Hindawi

BioMed Research International

Volume 2020, Article ID 2167160, 17 pages
https://doi.org/10.1155/2020/2167160

Review Article
A Survey on Recent Advances in Wearable Fall Detection Systems

1 2
Anita Ramachandran and Anupama Karuppiah
1
Department of Computer Science & Information Systems, BITS, Pilani, Bangalore, India
2
Deptartment of Electrical & Electronics Engineering, BITS, Pilani, KK Birla Goa Campus, Goa, India

Correspondence should be addressed to Anupama Karuppiah; anupkr@goa.bits-pilani.ac.in

Received 30 May 2019; Revised 8 October 2019; Accepted 23 November 2019; Published 13 January 2020

Academic Editor: Antonella Gigantesco

Copyright © 2020 Anita Ramachandran and Anupama Karuppiah. This is an open access article distributed under the Creative
Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.
With advances in medicine and healthcare systems, the average life expectancy of human beings has increased to more than 80 yrs.
As a result, the demographic old-age dependency ratio (people aged 65 or above relative to those aged 15–64) is expected to
increase, by 2060, from ∼28% to ∼50% in the European Union and from ∼33% to ∼45% in Asia (Ageing Report European
Economy, 2015). Therefore, the percentage of people who need additional care is also expected to increase. For instance, per
studies conducted by the National Program for Health Care of the Elderly (NPHCE), elderly population in India will increase to
12% of the national population by 2025 with 8%–10% requiring utmost care. Geriatric healthcare has gained a lot of prominence in
recent years, with specific focus on fall detection systems (FDSs) because of their impact on public lives. According to a World
Health Organization report, the frequency of falls increases with increase in age and frailty. Older people living in nursing homes
fall more often than those living in the community and 40% of them experience recurrent falls (World Health Organization, 2007).
Machine learning (ML) has found its application in geriatric healthcare systems, especially in FDSs. In this paper, we examine the
requirements of a typical FDS. Then we present a survey of the recent work in the area of fall detection systems, with focus on the
application of machine learning. We also analyze the challenges in FDS systems based on the literature survey.

1. Introduction then applied to ML algorithms to detect falls. The objective

of this paper is to bring out an extensive literature survey of
Intelligent IoT-based ambient assisted living systems the recent work in the area of fall detection systems, with
(AALS) for the elderly have been a major research focus area focus on the application of machine learning to wearable
in recent times. According to the studies conducted by sensor-based approaches.
National Program for Health Care of the Elderly (NPHCE), The rest of the paper is organized as follows. First, we
elderly population in India will increase to 12% of the na- examine the desirable requirements of a wearable fall de-
tional population by 2025 with 8%–10% requiring utmost tection system. Then we present an overview of FDSs based
care. Application of machine learning in areas of AALS such on environmental sensors, vision-based systems, and
as fall detection, therefore, has the potential to have a huge wearable sensors. Subsequently, we dwell a little deeper into
public impact. Much work has been done in the area of fall the recent advances in FDS based on wearable devices. In
detection systems and the application of machine learning to this, we present a literature survey on threshold-based
such systems to enable fall classification, detection, and mechanisms and machine learning-based algorithms in fall
prediction. We have been working on the development of an classification and detection. It was observed that different
FDS which applies the biological profile of a subject to papers examined different performance parameters of ML
classify him into a risk category pertaining to his fall algorithms, and we present a summary of the results here.
probability. The three categories we have defined are high Lastly, existing literature shows that various biological,
risk, medium risk, and low risk. The categorization thus physiological, and environmental parameters affect a sub-
derived, along with parameters from a wearable sensor, is ject’s risk of falling. We present results from our study on the
2 BioMed Research International

various biological factors that impact the probability of fall and machine learning techniques on the data collected over a
in elderly persons. We conclude our paper with the chal- period of time.
lenges we observed in the existing fall detection solutions.

2. Requirements for Fall Detection Systems 3. Fall Detection Systems: An Overview

The current research on the wearable system design for a
The purpose of FDSs is automatic detection of falls and
geriatric healthcare monitoring system for detecting falls can
enabling of assistance by caregivers when required. FDSs
be broadly classified as follows:
primarily find its application in geriatric care because falls
are more frequent and severe in the elderly. Because such (i) Environmental sensing-based systems
systems would be used by the elderly, it is important for the (ii) Wearable sensor-based systems
manner in which falls are detected to be nonintrusive. For
(iii) Vision-based systems
example, a wearable sensor that is heavy or causes incon-
venience to the subject may not be a popular solution. Environmental sensing-based systems work on input
Power consumption of the apparatus should be mini- from sensors placed in the environment. Some examples of
mized because there is a possibility of the subject forgetting such systems are infrared sensor and acoustic and passive
to charge it. This entails that the sensors and the network infrared (PIR) motion detector-based systems. Infrared
design of the system have to be optimized for power sensor-based systems sense certain characteristics of an ROI
consumption. to detect a fall, either by emitting or detecting infrared
An FDS should not restrict the subjects’ mobility, and radiation or by measuring the heat being emitted by an
they should be free to move around the area they want to. object and detecting motion. Microphone-based FDSs use
Camera-based or infrared- (IR-) based systems [3] may acoustic signals to detect sound in a room, locate the source,
restrict the subject to be within a certain region of interest and classify it as a fall or a nonfall condition. Motion de-
(ROI), while wearable sensor-based systems do provide tector-based systems identify falls based on detection of
more mobility to the subjects. An IR- or camera-based motion within an ROI. As an example of environmental
system may employ multiple techniques to increase the sensing-based systems, Taramasco et al. [4] explain one
coverage, such as sweep coverage-based deployments, but which uses very-low-resolution thermal sensors located on
this would mean increasing the cost of the system. two horizontal planes along the floor, for classifying falls.
Another important requirement of an FDS is that it The algorithms compared were three recurrent neural
should be able to distinguish between falls and activities of networks (RNN)—long short-term memory (LSTM), gated
daily life (ADLs) or near-fall conditions. This is to prevent recurrent unit, and Bi-LSTM, with Bi-LSTM giving an ac-
triggering of premature action in the case of false positives curacy of 93%. In [5], the authors rely on the existing
and adequate care not being provided in the case of false wireless infrastructure for detecting falls. The channel state
negatives. Therefore, the accuracy of detection is an im- information (CSI) from WiFi deployments in a given area
portant factor. was used for this purpose. Support vector machines (SVMs)
The manner in which triggers are generated on fall and Random Forest algorithms were applied to CSI matrix
detection is also important. Caregivers may be located re- for device-free fall detection. The experiments were con-
motely, and an FDS should support providing notifications ducted in controlled environments with falls simulated by
to remote personnel. The notification could be a short in- one subject at a time. The advantages of this system are that it
dication like a message or could be a descriptive account is nonintrusive, which means it does not expect the subjects
from the ROI, by way of images captured by a camera. The being monitored to be wearing or carrying any device.
former is easier to implement, while the latter has the ad- However, it may be high on false positives because of the
vantage that it can provide a clear picture of the impact to the impact of ambient parameters such as heat; for example,
observer before deciding a course of action. since the system relies on thermal sensors, its accuracy is
Latency is another factor that should be considered in affected by the presence of other exothermic devices such as
the design of FDSs. Delay in the detection of falls and that heaters. This study also observed that when multiple people
between the detection of falls and notification of the care- are in the region under experimentation, their combined
giver should be minimized for the FDS to be effective. This movements influenced the accuracy of the outcome. Also, in
implies that techniques used for fall detection should be this paper, false positives, though at acceptable levels, were
delay-sensitive. It also implies that the network design high, because of the small dataset used for training and
should provide high quality of service for data packets that testing. Ciabattoni et al. [6] describe a system consisting of a
are generated as a result of fall detection, in comparison with low-cost mobile robot and an RGB camera, deployed in a
messages for keepalive or periodic reporting of sensor room, that produces real-time video stream. The robot
readings. employs deep learning techniques for user detection, po-
It is also desirable for an FDS to keep track of a subject’s sition estimation to detect a fall, photo and video capture,
biological parameters and fall history so as to have the and interfacing with a Bot telegram. The accuracy reported
capability of predicting falls before their occurrence. This for fall detection was 93%.
would involve reporting of the biological parameters by the Vision-based systems do not perform any parameter
sensor nodes periodically and application of data analytics monitoring of the subjects; instead, they rely on image
BioMed Research International 3

processing techniques on the video frames or images cap- the subject, such as a wrist band. The parameters monitored
tured by cameras around the ROI. ML algorithms may be by such systems include the following: heart rate variability
applied over image processing techniques to enable more (HRV), electrocardiogram (ECG), pulse oximetry (SPO2),
accurate fall detection. In [7], convolutional neural networks and kinematic attributes measured by accelerometers, gy-
(CNNs) are trained on different datasets of optical flow roscope, and magnetometer. The data reported by wearable
images. This helps the network to detect different actions. sensors are fed as inputs to a threshold-based system or as
Transfer learning is then applied from action recognition to feature sets to a machine learning-based system to classify
fall detection. The experiment was conducted on 3 different and detect falls. Wearable sensor-based systems are less
datasets and reported an accuracy of above 95% in all cases. expensive, have low power consumption which reduces the
However, one stated drawback of this approach is that it is overheard on charging the system, and are usually in the
susceptible to inaccuracies resulting from ambient lighting form of a band that can be worn around the wrist or thigh,
changes. Zerrouki et al. [8] detail a comparative study of ML which is less susceptible to being separated from the subjects.
algorithms for fall detection with video sequences during Kaewkannate and Kim [14] provide a summary of com-
different daily and fall activities as input. They compared parison between four wearable wrist-band style devices
Naı̈ve-Bayes, k-nearest neighbors (kNN), neural network, currently available on the market with respect to their
and SVM algorithms and concluded that SVM performed features and cost. The power consumption of wearable
best among these, with respect to accuracy, sensitivity, devices is dependent on the device configuration, type of
specificity, precision, recall, F-measure, and area under the sensors, and communication technologies used. Oletic and
curve (AUC). Anishchenko [9] applies deep learning and Bilas [15] give an analysis of total power consumptions for
transfer learning techniques on data generated by surveil- different operating scenarios, for certain configurations of
lance cameras under realistic conditions, to detect falls. The wearable devices. There are also systems where the sensors
objective was to overcome the setbacks of simulated datasets are worn not on the wrist but on other parts of the body, or
collected under controlled environments. Bhandari et al. are embedded within a smartphone. Smartphone-based
[10] analyze video frames for fall detection in 3 step- systems expect the subject to charge their device as required
s—finding out interest points using Shi-Tomasi algorithm, and carry it with them to enable fall detection, which are not
calculating the distance between interest points from optical good prerequisites for geriatric healthcare systems.
flow calculation with Lucas–Kanade algorithm, and esti- There have been some papers that propose an end-to-
mating the speed and direction of motion to conclude end IoT-based system for fall detection. An example of such
whether a fall has taken place or not. The method is another a system in indoor environments is presented in [16]. This
example of application of unsupervised learning in fall design makes use of low-power wireless sensor networks,
detection. The accuracy reported is 95% for nonfall activities smart devices, big data, and cloud computing. A 3D-axis
and 96.67% for fall activities. In [11], data were collected accelerometer embedded into a 6LowPAN device wearable
using a Kinect camera and a triaxial accelerometer. The collects movement information and applies decision tree
video input was used to classify the accelerometer data into algorithms to detect falls.
falls or nonfalls, in the training phase. Time and frequency
domain analysis was performed on the data—the former 4. Relevance of Machine Learning in
using SVM and the latter using lifting wavelet transform. It is Fall Detection
reported that the frequency-based analysis exhibited an
accuracy of 100% in detecting falls, while the SVM-based Machine learning is a technique that applies mathematical
time domain analysis reported 98.31%. Yanfei et al. [12] models on datasets to analyze, classify, and discover new
analyze feeds from a Kinect camera and processes point cloud meanings from them, to enable the system to learn auto-
images to detect falls and reduce false positives. More recently, matically from the training it received. A model trained on a
the application of deep learning techniques to fall detection given dataset is capable of interpreting new input data and
has become an active area of research. Lu et al. [13] use video predicting outcome variables. Machine learning helps
feeds from ambient data and applies CNN and LSTM for achieve certain amount of task and decision automation for
feature extraction. The use of 3D CNN enables extraction of various domains. There are 3 types of machine learning
motion features from temporal sequence, in addition to approaches—supervised machine learning, unsupervised
spatial information, while LSTM-based visual attention machine learning, and reinforcement learning. In supervised
mechanism is used to locate the regions of interest. The machine learning, the training is done based on labelled
authors note that this approach works well on small datasets input data. For every input data, there is a corresponding
and that analysis of long motion sequences using this scheme outcome variable. Therefore, the input data is classified a
will increase the computational costs of the system. While priori, and when there is a new data point, it can be mapped
vision-based systems provide accurate details of abnormal to one of the defined classes. There are two approaches to
conditions to a remote caregiver via images or video feeds, supervised machine learning—regression and classification.
they tend to be more expensive and computationally intensive Unsupervised learning is when the algorithm itself tries to
and require higher processing time, in addition to being a find a pattern within a given dataset. Reinforcement learning
subtle intrusion to the privacy of the subjects. allows the system to adapt its behaviour based on feedback
In wearable device-based FDSs, the sensors used for fall or rewards from the environment [17]. In each of the cat-
detection are embedded within a wearable device worn by egories of machine learning approaches, there are multiple
4 BioMed Research International

algorithms. A simplified taxonomy of machine learning In [18], an algorithm based on first differences and first
algorithms is given in Figure 1. derivatives of sum of accelerometer readings along X, Y, and
Due to the advances in the field of medicine and changes Z directions is described. This algorithm is real-time and
in population demographics, geriatric healthcare has gained reliable and was capable of distinguishing jerky movements
a lot of significance. In the recent years, there has been from falls. Wu et al. [19] build a system with triaxial ac-
widespread application of technology in the healthcare celerometer and proposes an algorithm based on thresholds
domain. Some applications of machine learning in geriatric of sum acceleration and rotation angle information. This
healthcare include monitoring of vitals, analysis of sleep combines threshold values of acceleration with quaternion
patterns, behavioural studies, and fall detection—the fun- rotation, to conclude whether a fall has taken place or not.
damental objective of these applications being to detect and/ The sensitivity and specificity of this algorithm are reported
or predict abnormalities. Machine learning in fall detection to be better than pure threshold-based systems.
helps in intelligently detecting falls based on a subject’s In systems where only accelerometer is used, the ac-
activity patterns. It may be easy for a fall detection system to curacy of threshold-based fall detection may not hold true in
raise an alarm whenever a change in activity pattern is all conditions. Sensor fusion techniques have been experi-
observed; however, this would result in excessive alarms mented in some cases, where sensors other than acceler-
being triggered falsely. If the fall detection system is designed ometers have been applied. For example, in [20], the author
to be conservative in raising alarms, then it may not raise makes use of an accelerometer combined with an HRV
alarms when actual falls occur. Hence, it becomes important sensor. The signals from the accelerometer are analyzed for
that false positives and false negatives in a fall detection abnormalities in movements. The signals from HRV sensor
system are minimized, and the self-learning capability of are analyzed for abnormalities in heart rates induced by
machine learning algorithms plays a vital role here. Other anxiety at the time of fall. Both the analyses are threshold-
performance parameters of machine learning algorithms based and performed independently, and a fall is concluded
include specificity, sensitivity, and recall. Algorithms for fall to have occurred if both report the occurrence of a fall. The
detection work on datasets generated by camera, environ- accuracy of the ability to distinguish falls from identical
mental sensors, or wearable sensors, and the objective of activities in this study was reported to be between 96% and
research in this area is to improve the performance pa- 100%. In [21], a three-step algorithm is proposed based on
rameters of the algorithms when applied to fall detection. activity intensity analysis, posture analysis, and transition
In the context of fall detection, the outcome variables of analysis, with signals reported by accelerometer and gyro-
machine learning algorithms for binary classification would scope. Results show sensitivity of 91% and specificity of 92%,
be falls or ADLs. ADLs include various postures such as in being able to separate falls from ADLs and near-fall
sitting, standing, lying down, and slow or fast transitions conditions. In addition to the application of ML algorithms
between these activities. In order to ensure that the outcome to multiple wearable sensor node readings, there have also
variables predicted by a machine learning system are correct, been experiments on building context around the sensor
data cleaning and preprocessing are performed on the fall readings from the surroundings. In [22], the authors con-
dataset. Subsequently, feature extraction is performed so as sider acceleration, pulse, and oxygen saturation of the
to shortlist the right set of features that will characterize the subject via an Android phone, combined with context
dataset, and this set of features is used for creating a trained awareness being incorporated by PIR motion, door contact,
model (see Figure 2). For research in machine learning- pressure mats, and power usage detectors. Sensor fusion
based fall detection systems, new datasets for falls and ADLs among these disparate sources is achieved by Bayesian
are created by experimentations in controlled environments, networks to perform fall detection.
or publicly available datasets are used for analyses. Apart Chen et al. [23] use microelectromechanical system
from the direct application of machine learning algorithms (MEMS) accelerometers for fall detection. It says that in the
to such datasets to detect a fall or an ADL, existing literature experiments that the authors conducted, setting thresholds
also shows various techniques for feature extraction and separately for the 3 axes did not work well. Hence, the norm of
nullifying the errors induced by external factors such as the 3 axes was taken, and a threshold was set for the norm. The
misplacement of sensors. authors note that there is scope for improvement in perfor-
mance if the design is customized, since the acceleration
5. Fall Detection Using Wearable Sensors profiles vary from person to person depending on his physique.
Tsinganos Skodras [24] compiles various sensor fusion
There are two approaches to fall detection using wearable techniques applied for fall detection and also summarizes
sensors—threshold-based systems and machine learning- their performance results in the context of fall detection. In
based systems. most cases highlighted in this study, the sensors used were
accelerometers and gyroscopes. In our research too, we find
5.1. Threshold-Based Wearable Fall Detection Systems. that although there are cases where multiple sensors are
Threshold-based systems have been a widely researched used, most of the research studies use only IMU-based
area. The focus of such research has been on multiple as- sensors. The lack of application of threshold-based mech-
pects, such as ability to detect falls and classify falls from anisms with sensor fusion techniques could be because of the
ADLs and near-fall conditions and sensor fusion of readings limited capabilities that a pure threshold-based system
from multiple sensor nodes. presents in decision making under dynamic uncertainties.
BioMed Research International 5

kNN implement and are computationally less intensive. However,

Naïve Bayes there are some drawbacks in such system, as detailed above.
Supervised Decision trees
learning Linear regression
Setting a wrong threshold may lead to lower accuracies in fall
Support vector machines detection. Also, the thresholds themselves may be different
Neural networks across subjects because of differences in their ADL patterns.
The alternative machine learning-based approaches use
Machine K-means clustering supervised or unsupervised algorithms on large datasets to
learning Unsupervised Association rules train classifiers and thus build the ability to recognize a fall.
learning Hidden Markov model
algorithms
Neural networks
The model thus trained can be used to detect falls for other
input data. This is described below.

Reinforcement 5.2. Machine Learning-Based Wearable Systems for Fall

learning Detection. While threshold-based systems have been pop-
ular because of its low computational overhead, it could be
Figure 1: Taxonomy of machine learning algorithms.
prone to more false positives and false negatives, given that
the thresholds themselves may be affected by various factors.
Threshold-based algorithms are typically designed to As a result, machine learning algorithms for fall detection
minimize computational overhead. However, the threshold have been a much researched area. There has been extensive
values may vary with the placement of sensors and indi- research into the efficiencies of various machine learning
vidual activity patterns. In [25], the authors state that the techniques for fall detection. For example, de Quadros et al.
thresholds and positioning of sensors impact the accuracy of [28] compare threshold-based mechanism and machine
fall detection. They conducted experiments by placing the learning-based mechanisms for fall detection applied on
sensors on the shoulder, waist, and foot of the subjects. A data generated by accelerometer, gyroscope, and magne-
series of observations were made by adjusting both the tometer. The paper concludes that the machine learning-
thresholds of acceleration for fall detection and the place- based mechanism yielded much better results than the
ment of sensors for improving the performance of the threshold-based solutions.
system. Sensor placement on the waist resulted in lesser false Machine learning-based techniques differ from each
positives than that on the shoulder and foot. other in multiple factors—the feature set used, sensors
In order not to rely only on numerical thresholds set on employed, placement of sensors, algorithms applied, dataset
acceleration, Kostopoulos et al. [26] consider the subject’s used, performance parameters monitored, and so on. There
rebound and residual movements in the postfall phase for are studies that focus primarily on the application of su-
fall detection, in addition to a threshold-based analysis of pervised learning suited to fall detection using wearable
accelerometer data. The maximum and minimum thresholds sensors with good performance results. Özdemir and Bar-
of acceleration over a short duration of time are used to shan [29] take into account features of acceleration, rate of
determine occurrence of a fall. The rebound is calculated as turn, and the strength of the Earth’s magnetic field along
the difference between this max and min thresholds. The three perpendicular axes to detect a fall. The algorithm used
classification of fall is done subsequently, again based on the distinguishes falls from ADLs using six machine learning
acceleration values. The postfall analysis is used to determine techniques: kNN, least squares method (LSM), SVM,
the impact of the fall, which in turn decides whether an Bayesian decision making (BDM), dynamic time warping
alarm raised to the caretaker has to be cancelled or not. (DTW), and artificial neural networks (ANNs). The factors
There are also systems that have been proposed which monitored for performance comparison are sensitivity,
work based on sensors embedded in mobile phones. A specificity, accuracy, true positives, true negatives, false
disadvantage with such systems is that the subject should positives, and false negatives. The observation was that kNN
remember to charge and carry mobile phones as a pre- and LSM methods do not miss any falls and hence were
requisite for the system to detect falls. Another drawback is concluded as reliable classifiers. Choi et al. [30] compare ML
that not all mobile phones may come equipped with the algorithms for fall detection using a single node and two
required sensors. There is also the possibility of false posi- nodes. The results reported an accuracy of 99.4% for clas-
tives caused by mobile phone drops. Chaitep and Chawachat sification, with single node consisting of a 3-axis acceler-
[27] proposed a threshold-based detection method which ometer and a 2-axis gyroscope, worn at the chest level. With
makes use of G-force values derived from accelerometer 2 nodes, a second node with a 3-axis accelerometer and a 1-
readings to identify falls and smartphone drops. The algo- axis gyroscope was worn on the thigh, in addition to the first
rithm consists of 3 phases—checking for a smartphone drop, node worn at the chest. The accuracy in this case was 99.8%.
detecting a fall, and fall confirmation. Naı̈ve-Bayes classifier gave the best results in both cases. In
A snapshot of recent research on fall detection using [31], the dataset used was generated from accelerometer and
threshold-based wearable systems is presented in Table 1. In gyroscope, placed at the waist level. Feature extraction was
threshold-based systems, a maximum value for the param- performed using windowing technique, feature selection
eters read by sensors such as an accelerometer or gyroscope is using rank-based system, and classification using Naı̈ve-
predefined. A measured value beyond this threshold is an Bayes, LSM, ANN, SVM, and kNN algorithms. kNN, ANN,
indicator of an abnormal event. Such systems are simple to and SVM had the best performance results compared to
6 BioMed Research International

Data Data Feature Model

Training Inferences
acquisition preprocessing extraction evaluation

Figure 2: Flow diagram for machine learning-based model building.

Table 1: Threshold-based systems for fall detection using wearable devices.

Reference Year Dataset used Sensors used Sensor placement Methodology Observed performance
Three-step algorithm
based on activity
intensity analysis,
Generated
Accelerometer, posture analysis, and Sensitivity � 91%
[21] 2009 from Chest, thigh
gyroscope transition analysis, based Specificity � 92%
experiments
on signals reported by
accelerometer and
gyroscope
Algorithm based on first
differences and first
Generated
derivatives of sum of Algorithm is reliable,
[18] 2011 from 3D accelerometer Not specified
accelerometer readings simple, and real time
experiments
along X, Y, and Z
directions
Better sensitivity and
Generated Quaternion algorithm
specificity than
[19] 2015 from Accelerometer Waist using sum acceleration
threshold-based
experiments and angle information
algorithms
Analysis of signals from
Generated accelerometer for Accuracy � 96% to 100%
Accelerometer, HRV
[20] 2015 from Not specified movement detection and (depending on the type
sensor
experiments HRV sensor for stress of movement)
detection
Threshold-based Accuracy � 100%
method, applied to Specificity � 91.1%
Generated 3-Axis accelerometer, 3-
Shoulder, waist, and acceleration and Euler’s (shoulder), 100% (waist),
[25] 2017 from axis gyroscope, 3-axes
foot angle (yaw, pitch, and 78.5% (foot)
experiments magnetometer
roll), run on a mobile Sensitivity � 100% (for
phone all three placements)
Specificity � 72% when
3-Phase detection based
Generated compared to the
on thresholds to identify
[27] 2017 from G-force sensor Smartphone specificity of 31% with 2-
falls and smartphone
experiments phase threshold-based
drops
algorithm
Generated Waist + network of Signal analysis based on
MEMS accelerometers,
[23] 2017 from fixed motes within the threshold-based Not specified
RF signals
experiments home methods

LSM and Naı̈ve-Bayes. Results show an accuracy of 87.5%, terms of recall analysis. Machine learning algorithms, much
sensitivity of 90.70%, and specificity of 83.78%, for kNN. like threshold-based techniques, have also been applied to
Jefiza et al. [32] use backpropagation neural network sensors integrated with mobile phones. In [34], a method for
(BPNN) for fall detection, with data collected from 3-axis fall detection an classification by machine learning using
accelerometer and gyroscope, and reported an accuracy of mobile phones is proposed; the features used were acceler-
98.182%, precision of 98.33%, sensitivity of 95.161%, and ation and the algorithms compared were SVM, sparse mul-
specificity of 99.367%. Hossain et al. [33] also attempt to tinomial logistic regression (SMLR), kNN, decision trees, and
distinguish falls from ADLs and compares SVM, kNN, and Naı̈ve-Bayes. Results showed that both SVM and SMLR were
complex tree algorithms applied on data generated by ac- able to identify a fall with 98% accuracy and classify the type of
celerometers. The paper compared the performance of these fall with 99% accuracy.
algorithms with respect to accuracy, precision, and recall, on Despite supervised learning techniques finding more
ADLs and four types of falls (forward, backward, right, and application in fall detection, as detailed above, the appli-
left). It was observed that the accuracy and precision of SVM cation of unsupervised learning is also not uncommon. In
were the highest, while complex tree performed better in fact, Lee et al. [35] claim that supervised learning has
BioMed Research International 7

deficiencies in terms of abnormality detection and activity It applies Kalman filter to preprocess the raw data for noise
classification. The authors hence experimented with unsu- reduction and Bayes network classifier for fall detection. The
pervised learning for fall detection. Their algorithm creates algorithm presented an ability to distinguish simulated falls
an activity probability model of a subject’s past activity from ADLs with an accuracy of 95.67%, sensitivity of 99.0%,
information from accelerometer readings. This model is and specificity of 95.0%. Zhao et al. [41] also apply a win-
then used to determine whether an activity is abnormal or dowing technique to real-time data obtained from a triaxial
not. The advantage of this approach is that it achieves a gyroscope. The data were divided into a set of consecutive
certain level of personalization in fall detection since the and partially overlapping windows. Three time domain
probability density function which is central to activity features (resultant angle change, maximum resultant an-
comparison is developed per subject. gular acceleration, and fluctuation frequency) were extracted
One of the observed drawbacks of wearable sensors is from the data windows. Decision tree classifier was then
that the accuracy of fall classification and detection is im- used to classify each window as a fall or a nonfall event. The
pacted by the placement of the sensors. In [36], the authors detection algorithm gave accuracy of 99.52%, precision of
generated a dataset with accelerometer and gyroscope, worn 99.3%, and recall of 99.5%. Another recent research [42]
around the waist, and applied SVM, boosted and bagged compares the performance of 4 algorithms—ANN, kNN,
decision trees, kNN, k-mean, and hidden Markov model quadratic SVM, and ensemble bagged tree—in two steps.
(HMM). It was observed that fine kNN produced an ac- First, only acceleration and angular velocity data are used.
curacy of 99.4%. Yu et al. [37] attempt to reduce errors Then, new features that improve the performance of the
caused by incorrect sensor positions and details an HMM- classifier are extracted from the power spectral density of the
based fall detection system for the same. Sensor orientation acceleration. The accuracy of the algorithms is observed to
calibrations are applied on HMM classifiers to resolve issues have increased after applying feature extraction techniques.
arising out of misplaced sensor (3-axis accelerometer) lo- The objective of [43] was to test the impact of optimal
cations and misaligned sensor orientations. This paper re- feature selection on the accuracy of fall detection. The
ports sensitivity of 99.2% on an experimental dataset, 100% features of accelerations in different parts of the body are
for a real-world fall dataset. collected through wearable devices. Bayesian framework was
Guvensan et al. [38] focus on energy efficiency in fall applied to select the optimal features from the data generated
detection. A combination of threshold-based method and by the wearable devices, and the weight of each feature was
ML-based algorithms—K-Star, Naı̈ve-Bayes, and J48—was calculated, after which training was done based on the
applied to data generated from a 3D accelerometer attached optimal feature set. It was observed that improved classi-
to a smartphone. The algorithm employed three tiers—a pre- fication led to better accuracy, sensitivity, and specificity
elimination tier to apply initial filtering, a double thresh- when compared to Naı̈ve-Bayes and C4.5 classifications.
olding tier to detect harsh falls and physical activities oc- Tsinganos and Skodras [44] analyze accelerometer data to
curring at a slow pace, and a machine learning tier to extract a set of 14 features across time domain, statistical
recognize slow falls and fall-like events using ML techniques. measures, and continuous wavelet transform. ENN was
Energy saving was reported to be 62% compared with ML- applied to remove outliers and then trained using kNN
only techniques, while the accuracy with the hybrid model classifier to distinguish falls from ADLs. To negate indi-
was 93%. The hybrid approach was superior with respect to vidual-specific patterns, personalization was applied by
sensitivity and performed comparable to the threshold- appending the features of ADLs to the training dataset. The
based and ML-based approaches in terms of specificity. other models used for comparison were ANN, SVM, and J48
There have also been various techniques to improve the decision tree. The performance of kNN was the highest.
performance of the algorithms used for fall detection, by In [45], the authors propose EvenT-ML, in which a fall
optimizing preprocessing of data, influencing the feature event was aligned with three stages of falls (preimpact,
selection/extraction, and applying ensembles to fall detection. impact, and postimpact) using a finite state machine. The
In [39], an example of a system that applies intelligent experimentation was based on data generated by acceler-
preprocessing to the data before applying machine learning ometers, and features were extracted from each phase.
for fall detection is given. In this, the authors apply a Classification and regression tree (CART), kNN, logistic
windowing technique to divide the sensor signals into regression (LR), and SVM were used to train the classifiers.
windows or time segments. Classification algorithms were The authors observe better results for EvenT-ML than the
then applied to each window, to determine whether the commonly used data segmentation techniques of fixed-size
activity in that window corresponded to a fall. In this, two nonoverlapping sliding window (FNSW) or fixed-size
Sun SPOT sensors attached to the chest and thigh were used, overlapping sliding window (FOSW), where feature ex-
and it was observed that Naı̈ve-Bayes classifier gave 100% traction is performed on all data segments. The finite state
accuracy and 87.5% sensitivity. Other algorithms used were machine ensures that feature extraction gets executed only
SVM, OneR, C4.5 (J48), and neural networks. The objective when the subject is in the active state, and this reduces the
of [40] was to distinguish falls from ADLs. In this study, the computational complexity of this method.
wearable fall detection system comprises a wearable motion Recent research has also focused on the application of
sensor and a smartphone. The system works by analyzing not ensembles to fall detection. Hsieh et al. [46] use a combi-
instantaneous values of acceleration and angular velocities, nation of threshold-based and knowledge-based approach
but by applying sliding windows to analyze streams of data. based on SVM, on data from a triaxial accelerometer, to
8 BioMed Research International

detect a fall event. Absolute falls and ADLs are detected Our approach using SVM is also supported by the results
using thresholds on acceleration. In order to distinguish falls which show that activity recognition can be increased with
from ADLs in cases where the peak values of acceleration the accuracy level as high as 99%, when the combination of
overlap, a knowledge-based approach is applied. Using this acceleration, angular velocity, and orientation parameters
approach, sensitivity, specificity, precision, and accuracy are utilized compared to using them separately.
were over 99%. Genoud et al. [47] propose a system for soft The application of deep learning techniques for fall de-
fall detection using ML in wearable devices. The feature sets tection using wearable devices has been an area of recent
used were linear acceleration and gyroscope readings, and interest. Musci et al. [54] describe an RNN model with LSTM
the algorithms compared were decision tree, decision tree blocks on data generated by 3D accelerometers for fall de-
ensemble, kNN, and multilayer perceptrons (MLP). The tection. The paper observes that though it is difficult to
experiments showed that decision tree ensemble out- distinguish high dynamic activities from falls, the approach
performed the results obtained by the other algorithms. In described achieves a better overall classification. Fakhrulddin
[48], a comparison of Naı̈ve-Bayes classifier, decision trees, et al. [55] apply CNN to streaming time series accelerometer
random forests, random committee, and lazy learning (IBk) data, collected from body sensor networks (BSN), for fall and
algorithms for activity detection was done. This used data nonfall situations. In [56], a method of applying CNNs with 3
generated by motion, acceleration, or inertial sensors em- convolutional layers on data generated by accelerometers is
bedded in a smartphone worn by the subjects. Naı̈ve-Bayes described. The activation function in each layer is rectified
classifier performed reasonably well for a large dataset, with linear unit (ReLU). However, the study indicates that scarcity
79% accuracy, and it was also the fastest in terms of building of public datasets based on accelerometer and gyroscope
the model taking only 5.76 seconds. Random forest was makes it challenging to develop deep learning solutions for
better in terms of both accuracy and model building time, this kind of data. Also, deep learning techniques require high
with 96.3% accuracy and 14.65 seconds model building time. computational processors, which may not be well suited to the
k-Means clustering performed poorly with 60% classifica- constrained nature of wearable devices. Torti et al. [57] detail
tion accuracy and 582 seconds model building time. Kao the implementation of RNN architectures for constrained
et al. [49] use an ensemble of spectrum analysis, with GA- embedded devices on a microcontroller unit (MCU), for fall
SVM, SVM, and C4.5 classifiers. The sensor readings were detection with triaxial accelerometers. The work also provides
from 3-axis accelerometers. The best results were given by an abstraction of formulas for memory, computing power,
GA-SVM, with an accuracy of 94.1%, sensitivity of 94.6%, and power consumption for the embedding of a generic RNN
and specificity of 93.6%. Jahanjoo et al. [50] propose a fall architecture on an MCU.
detection algorithm based on data from 3-axis accelerom- Li et al. [58] describe fusion of data from a triaxial ac-
eters, using PCA for dimension analysis and a multilevel celerometer, a micro-Doppler radar, and a depth camera.
fuzzy (MLF) min-max neural network, and compared the The paper analyses the impact of sensor fusion on the
performance with MLP, kNN, and SVM. Using only 5 di- performance of classifiers. The classification accuracy
mensions of features, MLF performed better than the other attained by means of this fusion approach improves by
algorithms in terms of sensitivity, while the specificity was 11.2% compared to radar-only use and by 16.9% compared
comparable for all four algorithms. to the accelerometer. It was also observed that fusing in-
Hussain et al. [51] apply kNN, SVM, and Random Forest formation from three sensors increases the classification
algorithms to not just detect falls, but also to identify the accuracy to 86.9% with the quadratic-kernel SVM classifier,
falling pattern and identify the activity that may have caused and up to 91.3% using an ensemble classifier. Some studies
the fall. It is reported that the fall detection accuracy was have applied sensor fusion in combination with deep
highest for kNN, while the accuracy for recognizing different learning techniques for fall detection. For example, Dawar
activities was highest for random forest. Yet another re- and Kehtarnavaz [59] use CNN-based sensor fusion system
search [52] attempts to find a correlation between sampling to detect falls and ADLs. Signals from depth camera and
rate and performance accuracy of machine learning models. wearable sensors (acceleration and angular velocity) are fed
In this paper, the authors compare the performance of SVM, as inputs into separate CNNs. The algorithm then fuses the
Naive-Bayes, kNN, and decision trees with various sampling scores generated by these two CNNs to produce a classifi-
rates of sensors. It is concluded that with sampling rates of cation. Zhou et al. [60] also describe an approach of using 2
11.6 Hz and 5.8 Hz, SVM and radial basis function gives CNNs for initial processing of two types of inputs and
accuracies of 98% and 97%, respectively. The research subsequent merging of the results of the two CNNs to
suggests that a sampling rate of 22 Hz is sufficient for most produce the final detection results. In this, the inputs are
machine learning models to provide an accuracy of 97%. obtained from radar signals to detect velocities, acceleration
Hakim et al. [53] propose a hybrid approach between of human body parts, and images from optical camera.
threshold and ML-based fall detection algorithms. In this, a Other works along similar lines include references [61, 62].
threshold-based algorithm is implemented to detect falls Table 2 shows a snapshot of the recent research in the
while a supervised machine learning algorithm is used to application of machine learning to fall detection using
classify ADL. Data were collected from IMU sensors in a wearable systems. All experiments in the presented literature
smartphone. Four different classification algorithms were are based on analysis of public datasets or falls simulated
used for detection and classification: SVM, decision trees, under controlled environments. The table qualitatively
kNN, and discriminant analysis. compares different algorithms for fall detection and
BioMed Research International 9

Table 2: Machine learning-based systems for fall detection using wearable systems.
Sensor placement (if
Reference Year Dataset used Sensors/dataset used Methodology Observed performance
wearable system)
Accuracy of
classification � 99.8%,
Comparison of ML
with 2 nodes—one on
3-Axes accelerometer, algorithms for fall
[30] 2011 UCI dataset Chest, thigh the waist and one on
2-axis gyroscope detection using single
the knee
node and two nodes
Naı̈ve-Bayes classifier
gave best results
Support vector
machines and
regularized logistic
Comparison of SVM,
regression were able to
SMLR, Naive Bayes,
identify a fall with 98%
Generated from decision trees, kNN,
[34] 2012 Accelerometer Mobile phone accuracy and classify
experiments and regularized logistic
the type of fall (trips,
regression for fall
left lateral, slips, right
detection
lateral) with 99%
accuracy. Naı̈ve-Bayes
reported least accuracy
k-NN classifier and
Comparison of k-NN
Accelerometer, LSM gave above 99%
Generated from 6 different positions on classifier, LSM, SVM,
[29] 2014 gyroscope, for sensitivity,
experiments the body BDM, DTW, and
magnetometer specificity, and
ANNs algorithms
accuracy
Accelerometer data
from wearable sensors
to generate alarms for
falls, combined with Provides statistical
Generated from context recognition information regarding
[22] 2014 Accelerometer Mobile phone
experiments using sensors in an the fall risk probability
apartment, for for a subject
inferring regular
ADLs, using Bayesian
networks
Naive Bayes classifier
performs reasonably
well for a large dataset,
with 79% accuracy, and
it is fastest in terms of
Comparison of Naive
building the model
Bayes classifier,
taking only.5.76 seconds
decision trees, random
Random forests are
forests, classifiers
Publicly available better in terms of both
Accelerometer, based on ensemble
[48] 2015 activity recognition Smartphone accuracy and model
gyroscope learning (random
dataset building time, with
committee), and lazy
96.3% accuracy and
learning (IBk)
14.65 seconds model
algorithms for activity
building time. k-Means
detection
clustering performs
poorly with 60%
classification accuracy
and 582 seconds model
building time
Comparison of
decision tree, decision Decision tree ensemble
Generated from tree ensemble, kNN, was able to detect soft
[47] 2016 3-Axes accelerometer Not specified
experiments neural networks, MLP falls at more than
algorithms for soft fall 0.9 AUC
detection
10 BioMed Research International

Table 2: Continued.
Sensor placement (if
Reference Year Dataset used Sensors/dataset used Methodology Observed performance
wearable system)
k-NN, ANN, SVM had
the best
Comparison of Naı̈ve-
accuracy—results for
Accelerometer, Bayes, LSM, ANN,
[31] 2016 MobiFall dataset User’s trouser pocket kNN:
gyroscope SVM, kNN algorithms
Accuracy � 87.5
for fall detection
Sensitivity � 90.70
Specificity � 83.78
Generated from Threshold-based
[26] 2016 3-Axis accelerometer Smartwatch Accuracy � 96.01%
experiments analysis of acceleration
Kalman filter for noise
With Kalman filter
reduction, sliding
Generated from Accelerometer, Accuracy � 95.67%,
[40] 2017 Vest window, and Bayes
experiments gyroscope Sensitivity � 99.0%
network classifier for
Specificity � 95.0%
fall detection
Energy saving � 62%
compared with ML-
only techniques
Sensitivity � 77%
(thresholding only),
Combination of 82% (ML only), 86%
threshold-based and (hybrid)
Generated from
[38] 2017 3D accelerometer Smartphone ML-based Specificity � 99.8%
experiments
algorithms—K-Star, (thresholding only),
Naive Bayes, J48 98% (ML only), 99.5%
(hybrid)
Accuracy � 88.4%
(thresholding only),
90% (ML only), 92.75%
(hybrid)
Combination of Using a knowledge-
threshold-based and based algorithm:
Generated from knowledge-based Sensitivity � 99.79%
[46] 2017 3-Axes accelerometer Waist
experiments approach based on Specificity � 98.74%
SVM to detect a fall Precision � 99.05%
event Accuracy � 99.33%
GA-SVM gave best
Spectrum analysis,
results with
Generated from combined with GA-
[49] 2017 3-Axes accelerometer Smartwatch Accuracy � 94.1%
experiments SVM, SVM, and C4.5
Sensitivity � 94.6%
classifiers
Specificity � 93.6%
Comparison of Multilevel fuzzy min-
multilevel fuzzy min- max neural network
[50] 2017 MobiFall dataset 3-Axes accelerometer Not specified max neural network, gave best results with
MLP, KNN, SVM, Sensitivity � 97.29%
PCA for fall detection Specificity � 98.70%
Sensor orientation
calibration algorithm
5 locations on the to resolve issues arising Sensitivity � 99.2%
FARSEEING upper body - neck, out of misplaced (experimental dataset),
[37] 2017 3-Axes accelerometer
dataset chest, waist, right side, sensor locations and 100% (real-world fall
and left side misaligned sensor dataset)
orientations, HMM
classifiers
LWT-based frequency
domain analysis and Accuracy � 100%
Generated from
[11] 2017 3-Axes accelerometer Chest SVM-based time Sensitivity � 100%
experiments
domain analysis of Specificity � 100%
RMS of acceleration
BioMed Research International 11

Table 2: Continued.
Sensor placement (if
Reference Year Dataset used Sensors/dataset used Methodology Observed performance
wearable system)
Backpropagation Accuracy � 98.182%
Generated from 3-Axis accelerometer, neural network Precision � 98.33%
[32] 2017 Waist
experiments 3-axis gyroscope (BPNN) for fall Sensitivity � 95.161%
detection Specificity � 99.367%
Naı̈ve-Bayes gave best
Naı̈ve-Bayes, SVM,
Generated from results
[39] 2010 Accelerometer Chest, thigh OneR, C4.5 (J48),
experiments Accuracy � 100%
neural networks
Sensitivity � 87.5%
Better accuracy with
Bayesian framework
Generated from Different parts of the improved classification
[43] 2016 Accelerometer for feature selection,
experiments body than Naı̈ve-Bayes and
Naı̈ve-Bayes, C4.5
C4.5
Accuracy and precision
SVM, kNN, complex
of SVM were the
Generated from tree algorithms applied
[33] 2016 3D accelerometer Chest highest
experiments on data generated by
Recall was highest for
accelerometers
complex tree
ENN + kNN (where For ENN + kNN:
Generated from Accelerometer ENN was applied to Sensitivity � 95.52%
[44] 2017 Not applicable
experiments (MobiAct dataset) remove outliers), Specificity � 97.07%
ANN, SVM, and J48 Precision � 91.83%
Accuracy � 99.52%
Generated from
[41] 2018 Triaxial gyroscope Waist Decision tree Precision � 99.3%
experiments
Recall � 99.5%
3D accelerometer, 3D
Event-ML, Better precision and F-
gyroscope-Cogent
classification and scores with Event-ML
Cogent dataset, dataset
[45] 2018 Chest, waist regression tree than FOSW and
SisFall dataset Accelerometer,
(CART), kNN, logistic FNSW-based
gyroscope-SisFall
regression, SVM approaches
dataset
Extraction of new
features from
acceleration and
ANN, kNN, QSVM,
Accelerometer, angular velocity
[42] 2019 Public datasets Chest, thigh ensemble bagged tree
gyroscope improved the accuracy
(EBT)
of all 4 classifiers.
Accuracy of EBT was
highest (97.7%)
Accuracy for fall
detection was the
highest for kNN
Accelerometer, kNN, SVM, random (99.8%). Accuracy for
[51] 2019 SisFall dataset Waist
gyroscope forest recognizing fall
activities was the
highest for random
forest (96.82%)
Accuracy and
SisFall dataset, sensitivity of SVM were
SVM, kNN, Naı̈ve-
[52] 2018 generated from Accelerometer Chest/thigh, waist the highest (97.6% and
Bayes, decision tree
experiments 98.3%, respectively) for
both datasets
Without risk
categorization: 81% for
Accelerometer, kNN, Naı̈ve-Bayes,
Wrist, waist, chest, decision tree
[63] 2018 UMA dataset gyroscope, SVM, ANN, decision
ankle With risk
magnetometer tree
categorization: 85% for
decision tree
12 BioMed Research International

Table 2: Continued.
Sensor placement (if
Reference Year Dataset used Sensors/dataset used Methodology Observed performance
wearable system)
CNN-based models for Highest accuracy
[56] 2019 Public datasets Accelerometer Not specified
feature extraction reported � 99.86%
Highest accuracy
reported for fall
SisFall dataset- detection: 83.68%
[57] 2018 original and Accelerometer Not specified RNN (before manual
manually labelled labelling), 98.33%
(after manual
labelling)
Accelerometer,
Generated from
[36] 2018 gyroscope, Near the waist kNN Accuracy � 99.4%
experiments
magnetometer
Generated from Accuracy � 91.67%
[16] 2018 Accelerometer Waist Decision tree
experiments Precision � 93.75%
Highest accuracy (after
[54] 2018 SisFall dataset Accelerometer Waist RNN with LSTM hyperparameter
optimization) � 97.16%
Accelerometer, SVM, decision tree, Highest
Generated from Right, left, and front
[53] 2017 gyroscope, proximity kNN, discriminant accuracy � 99% for
experiments pockets
sensor, compass analysis SVM
Generated from Depth camera, Accuracy of fall
[59] 2018 Waist CNN
experiments accelerometer detection � 100%
CNN-based analysis
on time series
[55] 2017 Public datasets Accelerometer Not specified Accuracy � 92.3%
accelerometer data
converted to images
Overall accuracy of
ensemble classifier was
the highest, after fusion
of radar,
Ensemble subspace accelerometer, and
Generated from Accelerometer, radar, discriminant, linear camera � 91.3%. This is
[58] 2017 Wrist
experiments depth camera discriminant, kNN, an improvement of
SVM 11.2% compared to
radar-only and 16.9%
compared to
accelerometer-only
results
Without sensor
fusion:
Accelerometer
precision � 86.23%
Accelerometer,
Generated from Accelerometer
[62] 2018 gyroscope, Hip SVM, random forest
experiments recall � 87.46%
magnetometer
With sensor fusion:
precision � 94.78%,
recall � 94.37%, with
random forest

summarizes their performance parameters such as accuracy, survey on the factors that impact fall in the elderly. Existing
sensitivity, and specificity where available. literature shows that the risk of fall based on various factors
is presented in the form of odds ratio, based on actual
6. Biological Risk Factors on Falls observations. OR is the ratio of the probability of an event of
interest occurring to the probability of that event not oc-
The risk of fall exhibited by a subject could be influenced by curring. It helps in estimating the relationship between two
multiple factors such as age, biological and physiological binary variables. Andrade [64] explains the significance of
health profile, and environmental conditions. We did a OR in the medical field, especially in cases where logistic
BioMed Research International 13

regression analyses are applied, to find out the impact of a having high impact in men were different from those in
risk factor on an outcome variable. women. For example, in men, history of stroke presented an
Bird et al. [65] study the impact of decrease in postural OR of 1.91, while in women, it was 1.51, and the OR for
stability over short time frame on fall rates and observes that arthritis was 1.27 in men and 1.36 in women. The study
fall rates increase when postural stability decreases, despite highlights the differences between men and women in the
maintaining leg strength. Environmental factors also play a associations between falls and various biological and medical
role in a subject’s fall rate. In some of the experiments that factors. However, due to various limitations, further re-
we had performed, we observed that the accuracy of fall search is required to better understand these gender dif-
detection increased with the addition of health profile as a ferences and their implications for risk assessment. Gale
feature set. We present a survey on the biological risk factors et al. [74] report that there are certain gender-specific risk
on falls. factors, such as incontinence (OR � 1.48) and frailty
Current research indicates the correlation between a (OR � 1.69) in women, and older age (OR � 1.02), high levels
person’s biological health and the risk of his/her falling. of depressive symptoms (OR � 1.33), and being unable to
Kronfol [66] details results from comprehensive studies perform a standing balance test (OR � 3.32) in men. The
conducted on falls and cites causes of falls to be primarily authors suggest that although some homogeneity between
environment-related, gait disorders, vertigo, drop attack, the genders in the risk factors that were associated with falls
confusion, postural hypotension, and visual disorders. The were observed, gender should be taken into account in
study also specifies the odds ratio of various risk factors such designing fall-prevention strategies because of the existence
as weakness, balance and gait deficits, mobility, and cog- of several gender-specific risk factors. Table 3 summarizes
nitive impairments, on the risk of falling. The impact of the various biological factors that impact a subject’s prob-
behavioural risk factors such as ADL characteristics and ability of falling.
environment on falls is also explained. In [67], the authors
performed experiments on 163 elderly men and women aged 7. Challenges in the Design of Fall
60–95 years and found that history of falls, poor vision, use Detection Systems
of multiple medications, chronic diseases, use of walking
aids, vertigo, and balance problems were associated with falls From our analysis, an observation is that machine learning
among the elderly population living in long-term care algorithms applied to various datasets in the literature survey
homes. Graafmans et al. [68] constructed a risk profile for produce varying degrees of accuracy. This indicates that the
recurrent falls that included five risk factors: mobility im- performance of the algorithms is dependent on various
pairment, dizziness upon standing, history of stroke, poor factors such as the type and placement of the sensors, the fall
mental state, and postural hypotension and found out the pattern, related thresholds if any, the characteristics of the
probability of recurrent falls on people exhibiting one or dataset, and possibly the preprocessing that has been applied
more of these risk factors. In [69], a method based on to it. Most literature on wearable sensor-based methods
performing chi-square tests to compare fall risk and overall indicates that the performance of the algorithms varied with
injury risk with various demographic, behavioural, and the position of the sensors. The thresholds are dependent on
health-related variables is proposed. Odds ratios for the the subject’s physical parameters, and hence, the perfor-
association of each risk factor with the outcome were es- mance of threshold-based methods depends on the cus-
timated using bivariate analyses and multivariate logistic tomizations set according to features of the experimentation
regression models. environments.
There have also been studies on risk factors of falls on The lack of datasets that support research in this do-
subjects with specific conditions. For example, Stanmore main is also to be noted. Some research depends on public
et al. [70] observe that, in subjects with rheumatoid arthritis, datasets—while these are useful to perform initial com-
risk factors include swollen joints, use of psychotropic parative studies, the fact that no more information than
medications and steroids, poor balance, and VAS fatigue what is provided by the dataset would be available is a
score. This study did not uncover any relation between hindrance in proceeding further with research using such
gender or age and risk of falling, which may be indicative of datasets. Most research has been done on datasets gener-
the fact that risks arising from rheumatoid arthritis override ated via experimentations. For obvious reasons, the ex-
those specific to age or gender. periments for simulating falls in all these cases were
According to Li et al. [71], the major risk factors resulting conducted in controlled environments, which may not
in fall-related injuries are intrinsic and not situational or reflect real-life situations accurately. Again, for obvious
environment-related. Vertigo, weakness of the legs, and reasons, real datasets on fall patterns of the elderly are not
history of cancer were found to be risk factors. Cattelani et al. available.
[72] propose a system for fall risk estimation in the elderly.
This study leverages the already existing analysis of fall risk 8. Conclusions and Future Directions
and deduces a fall probability for a subject based on sta-
tistical methods. In this paper, we performed a brief comparison of fall de-
A study on the impact of gender on the probability of fall tection systems that rely on environmental sensor-based,
was inconclusive. Chang and Do [73] study the implications vision-based, and wearable sensor-based techniques. We
of gender on risk factors for falls among seniors. The factors then did a comprehensive survey of application of machine
14 BioMed Research International

Table 3: Biological risk factors on falls.

Reference Year Population demographics Relevant parameters [odds ratio]
Activity limitation due to health problems [1.13]; use
Adults 65 years and older, with focus on adults 85
[69] 2012 of assistive devices [2.18]; diabetes [1.2]; history of
years and older
stroke [1.32]
Swollen or tender lower extremity joints [2.0]; history
of stroke or Parkinson’s disease [1.8]; history of ≥2
[70] 2013 Adults 18–88 years with RA
falls in previous 12 months [4.3]; symptoms of feeling
dizzy or unsteady [1.8]
Weakness [8]; balance deficit [5]; gait deficit [5];
visual deficit [9]; mobility limitation [8]; cognitive
[66] 2007 Older adults
impairment [5]; impaired functional status [4];
postural hypotension [5]
Visual deficit [1.851]; chronic conditions [1.633];
vertigo [2.237]; imbalance [3.105]; fear of falling
[67] 2016 Adults aged 60–95 years [3.227]; history of previous falls [5.661]; postural
hypotension [0.804]; use of assistive devices [2.139];
hearing impairment [1.543]
Mobility impairment [5.0]; dizziness upon standing
[2.3]; history of stroke [3.4]; postural hypotension
[68] 1996 Adults 70 years or older living in homes for elderly
[2.0]; urinary incontinence [2.6]; use of walking aid
[3.2]; visual deficit [1.7]; history of falls [3.5]
Cancer [2.71]; vertigo [4.35]; weakness of lower legs
[71] 2016 Medical records of elderly hospitalized patients
[2.15]

learning in wearable sensor-based FDSs. The survey inputs—for example, those that support activation of
was done by taking into account the type of sensors cameras to take pictures in the event of detection of a fall,
used, their positioning, the dataset used for analysis, the to assist a remote caregiver.
machine learning algorithms employed, and their perfor- The literature survey does not show any indication of
mance summary results. We also presented a survey of having considered a person’s biological parameters or health
biological factors affecting a person’s probability of fall. history into a fall detection algorithm. We observe that while
Our findings indicate that wearable systems for fall de- there has been considerable research into finding out the
tection have the advantage of being less intrusive, especially correlation between a person’s health profile and his
for elderly people, and ML techniques have the ability to probability of fall (as indicated by the odds ratio), not much
detect falls to a reasonable accuracy level. However, a has been explored in evaluating the impact of this odds ratio
wearable system consisting of a device such as a wrist band on the performance parameters of various ML algorithms
alone is insufficient to meet the requirements of a com- for fall detection.
prehensive FDS. We also note that there is no integrated system that
From our study, we have observed that systems capable considers the fall risk to detect falls and generate alerts and
of generating alerts on detecting falls have been designed, camera-assisted observations. In addition, while the
but they fall short in the ability to activate adequate alerts, existing systems focus extensively on fall detection, there is
while minimizing the cost and power requirements. For scope for building an FDS that implements a closed-loop
example, there are vision-based systems which work based feedback—one that learns from a subject’s fall patterns/
on videos/images captured by a camera. Such systems may history and change in physiological parameters at the time
be expensive and have higher power and bandwidth re- of fall and trains the fall detection algorithm based on these
quirements. Also, parameters such as pulse, heart rate, factors, to enable accurate profiling.
temperature, SPO2, and electrodermal activity, which Our team is working on building a fall detection system
would increase the accuracy of the fall detection algo- that applies machine learning techniques for fall detection of
rithm, are not taken into account in such systems because the elderly. Our work on the comparison of the performance
they rely solely on image processing mechanisms. Non- of various machine learning algorithms on a public dataset
vision-based systems (e.g., wearable systems) detect falls, for fall detection is given in [63]. As next steps, we plan to
but they are decoupled from video/image inputs such as work on data generated by a combination of IMU and vital
those generated by a camera. The flip side is that false signs sensors which are designed to be integrated into a wrist
positives and false negatives may tend to be ignored; for band. These wrist bands would be worn by elderly people
example, a false positive would trigger unnecessary action staying in old-age care homes, where the end-to-end system
from the caregivers, and a false negative may not get the is meant to be deployed. The system is also being designed to
attention it deserves, leading to potentially dangerous have the capability to generate alerts accurately and in a
situations. We observe that there is scope for develop- timely manner, in the event of an abnormality. The mod-
ing systems that combine sensor readings with image erately mobile environment that the deployment pattern
BioMed Research International 15

presents has an impact on the network design to support the [14] K. Kaewkannate and S. Kim, “A comparison of wearable
end-to-end functionality. fitness devices,” BMC Public Health, vol. 16, no. 1, p. 433, 2016.
[15] D. Oletic and V. Bilas, “System-level power consumption
analysis of the wearable asthmatic wheeze quantification,”
Conflicts of Interest Journal of Sensors, vol. 2018, Article ID 6564158, 18 pages,
2018.
The authors declare that there are no conflicts of interest [16] D. Yacchirema, J. S. de Puga, C. Palau, and M. Esteve, “Fall
regarding the publication of this paper. detection system for elderly people using IoT and big data,”
Procedia Computer Science, vol. 130, pp. 603–610, 2018.
References [17] T. M. Mitchell, Machine Learning, McGraw-Hill, New York,
NY, USA, 1997, ISBN978-0-07-042807-2.
[1] The 2015 Ageing Report European Economy—Economic and [18] T. I. Iliev, D. S. Tabakov, and A. I. Dotswinsky, “Automatic fall
Budgetary Projections for the 28 EU Member States (2013- detection of elderly living alone at home environment,”
2060), ISSN 1725-3217, ISSN 0379-0991. Global Journal of Medical Research, vol. 11, no. 4, 2011.
[2] World Health Organization, Publications on Ageing and Life [19] F. Wu, H. Zhao, Y. Zhao, and H. Zhong, “Development of a
Course, World Health Organization, Geneva, Switzerland, wearable-sensor-based fall detection system,” International
2007, http://www.who.int/ageing/publications. Journal of Telemedicine and Applications, vol. 2015, Article ID
[3] N. Pannurat, S. Thiemjarus, and E. Nantajeewarawat, “Au- 576364, 11 pages, 2015.
tomatic fall monitoring: a review,” Sensors, vol. 14, no. 7, [20] M. Shahiduzzaman, “Fall detection by accelerometer and
pp. 12900–12936, 2014. heart rate variability measurement,” Global Journal of Com-
[4] C. Taramasco, T. Rodenas, F. Martinez et al., “A novel puter Science and Technology, vol. 15, no. 3, 2015.
monitoring system for fall detection in older people,” IEEE [21] J. A. Li, M. A Stankovic, A. T. Hanson, J. L. Barth, and
Access, vol. 6, pp. 43563–43574, 2018. G. Zhou, “Accurate, fast fall detection using gyroscopes and
[5] C. Han, K. Wu, Y. Wang, and L. M. Ni, “WiFall: device-free accelerometer-derived posture information,” in Proceedings of
fall detection by wireless networks,” in Proceedings of the IEEE the 2009 Sixth International Workshop on Wearable and
INFOCOM 2014—IEEE Conference on Computer Communi- Implantable Body Sensor Networks, pp. 138–143, Berkeley,
cations, pp. 271–279, Toronto, ON, Canada, April 2014. CA, USA, June 2009.
[6] L. Ciabattoni, G. Foresi, A. Monteriù, D. P. Pagnotta, and [22] G. Koshmak, M. Linden, and A. Loutfi, “Dynamic Bayesian
L. Tomaiuolo, “Fall detection system by using ambient in- networks for context-aware fall risk assessment,” Sensors,
telligence and mobile robots,” in Proceedings of the 2018 vol. 14, no. 5, pp. 9330–9348, 2014.
Zooming Innovation in Consumer Technologies Conference [23] J. Chen, K. Kwong, D. Chang, J. Luk, and R. Bajcsy, “Wearable
(ZINC), pp. 130-131, Novi Sad, Serbia, May 2018. sensors for reliable fall detection,” in Proceedings of the IEEE
[7] A. Núñez-Marcos, G. Azkune, and I. Arganda-Carreras, Engineering in Medicine and Biology 27th Annual Conference,
“Vision-based fall detection with convolutional neural net- pp. 3551–3554, Shanghai, China, January 2005.
works,” Wireless Communications and Mobile Computing, [24] P. Tsinganos Skodras, “On the comparison of wearable sensor
vol. 2017, Article ID 9474806, 16 pages, 2017. data fusion to a single sensor machine learning technique in
[8] N. Zerrouki, F. Harrou, A. Houacine, and Y. Sun, “Fall de- fall detection,” Sensors (Basel), vol. 18, no. 2, p. 592, 2018.
tection using supervised machine learning algorithms: a [25] A. Mao, X. Ma, Y. He, and J. Luo, “Highly portable, sensor-
comparative study,” in Proceedings of the 8th International based system for human fall monitoring,” Sensors, vol. 17,
Conference on Modelling, Identification and Control (ICMIC), no. 9, p. 2096, 2017.
pp. 665–670, Algiers, Algeria, November 2016. [26] P. Kostopoulos, A. I. Kyritsis, M. Deriaz, and D. Konstantas,
[9] L. Anishchenko, “Machine learning in video surveillance for “F2D: a location aware fall detection system tested with real
fall detection,” in Proceedings of the 2018 Ural Symposium on data from daily life of elderly people,” in Proceedings of the
Biomedical Engineering, Radioelectronics and Information 17th International Conference on E-Health Networking, Ap-
Technology (USBEREIT), pp. 99–102, Yekaterinburg, Russia, plication & Services (HealthCom), pp. 397–403, Boston, MA,
May 2018. USA, October 2015.
[10] S. Bhandari, N. Babar, P. Gupta, N. Shah, and S. Pujari, “A [27] T. Chaitep and J. Chawachat, “A 3-phase threshold algorithm
novel approach for fall detection in home environment,” in for smartphone-based fall detection,” in Proceedings of the
Proceedings of the 2017 IEEE 6th Global Conference on 14th International Conference on Electrical Engineering/
Consumer Electronics (GCCE), pp. 1–5, Nagoya, Japan, Oc- Electronics, Computer, Telecommunications and Information
tober 2017. Technology (ECTI-CON), pp. 183–186, Phuket, Thailand, June
[11] H. Liang and W. Usaha, “Fall detection using lifting wavelet 2017.
transform and support vector machine,” in Proceedings of the [28] T. de Quadros, A. E. Lazzaretti, and F. K. Schneider, “A
2017 Federated Conference on Computer Science and Infor- movement decomposition and machine learning-based fall
mation Systems (FedCSIS), pp. 877–883, Prague, Czech Re- detection system using wrist wearable device,” IEEE Sensors
public, September 2017. Journal, vol. 18, no. 12, pp. 5082–5089, 2018.
[12] P. Yanfei, P. Jianjun, L. Jiping, P. Yan, and B. Hu, “Design and [29] A. Özdemir and B. Barshan, “Detecting falls with wearable
development of the fall detection system based on point sensors using machine learning techniques,” Sensors, vol. 14,
cloud,” Procedia Computer Science, vol. 147, pp. p271–p275, no. 6, pp. 10691–10708, 2014.
2019. [30] Y. Choi, A. S. Ralhan, and S. Ko, “A study on machine
[13] N. Lu, Y. Wu, L. Feng, and J. Song, “Deep learning for fall learning algorithms for fall detection and movement classi-
detection: three-dimensional CNN combined with LSTM on fication,” in Proceedings of the International Conference on
video kinematic data,” IEEE Journal of Biomedical and Health Information Science and Applications, pp. 1–8, Jeju Island,
Informatics, vol. 23, no. 1, pp. 314–323, 2019. South Korea, April 2011.
16 BioMed Research International

[31] P. Vallabh, R. Malekian, N. Ye, and D. C. Bogatinoska, “Fall [45] I. Putra, J. Brusey, E. Gaura, and R. Vesilo, “An event-trig-
detection using machine learning algorithms,” in Proceedings gered machine learning approach for accelerometer-based fall
of the 24th International Conference on Software, Telecom- detection,” Sensors, vol. 18, no. 2, p. 20, 2017.
munications and Computer Networks (SoftCOM), pp. 1–9, [46] C.-Y. Hsieh, K.-C. Liu, C.-N. Huang, W.-C. Chu, and
Split, Croatia, September 2016. C.-T. Chan, “A Novel hierarchical fall detection algorithm
[32] A. Jefiza, E. Pramunanto, H. Boedinoegroho, and using a multiphase fall model,” Sensors, vol. 17, no. 2, p. 307,
M. H. Purnomo, “Fall detection based on accelerometer and 2017.
gyroscope using back propagation,” in Proceedings of the 4th [47] D. Genoud, V. Cuendet, and J. Torrent, “Soft fall detection
International Conference on Electrical Engineering, Computer using machine learning in wearable devices,” in Proceedings of
Science and Informatics (EECSI), pp. 1–6, Yogyakarta, the IEEE 30th International Conference on Advanced Infor-
Indonesia, September 2017. mation Networking and Applications (AINA), pp. 501–505,
[33] F. Hossain, M. L. Ali, M. Z. Islam, and H. Mustafa, “A di- Crans-Montana, Switzerland, March 2016.
rection-sensitive fall detection system using single 3D ac- [48] G. Chetty, M. White, and F. Akther, “Smart phone based data
celerometer and learning classifier,” in Proceedings of the 2016 mining for human activity recognition,” Procedia Computer
International Conference on Medical Engineering, Health Science, vol. 46, pp. 1181–1187, 2015.
Informatics and Technology (MediTec), pp. 1–6, Dhaka, [49] H.-C. Kao, J.-C. Hung, and C.-P. Huang, “GA-SVM applied to
Bangladesh, December 2016. the fall detection system,” in Proceedings of the 2017 Inter-
[34] M. V. Albert, K. Kording, M. Herrmann, and A. Jayaraman, national Conference on Applied System Innovation (ICASI),
“Fall classification by machine learning using mobile phones,” pp. 436–439, Sapporo, Japan, May 2017.
PLoS One, vol. 7, no. 5, Article ID e36556, 2012. [50] A. Jahanjoo, M. N. Tahan, and M. J. Rashti, “Accurate fall
[35] M.-S. Lee, J.-G. Lim, and K.-R. Park, “Unsupervised clustering detection using 3-axis accelerometer sensor and MLF algo-
for abnormality detection based on the tri-axial accelerom- rithm,” in Proceedings of the 2017 3rd International Confer-
eter,” in Proceedings of the ICCAS-SICE, pp. 134–137, ence on Pattern Recognition and Image Analysis (IPRIA),
Fukuoka City, Japan, August 2009. pp. 90–95, Shahrekord, Iran, April 2017.
[36] T. B. Rodrigues, D. P. Salgado, M. C. Cordeiro et al., “Fall [51] F. Hussain, F. Hussain, M. Ehatisham-ul-Haq, and
detection system by machine learning framework for public M. A. Azam, “Activity-aware fall detection and recognition
health,” Procedia Computer Science, vol. 141, pp. 358–365, based on wearable sensors,” IEEE Sensors Journal, vol. 19,
2018. no. 12, pp. 4528–4536, 2019.
[37] S. Yu, H. Chen, and R. A. Brown, “Hidden Markov model- [52] K.-C. Liu, C.-Y. Hsieh, S. J.-P. Hsu, and C.-T. Chan, “Impact
based fall detection with motion sensor orientation calibra- of sampling rate on wearable-based fall detection systems
tion: a case for real-life home monitoring,” IEEE Journal of based on machine learning models,” IEEE Sensors Journal,
Biomedical and Health Informatics, vol. 22, no. 6, pp. 1847– vol. 18, no. 23, pp. 9882–9890, 2018.
1853, 2018. [53] A. Hakim, M. S. Huq, S. Shanta, and B. S. K. K. Ibrahim,
[38] M. Guvensan, A. Kansiz, N. Camgoz, H. Turkmen, A. Yavuz, “Smartphone based data mining for fall detection: analysis
and M. Karsligil, “An energy-efficient multi-tier architecture and design,” Procedia Computer Science, vol. 105, pp. 46–51,
for fall detection on smartphones,” Sensors, vol. 17, no. 7, 2017.
p. 1487, 2017. [54] M. Musci, D. Martini, N. Blago, T. Facchinetti, and M. Piastra,
[39] X. Yang, A. Dinh, and L. Che, “A wearable real-time fall “Online fall detection using recurrent neural networks,” 2018,
detector based on Naive Bayes Classifier,” in Proceedings of the https://arxiv.org/abs/1804.04976.
23rd Canadian Conference on Electrical and Computer En- [55] A. H. Fakhrulddin, X. Fei, and H. Li, “Convolutional neural
gineering (CCECE 2010), pp. 1–4, Calgary, AB, Canada, May networks (CNN) based human fall detection on body sensor
2010. networks (BSN) sensor data,” in Proceedings of the 2017 4th
[40] J. He, S. Bai, and X. Wang, “An unobtrusive fall detection and International Conference on Systems and Informatics (ICSAI),
alerting system based on Kalman filter and Bayes network pp. 1461–1465, Hangzhou, China, November 2017.
classifier,” Sensors, vol. 17, no. 6, p. 1393, 2017. [56] G. Leoni, P. T. Endo, K. Monteiro, E. Rocha, I. Silva, and
[41] S. Zhao, W. Li, W. Niu, R. Gravina, and G. Fortino, “Rec- T. G. Lynn, “Accelerometer-based human fall detection using
ognition of human fall events based on single tri-axial gy- convolutional neural networks,” Sensors (Basel), vol. 19, no. 7,
roscope,” in Proceedings of the IEEE 15th International p. E1644, 2019.
Conference on Networking, Sensing and Control (ICNSC), [57] E. Torti, A. Fontanella, M. Musci et al., “Embedded real-time
pp. 1–6, Zhuhai, China, March 2018. fall detection with deep learning on wearable devices,” in
[42] A. Chelli and M. Patzold, “A machine learning approach for Proceedings of the 21st Euromicro Conference on Digital
fall detection and daily living activity recognition,” IEEE System Design (DSD), pp. 405–412, Prague, Czech Republic,
Access, vol. 7, pp. 38670–38687, 2019. August 2018.
[43] H. Wang, M. Li, J. Li, J. Cao, and Z. Wang, “An improved fall [58] H. Li, A. Shrestha, F. Fioranelli et al., “Multisensor data fusion
detection approach for elderly people based on feature weight for human activities classification and fall detection,” in
and Bayesian classification,” in Proceedings of the IEEE In- Proceedings of the 2017 IEEE Sensors, pp. 1–3, Glasgow, UK,
ternational Conference on Mechatronics and Automation, November 2017.
pp. 471–476, Harbin, China, August 2016. [59] N. Dawar and N. Kehtarnavaz, “A convolutional neural
[44] P. Tsinganos and A. Skodras, “A smartphone-based fall network-based sensor fusion system for monitoring transition
detection system for the elderly,” in Proceedings of the 10th movements in healthcare applications,” in Proceedings of the
International Symposium on Image and Signal Processing 2018 IEEE 14th International Conference on Control and
and Analysis, pp. 53–58, Ljubljana, Slovenia, September Automation (ICCA), pp. 482–485, Anchorage, AK, USA, June
2017. 2018.
BioMed Research International 17

[60] X. Zhou, L.-C. Qian, P.-J. You, Z.-G. Ding, and Y.-Q. Han,
“Fall detection using convolutional neural network with
multi-sensor fusion,” in Proceedings of the 2018 IEEE Inter-
national Conference on Multimedia & Expo Workshops
(ICMEW), pp. 1–5, San Diego, CA, USA, July 2018.
[61] R. Ramezani, Y. Xiao, and A. Naeim, “Sensing-Fi: Wi-Fi CSI
and accelerometer fusion system for fall detection,” in Pro-
ceedings of the 2018 IEEE EMBS International Conference on
Biomedical & Health Informatics (BHI), pp. 402–405, Las
Vegas, NV, USA, March 2018.
[62] T.-L. Nguyen, T.-A. Le, and C. Pham, “The internet-of-things
based fall detection using fusion feature,” in Proceedings of the
2018 10th international conference on knowledge and systems
engineering (KSE), pp. 129–134, Ho Chi Minh City, Vietnam,
November 2018.
[63] A. Ramachandran, R. Adarsh, P. Pahwa, and K. R. Anupama,
“Machine learning-based fall detection in geriatric healthcare
systems,” in Proceedings of the 2018 IEEE International
Conference on Advanced Networks and Telecommunications
Systems (ANTS), pp. 1–6, Indore, India, December 2018.
[64] C. Andrade, “Understanding relative risk, odds ratio, and
related terms: as simple as it can get,” The Journal of Clinical
Psychiatry, vol. 76, no. 07, pp. e857–e861, 2015.
[65] M.-L. Bird, J. Pittaway, I. Cuisick, M. Rattray, and K. Ahuja,
“Age-related changes in physical fall risk factors: results from
a 3 year follow-up of community dwelling older adults in
Tasmania, Australia,” International Journal of Environmental
Research and Public Health, vol. 10, no. 11, pp. 5989–5997,
2013.
[66] N. Kronfol, Biological, Medical and Behavioral Risk Factors on
Falls, World Health Organisation, Geneva, Switzerland, 2012,
http://www.who.int/ageing/project/falls_prevention_older_age/
en/inded.html.
[67] P. Dhargave and R. Sendhilkumar, “Prevalence of risk factors
for falls among elderly people living in long-term care
homes,” Journal of Clinical Gerontology and Geriatrics, vol. 7,
no. 3, pp. 99–103, 2016.
[68] W. C. Graafmans, M. E. Ooms, H. M. A. Hofstee,
P. D. Bezemer, L. M. Bouter, and P. Lips, “Falls in the elderly: a
prospective study of risk factors and risk profiles,” American
Journal of Epidemiology, vol. 143, no. 11, pp. 1129–1136, 1996.
[69] A. C. Grundstrom, C. E. Guse, and P. M. Layde, “Risk factors
for falls and fall-related injuries in adults 85 years of age and
older,” Archives of Gerontology and Geriatrics, vol. 54, no. 3,
pp. 421–428, 2012.
[70] E. K. Stanmore, J. Oldham, D. A. Skelton et al., “Risk factors
for falls in adults with rheumatoid arthritis: a prospective
study,” Arthritis Care & Research, vol. 65, no. 8,
pp. 1251–1258, 2013.
[71] I.-F. Li, Y. Hsiung, H.-F. Hsing, M.-Y. Lee, T.-H. Chang, and
M.-Y. Huang, “Elderly Taiwanese’s intrinsic risk factors for
fall-related injuries,” International Journal of Gerontology,
vol. 10, no. 3, pp. 137–141, 2016.
[72] L. Cattelani, F. Chesani, P. Palumbo et al., “FRAT-Up, a rule-
based system evaluating fall risk in the elderly,” in Proceedings
of the IEEE 27th International Symposium on Computer-Based
Medical Systems, pp. 38–41, New York, NY, USA, May 2014.
[73] V. C. Chang and M. T. Do, “Risk factors for falls among
seniors: implications of gender,” American Journal of Epi-
demiology, vol. 181, no. 7, pp. 521–531, 2015.
[74] C. R. Gale, C. Cooper, and A. Aihie Sayer, “Prevalence and
risk factors for falls in older men and women: the English
longitudinal study of ageing,” Age and Ageing, vol. 45, no. 6,
pp. 789–794, 2016.
 MEDIATORS of

INFLAMMATION

The Scientific Gastroenterology Journal of

World Journal
Hindawi Publishing Corporation
Research and Practice
Hindawi
Hindawi
Diabetes Research
Hindawi
Disease Markers
Hindawi
www.hindawi.com Volume 2018
http://www.hindawi.com
www.hindawi.com Volume 2018
2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal of International Journal of

Immunology Research
Hindawi
Endocrinology
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Submit your manuscripts at

www.hindawi.com

BioMed
PPAR Research
Hindawi
Research International
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal of
Obesity

Evidence-Based
Journal of Stem Cells Complementary and Journal of
Ophthalmology
Hindawi
International
Hindawi
Alternative Medicine
Hindawi Hindawi
Oncology
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2013

Parkinson’s
Disease

Computational and
Mathematical Methods
in Medicine
Behavioural
Neurology
AIDS
Research and Treatment
Oxidative Medicine and
Cellular Longevity
Hindawi Hindawi Hindawi Hindawi Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018