WK(XURSHDQ6LJQDO3URFHVVLQJ&RQIHUHQFH (86,3&2

A Highly Reliable Wrist-Worn Acceleration-Based

Fall Detector
Majd SALEH∗ , Nawras GEORGI∗ , Manuel ABBAS∗ and Régine LE BOUQUIN JEANNÈS∗
∗
Univ Rennes, Inserm, LTSI - UMR 1099, F-35000 Rennes, France
majd.saleh@univ-rennes1.fr nawras.georgi@univ-rennes1.fr
manuel.abbas@etudiant.univ-rennes1.fr regine.le-bouquin-jeannes@univ-rennes1.fr

Abstract—Automatic fall detection for the elderly is one of the of classifiers. Thanks to the small size of feature space, an ef-
most important health-care applications since it enables a rapid ficient implementation was proposed where the trained model
medical intervention preventing serious consequences of falls. was not required in the embedded algorithm but a table that
Wrist-worn fall detectors represent one of the most convenient
solutions. However, power consumption has a notable impact on contains all the possible feature/response pairs was embed-
the acceptability of such devices since it affects the size and weight ded instead. This algorithm achieved an accuracy of 99.65%
of the required battery and the rate of replacing/recharging it. In with extremely low computational complexity [3]. Recently,
this paper, an acceleration-based fall detection system is proposed two machine learning-based algorithms for acceleration-based
for wrist-worn devices. It consists of two stages. The first one waist-mounted fall detectors have been proposed in [4], both
is a highly-sensitive low computational complexity algorithm to
be embedded in the wearable device. When a potential fall is achieving an accuracy greater than 99.9% when evaluated on
detected, raw data are transmitted to a remote server for accurate a large open dataset. This superior accuracy has been satisfied
analysis in order to reduce the number of false alarms. The with quite low computational complexity which enables a fall
second stage algorithm is based on machine learning and applied detector to work for years with a 1000 mAh-battery. The main
to highly discriminant features. The latter are selected using reasons underlying the high performance of waist-mounted
powerful feature selection algorithms where the input is 12 000
features extracted from each entry of a large activity dataset. The fall detectors is that the waist is a good place to capture
proposed system achieved an accuracy of 100% when evaluated the activity of the user since it is close to the center of the
on a 2400-file dataset. Moreover, the feasibility of the proposed body mass and the device could be tightly fixed in order to
system has been validated in real world conditions where it has avoid oscillations. In spite of their perfect performance, waist
been realized and tested using a smart watch and a server. devices could not easily accompany the elderly in any place,
Index Terms—fall detection, machine learning, elderly health-
care, wearable sensors, feature selection. such as under shower, in bed, . . . etc. Therefore, alternative
positions have been considered by researchers. Among the
I. I NTRODUCTION alternatives, the wrist is one of the most acceptable positions.
A notable demographic shift has been evaluated by the Recently, Quadros et al. [5] have proposed an algorithm for
World Health Organization (WHO) [1] where the number of wrist-worn fall detectors with an accuracy of 99.0%. Using
people older than 60 years is expected to exceed the number an accelerometer, gyroscope and magnetometer, the orientation
of children younger than 5 years by 2020. This problem leads of the fall detector is estimated using Madgwick sensor fusion
to major challenges to the health systems in all countries algorithm [6]. The extracted features are based on Euler angles
worldwide [1]. In a related context, WHO also showed that that represent the orientation of the fall detector with respect
an average of 100 000 falls that require medical intervention to the Earth frame and the rest of features are extracted
occur every day worldwide and, as expected, the majority of from vertical acceleration, velocity and displacement that also
fatal falls occur with people older than 65 years [2]. Therefore, require estimating the orientation of the device. The classifier
automatic fall detection for the elderly is one of the most that shows the aforementioned result is k-NN. Despite the
important health-care applications since it enables a rapid impressive accuracy of this algorithm, its complexity could
medical intervention and thus prevents serious consequences considerably limit the battery life of a wearable fall detector.
of falls. Thanks to the new technology of microscopic de- The complexity includes: 1) the hardware complexity due to
vices, namely Micro-Electro-Mechanical Systems (MEMS), a the need for three sensors, 2) the need for fusion algorithm and
multitude of small-size light-weight wearable fall detectors 3) the complexity of the classifier where making a decision
have been developed over the last two decades. These devices using a k-NN classifier requires calculating the distances
use MEMS-based accelerometers, gyroscopes, magnetometers between the extracted features vector and all the stored training
and/or barometers to capture the activity of the user. Waist- vectors [4].
mounted fall detectors have shown high performance levels Power consumption of a wearable fall detector has a consid-
in the last years. Even using only an accelerometer, highly- erable impact on its acceptability for the following reasons:
accurate low-complexity solutions have been proposed in the 1) the need for replacing/recharging the battery decreases
literature. For instance, in [3] local binary features have been with lower power consumption and 2) using small batteries
proposed to discriminate between falls and Activities of Daily enables producing small-size light-weight fall detectors. In
Living (ADLs). These features were used to train several types order to minimize the power consumption, two factors are
to be considered: (i) the computational complexity of the

k,(((
 WK(XURSHDQ6LJQDO3URFHVVLQJ&RQIHUHQFH (86,3&2

embedded algorithm should be low and (ii) since the power (a) (b)
‫ݖ‬
consumption of accelerometers is very low (few μAh) espe-
cially in comparison with gyroscopes [7], acceleration-based ʹ
solutions are to be preferred. To this end, the objective of
the current work is to investigate the feasibility of detecting ‫ݖ‬
ࢍ
falls using only an accelerometer in a wrist-worn device. The
ࢇ ࢇᇱ
strategy proposed to tackle the aforementioned problem is to ‫ݔ‬ ‫ݕ‬ ʹ ‫ݕ‬
describe acceleration signals of falls and ADLs using a large
ʹ
number of features and then to apply a variety of feature
selection methods to discover the most discriminant features. ʹ
In order to decrease the computational load on the wearable ‫ݔ‬
device, feature extraction will be executed on a remote server
only when a potential fall is detected by the wearable device Fig. 1. A wrist-worn fall detector: a) the senor coordinate frame, b) an
as will be explained later. example illustrating the relationship between the measured, dynamic and
gravity accelerations
Feature selection methods could be divided into filters, wrap-
pers and embedded methods [8]. Wrappers and embedded could carry important indicators for falls. However, it is a
methods employ a predictor to select the features while demanding task as it requires a gyroscope and a sensor fusion
filters select features without optimizing the performance of a algorithm like in [5]. On the other hand, working directly
predictor [8]. The difference between wrappers and embedded with the measured acceleration is a challenge because of the
methods is that the former use the predictor as a black ambiguity discussed above. So, our objective is to use only the
box while in the latter, feature selection is involved in the measured acceleration, a, in order to discover features that can
training process. In this paper, the considered wrappers are: discriminate between falls and ADLs.
feature ranking based on individual feature performance using
a logistic regression classifier, Sequential Forward Selection III. M ETHODOLOGY
(SFS) [9] and Sequential Backward Floating Selection (SBFS) A. Motivation
[10]. Note that the considered criterion to be maximized in The presence of a high acceleration peak followed by
both SFS and SBFS is the classification accuracy of a linear inactivity is a strong indicator to detect falls [3]. However,
SVM classifier evaluated using 10-fold cross validation. The it is not sufficient to avoid false alarms. For instance, laying
considered embedded methods are: Support Vector Machine down on a bed could easily satisfy the aforementioned sit-
based on Recursive Feature Elimination (SVM-RFE) [11], uation. Therefore, sophisticated features should be extracted
stepwise multi-linear regression, random forest and Neigh- from acceleration signals in order to achieve an accurate
borhood Component Feature Selection (NCFS) [12]. The solution. Studies on time series analysis in different application
considered filters are: feature ranking based on correlation co- fields generally result in thousands of useful features. For a
efficients, feature ranking using Fisher ratio and the minimum challenging problem like acceleration-based fall detection for
Redundancy Maximal Relevance method (mRMR) [13]. wrist-worn devices, investigating the feasibility of employing
In the next section, we explain the challenges of detecting these features is attractive. Thanks to the recent MATLAB
falls using acceleration-based wrist-worn devices before ex- tool, hctsa [14], [15], an automatic massive feature extraction
plaining the proposed methodology in Section III. Experimen- could be applied to any time series. More than 7.700 time
tal results are described in Section IV before giving conclusion series features that encapsulate several decades of research in
in Section V. feature extraction could be employed. In our strategy, both of
II. C HALLENGES WITH ACCELERATION - BASED FALL the last massive-feature-based solution and the simple low-cost
DETECTION USING WRIST- WORN DEVICES criteria mentioned above are exploited.
Capacitive MEMS accelerometers are widely used in the B. The proposed fall detection algorithm
context of fall detection thanks to their low cost, small size, The activity of the user is captured using a 3-axial ac-
light weight and low power consumption. Moreover, they can celerometer built in the wrist-worn device. In order to analyze
measure static accelerations contrary to the AC accelerometers the acceleration signal during a sufficient period, the latter
like e.g. piezoelectric ones. For these reasons, the focus in this is buffered in a 3-second sliding window. More precisely,
work is on capacitive MEMS accelerometers.
The measured tri-axial acceleration a in the sensor frame  a = [ax ay az ] denotes the acceleration, ax and
given that
represents a vector combination of the dynamic acceleration a = a2x + a2y + a2z are buffered in the 3-second window
a caused by the body movement and the gravity acceleration A ∈ R2×3fs where fs denotes the sampling frequency. We
g i.e. a = a + g. Theoretically, the measured acceleration only consider ax and a since these signals are not affected
represents the center of a unit sphere (the radius is one when the device rotates around the wrist, recalling that ax
gravity unit) while the dynamic acceleration could be any corresponds with the axis of the arm as shown in Figure 1.a.
point on the surface of this sphere. Figure 1.a illustrates The choice of the sliding window length is based on the
the sensor coordinate frame while Figure 1.b illustrates an time structure of falls. Indeed, falls consist of three phases
example of the aforementioned ambiguity. Separating a and namely pre-fall, critical and post-fall phases [16] as illustrated
g could be quite useful for fall detection because each of them in Figure 2 where we represent the acceleration of a forward
 WK(XURSHDQ6LJQDO3URFHVVLQJ&RQIHUHQFH (86,3&2

hctsa as introduced in III-A and then to select the smallest

set of features that satisfy the maximum attainable accuracy.
To implement the aforementioned strategy, we recorded 2400
activity signals from 7 subjects. These activities covered a
variety of ADLs and simulated falls (further details on the
dataset are explained in Section IV). From each activity signal,
12 000 features were separately extracted from ax and a.
Let B ∈ R2400×12000 denotes the set of all the extracted
Fig. 2. Major phases of falls illustrated by 3-axial acceleration. Example of features. Using 10-fold cross validation, the input of a feature
a forward fall caused by a stumble while walking. selection algorithm is a subset B i ⊂ B ∀i ∈ {1, . . . , 10}
Wrist-worn fall detector Remote server
where the dimensions of B i are 2160 × 12000. For com-
Start Start
prehensive feature selection analysis, 10 methods have been
ܽ௫ ǡ ܽ௬ ǡ ܽ௭ ࡭ considered including embedded methods, wrappers and filters
as mentioned in Section I. The minimum set of selected
ࢇ ൌ ܽ௫ଶ ൅ ܽ௬ଶ ൅ ܽ௭ଶ Feature extraction features that satisfy 100% of accuracy has been chosen for the
ሺଵሻ ሺேሻ
machine learning-based algorithm. When the server receives
ܽ௫ ܽ௫ ‫ܽ ڮ‬௫
ࢇ
՜࡭ൌ
ࢇ ሺଵሻ ‫ڮ‬ ࢇ ሺேሻ Linear SVM classifier A, it extracts the pre-defined set of most relevant features.
These features represent the input of a trained linear SVM
No ࢇ ሺ௜ሻ ൐ ܶଵ Fall Yes Send classifier. When the latter detects a fall, an alarm is sent to an
‫݂ א ݅ ׊‬௦ ǡ ʹ݂௦ detected
?
Alarm authorized party as shown in Figure 3.
Yes No The performance of the considered feature selection algo-
ሺ௜ሻ
‫ ݏݒ‬ൌ ‫ ࢇ ݎܽݒ‬ሺ௜ሻ ൅ ‫ܽ ݎܽݒ‬௫ End rithms, the predictive power of the extracted features and
‫݂ʹ א ݅ ׊‬௦ ǡ ͵݂௦ the performance of the proposed algorithm in a real world
No
‫ ݏݒ‬൏ ܶଶ Network
application are explored in the next section.
Yes
Send raw data ࡭ Router
IV. E XPERIMENTS
For training and performance evaluation, 7 young adults
Fig. 3. Flowchart of the proposed fall detection system (4 males and 3 females) were asked to perform 19 types
of ADLs and to simulate 35 types of falls. These activi-
fall caused by a stumble while walking. The critical phase ties were recorded using a wrist-worn data-logger equipped
includes a free fall followed by an impact when the body with a 3-axial accelerometer. The latter is a DC capacitive
hits the ground and then an adjustment until the body takes MEMS accelerometer. Its measurement range is ±8 g and the
its final position. Thus, at some instant of activity monitoring sampling frequency is fs = 238 Hz. Raw acceleration data
using overlapped windows, the sliding window captures all of were segmented into 2400 files (608 falls and 1792 ADLs).
the three phases of fall. The considered performance criteria are accuracy, specificity
The proposed strategy is to detect falls in two stages as shown and sensitivity [3], [4]. Note that the specificity is inversely
in Figure 3. The first one is a threshold-based algorithm to be proportional to the false alarm rate.
embedded in the wearable device. This algorithm is designed
to satisfy two objectives: 1) low computational cost in order A. Analyzing the classification performance with a small set
to minimize the power consumption of the wearable device of features
and 2) highly sensitive fall detection. When a fall is detected In order to show the discriminant power of a small set of
using this algorithm, the buffered acceleration A is sent to features, the top ranked three features selected using the SVM-
a remote server that performs machine learning-based fall RFE method are discussed in detail. The features, denoted as
detection where the objective is to minimize the number of F 1, F 2 and F 3, are extracted from ax . Figure 4 shows how
false alarms. Both stages are discussed hereafter. these features are extracted where an example of a backward
The embedded algorithm: as illustrated in Figure 3, fall occurred while walking and caused by a slip is considered.
a(i) ∀i ∈ [fs , 2fs ] is compared with a predefined threshold Feature F 1 represents the discrepancy between the uniform
T1 in order to detect the presence of an impact. If an impact distribution fitted to ax and the kernel-smoothed distribution
is detected, the signal in the post-fall phase is inspected in of the signal ax denoted as fˆh (u) where h represents the
order to detect inactivity. Particularly, the sum of variance of bandwidth of the scaled normal kernel used for smoothing.
(i)
a(i) and ax ∀i ∈ [2fs , 3fs ] is compared with a predefined This discrepancy is measured as the distance between the left
threshold T2 . If inactivity is detected, the buffered acceleration edge of the uniform PDF and the argmaxu fˆ(u) as illustrated
A is sent to a remote server for further analysis. in Figure 4.a. Features F 2 and F 3 are extracted from the
The machine learning-based algorithm: in order to minimize standardized acceleration ãx . In order to extract the feature F 2,
the number of false alarms, features of higher discriminant the 3-second sliding window is divided into 5 sub-windows.
power are required. Such features are to be extracted from A The local sample entropy of ãx is calculated for each sub-
on the remote server. The question is: what are those features window resulting in a vector s = [s1 s2 s3 s4 s5 ]. Now, feature
to be used in the machine learning-based algorithm? The pro- F 2 is calculated as the standard deviation of s as illustrated
posed strategy is to perform a massive feature extraction using in Figure 4.b. Thus, F 2 is a measure of stationarity of ax . In
 WK(XURSHDQ6LJQDO3URFHVVLQJ&RQIHUHQFH (86,3&2

(a) (b) (c)

Fig. 4. How to calculate the top 3 features selected using SVM-RFE

Fig. 5. The discriminative power of the top three features selected using SVM-RFE. Violin plots of features F 1, F 2 and F 3 are illustrated in (a),(b) and
(c), respectively. Sub-figure (d) illustrates a 3D representation of the falls and ADLs as well as the separating hyperplane achieved by a quadratic SVM.

a similar manner, F 3 is extracted by dividing the 3-second lected features using the i-th method. A 10-fold cross-validated
sliding window into 7 sub-windows and calculating the local linear SVM classifier is used to evaluate the performance of
(i) (i)
features di = | max (ãx ) − | min (ãx )||, ∀i ∈ {1, 2, . . . , 7}, the selected features. Figure 6 shows that embedded methods
as illustrated in Figure 4.c. Thus, F 3 reflects the variation of display superior performance in comparison with wrappers
local maxima and minima across ax . The individual discrim- and filters. Particularly, SVM-RFE presents the best result
ination performance of features F 1, F 2 and F 3 is illustrated achieving a classification accuracy of 100% using the top 121
in Figures 5.a,b and c, respectively where violin plots show features, recall that the total number of extracted features is
the distribution of the aforementioned features when extracted 12 000. The other methods show lower quality especially in
from falls and ADLs. From these violins, it is clear that none terms of specificity as shown in Figure 6.
of these features is sufficient to discriminate between falls and The top 121 features selected using SVM-RFE are based
ADLs. However, when considering a combination of these on entropy and mutual information, correlation, stationarity
features, a quadratic-kernel-based SVM is able to separate analysis, model fitting and spectral analysis. Explaining these
the two classes with an accuracy of 92.35% as illustrated in features in detail is out of the scope of this paper. The
Figure 5.d. More precisely, a 10-fold cross validation is applied last result satisfies the objective of this work where the
to evaluate the performance of the aforementioned SVM problem of acceleration-based fall detection using wrist-worn
classifier. The resulting accuracy, sensitivity and specificity devices is solved with 100% accuracy. It is also clear from
are 92.35%, 96.25% and 80.78%, respectively. A linear SVM Figure 6 that a trade-off between the classification accuracy
shows comparable results i.e. 91.58%, 95.31% and 80.56% of and the number of selected features could be controlled in
accuracy, sensitivity and specificity, respectively, as shown in order to reduce the complexity of extracting features. Table
Table I. These results show that higher dimensional features I shows some interesting levels of the performance of top
are needed basically to improve the specificity. features selected using SVM-RFE. For instance, using the top
30 features, all performance criteria exceed 98%. However,
B. Analyzing the classification performance as a function of this accuracy/complexity trade-off is not necessary as feature
the number of selected features extraction is executed on a remote server which is able to
The performance of the 10 feature selection algorithms calculate the 121 features rapidly.
considered in this paper is evaluated as a function of the
number of selected features as illustrated in Figure 6. For C. Proof of Concept
each method, the number of selected features varies from 1 to The proposed first-stage fall detection algorithm was im-
min(300, sizeof (Si )) where Si is the total number of the se- plemented in Java using Android Studio IDE and embedded
 WK(XURSHDQ6LJQDO3URFHVVLQJ&RQIHUHQFH (86,3&2

Fig. 6. Classification accuracy, specificity and sensitivity vs. the number of selected features. A comparative study between 10 feature selection methods

TABLE I required by the server. In a future work, the feasibility of

P ERFORMANCE OF A LINEAR SVM CLASSIFIER APPLIED TO THE TOP x tackling the same problem using low complexity features will
FEATURES SELECTED USING SVM-RFE
be investigated in order to develop a standalone fall detector.
x→ 3 10 20 30 40 50 121 ACKNOWLEDGMENT
Accuracy % 91.58 95.75 97.71 98.42 98.83 99.50 100 This publication is supported by the European Union
Sensitivity % 95.31 97.32 98.27 98.55 99.10 99.67 100 through the European Regional Development Fund (ERDF),
Specificity % 80.56 91.10 96.05 98.02 98.02 99.01 100 the Ministry of Higher Education and Research, the French
in Huawei Watch 2. It is a smart-watch running Wear OS, a region of Brittany and Rennes Métropole.
R EFERENCES
modified version of Android OS for wearable. The algorithm
used the IMU which is a LSM6DS3 where the acceleration [1] World Health Organization (WHO). Ageing and health, (feb. 2018)
https://www.who.int/news-room/fact-sheets/detail/ageing-and-health.
range is set to ±8 g and the sampling frequency is 53 [2] World Health Organization (WHO). Falls, (jan. 2018)
Hz. The proposed second-stage fall detection algorithm was https://www.who.int/news-room/fact-sheets/detail/falls.
implemented in MATLAB and installed on a server running [3] M. Saleh and R. Le Bouquin Jeannès. An efficient machine learning-
based fall detection algorithm using local binary features. In EUSIPCO,
a 64-bit Linux Debian Jessie 8 operating system on an Intel Rome, Italy, Sept. 3-7 2018.
Xeon Processor 2.1 GHz and 16 GBytes of RAM. The system [4] M. Saleh and R. Le Bouquin Jeannès. Elderly fall detection using
was tested by one subject where 5 types of ADLs i.e. walking, wearable sensors: A low cost highly accurate algorithm. IEEE Sensors
Journal, 19(8):3156–3164, 2019.
sitting down, standing up, laying down and rising up were [5] T. Quadros, A. E. Lazzaretti, and F. K. Schneider. A movement
performed and 4 falls in different directions were simulated. decomposition and machine learning-based fall detection system using
All falls were detected while no false alarm was generated. wrist wearable device. IEEE Sensors Journal, 18(12):5082 – 5089, 2018.
[6] S.O.H. Madgwick, A.J.L. Harrison, and R. Vaidyanathan. Estimation
This validates the perfect accuracy of the proposed solution. of IMU and MARG orientation using a gradient descent algorithm. In
IEEE ICORR, Zurich, Switzerland, 29 June-1 July 2011.
V. C ONCLUSION AND F UTURE W ORK [7] Q. Liu, J. Williamson, K. Li, W. Mohrman, Q. Lv, R.P. Dick, and
L. Shang. Gazelle: Energy-efficient wearable analysis for running. IEEE
In this paper, the feasibility of detecting falls using a wrist- Transactions on Mobile Computing, 16(9):2531 – 2544, 2017.
worn device that acquires the users’ activities with only an [8] I. Guyon, S. Gunn, M. Nikravesh, and L. A. Zadeh. Feature Extraction-
accelerometer has been investigated. The problem of finding Foundations and Applications. Springer-Verlag Berlin Heidelberg, 1
edition, 2006.
a set of features that can accurately discriminate between falls [9] A.W. Whitney. A direct method of nonparametric measurement selec-
and ADLs has been tackled. Starting from a dataset of 2400 tion. IEEE Transactions on Computers, C-20(9):1100–1103, 1971.
files, 12 000 features have been extracted from each file. They [10] P. Pudil, J. Novovicova, and J. Kittler. Floating search methods in feature
selection. Pattern Recognition Letters, 15(11):1119–1125, 1994.
have been used as an input for a multitude of feature selection [11] I. Guyon, J. Weston, S. Barnhill, and V. Vapnik. Gene selection for
algorithms. SVM-RFE succeeded to reduce the number of cancer classification using support vector machines. Machine Learning,
features to 121 while keeping the best accuracy. The proposed 46(1-3):389–422, 2002.
[12] W. Yang, K. Wang, and W. Zuo. Neighborhood component feature
fall detection system consists of two stages. The first represents selection for high-dimensional data. Journal of Computers, 7(1):161–
a low complexity highly-sensitive algorithm to be embedded 168, 2012.
in the wearable device. When this algorithm detects a fall, it [13] H. Peng, F. Long, and C. Ding. Feature selection based on mu-
tual information criteria of max-dependency, max-relevance, and min-
sends raw data to a remote server for further accurate analysis. redundancy. IEEE Transactions on Pattern Analysis and Machine
Particularly, the 121 more discriminant features are extracted Intelligence, 27(8):1226 – 1238, 2005.
and supplied to a trained SVM classifier that predicts the class [14] B.D. Fulcher and N.S. Jones. hctsa: A computational framework for
automated time-series phenotyping using massive feature extraction.
of the activity. The proposed system achieved an accuracy Cell Systems, 5(5):527–531, 2017.
of 100% and its performance has also been validated in real [15] B.D. Fulcher, M.A. Little, and N.S. Jones. Highly comparative time-
world conditions. Currently, the system subjects to a long-term series analysis: the empirical structure of time series and their methods.
Journal of the Royal Society Interface, 10(83), 2013.
test where the objectives are to evaluate the average power [16] N. Noury, P. Rumeau, A.K. Bourcke, G. Olaighin, and J.E. Lundy. A
consumption of the embedded algorithm and to evaluate the proposal for the classification and evaluation of fall detectors. IRBM,
average number of events per day where accurate analysis is 29(6):340–349, 2008.