual falls wererecorded, although the system registered a total of 42 false alarms(i.e., false positives).Recently, Bourke  et al.  [32] evaluated 21 fall-detectionalgorithms of varying degrees of complexity for a waist-mountedaccelerometer based system. The algorithms were tested againsta comprehensive data-set recorded from 10 young healthy volunteers performing 240 simulated falls and 120 ADL and 10elderly healthy volunteers performing 240 scripted ADL and 52.4waking hours of continuous unscripted normal ADL. Here, weevaluated the algorithm (Bourke3) VELOCITY + IMPACT + POS-TURE that achieved 100% sensitivity and specificity and with thelowest false-positive rate (0.6 false positive per day) when appliedto simulated falls and tested it on the real-world falls database. Thealgorithm is based on the detection of the four distinct phases of a fall [48] (pre-fall, critical phase, post-fall phase and recovery)when the SV exceeds the LFT (0.65  g   ) and the UFT (2.8  g   )thresholds. Two temporal features and their related thresholds areconsidered: the falling-edge time,  t FE  , is from the SV signal lastgoing below the LFT until it exceeds the UFT (threshold set to600  ms   ), and the rising-edge time,  t RE  , is the last time when theLFT is exceeded until the UFT is exceeded (threshold set to350  ms   ). The vertical velocity is further considered as an indicatorof a fall when it overcomes the threshold  V  T   (  2 0.7  m/s   ). It isevaluated through the numerical integration of the SV signal withthe gravity component subtracted. The post-fall posture isdetermined taking the dot product of the gravity vector  g REF  and the current gravity vector estimated relative to the bodysegment  g SEG  ( t ) . Lying is detected if the waist posture,  q ( t ) , fromt + 1 s to t + 3 s exceeds 60 u  for more than 75% of the duration. As summarized in Table S1, the  SV   is a common feature among all the algorithms. An example of prototypical signal of the  SV   isshown in Fig. 1. The signal reflects a forward real-world fall in Table 1.  Description of real-world falls ( n =32). Number of falls per condition Location Indoor ( n =30), outdoor ( n =2)Activity before the fall Standing ( n =16), walking forward ( n =8), walking backward ( n =1), sit-to-stand ( n =5), stand-to-sit ( n =2)Reported direction of fall Forward ( n =8), backward ( n =18), sideward ( n =6)Impact spot Floor ( n =23), against wall/locker before hitting the floor ( n =4), bed/sofa ( n =4), desk ( n =1)doi:10.1371/journal.pone.0037062.t001 Fall Detection Algorithms on Real-World FallsPLoS ONE | 3 May 2012 | Volume 7 | Issue 5 | e37062  which the subject fell directly on the floor while bending to pick upan object. The typical trough before the impact, the impact andthe maximum magnitude due to the impact are also indicated.The 29 accelerometer fall recordings were used to test theperformance of the algorithms in terms of sensitivity (SE,percentage of falls correctly detected as such). Further signalanalysis was performed in order to evaluate the specificity (SP,percentage of ADL correctly identified as non-falls).Previous studies tested the specificity of ADL performed in thelaboratory environment by the same subjects who simulated falls(generally healthy young subjects) or community-dwelling elderlysubjects. These data could be biased, since subjects are forced toperform activities, which are typically spontaneous. To avoidbiased results for specificity, we extracted ADL based on theindividual physical activity recordings from each subject excluding the 60 second fall-time-windows. The remaining observation timewas also separated into 60 second time-windows.The recordings of 8 of the 15 fallers were carried out using thesensor with range  6 2  g   and therefore were excluded from thespecificity evaluation. We collected, for the remaining 7 subjects,168  h   of accelerometer recordings, i.e., 10,050 time-windows of 60 seconds (the 29 time-windows related to falls were excluded).These time windows could be related to resting periods. In allthese cases, the fall detection algorithms correctly identify 100% of  ADL as not-falls. Thus, the SP will show high values because of thehigh number of time windows with inactivity included in theanalysis. According to these considerations, the time windowsrelated to resting periods were excluded and those related toactivity periods were considered in the study according witha simple procedure. We assumed that an activity is performed if the dynamics of the signal (the difference between the maximumand the minimum value) in a 60 second time window overcomesa fixed threshold TH. This was selected from the following steps: – the difference  M  i  ~ max( SV  i  ) { min( SV  i  )  (i=1,…, 10,050)was evaluated from the accelerometer outputs for each of the10,050 time-windows; – the 10,050 time-windows were tested by the 13 algorithms; – if the  k  -th time window was wrongly identified as fall, the value M  k   was allocated in a vector  M ; – after testing the 13 algorithms, the minimum element of   M  wasconsidered as the threshold TH for discriminating resting fromactivity periods. All the time-windows with  M  i  w TH   were considered as ADLand thus selected for the analysis. The threshold evaluated by theprocedure was TH=1.01  g  . The total number of time-windowsconsidered was 1,170.The accuracy (ACC, the ratio between the number of correctassessments, falls and ADL, and the number of all assessments), thepositive predictive value (PPV, the probability that a time windowwith a positive test result, fall detected, really does have thecondition for which the test was conducted) and the negativepredictive value (NPV, the probability that a time window witha negative result, fall undetected, really does have the condition forwhich the test was conducted) were evaluated for each algorithm.Moreover, the performance of the tested algorithms wereevaluated on 24 hour accelerometer recordings for three of thePSP fallers, in order to evaluate the number of false alarms (ADLdetected as falls) generated by the different algorithms.Data analysis was performed using MATLAB 7.9.0 (R2009B). Results In order to show an example of real-world fall signals, the sum vector of a backward fall and its detail is reported in Fig. 2(a). Thesum vector related to one of the randomly extracted ADL is shownin Fig. 2(b). Figure 1. Prototypical acceleration sum vector of a fall.  This real-world example illustrates components that are common to many falls.doi:10.1371/journal.pone.0037062.g001Fall Detection Algorithms on Real-World FallsPLoS ONE | 4 May 2012 | Volume 7 | Issue 5 | e37062


Jul 23, 2017
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks