Fall Detection System using Wearable Sensor Devices and Machine

Learning: A Review
Nur Izdihar Muhd Amir1 , Rudzidatul Akmam Dziyauddin1 , Norliza Mohamed1 , Nor
Syahidatul Nadiah Ismail1 , Hazilah Mad Kaidi1 , Norulhusna Ahmad1 , and Mohd Azri
Mohd Izhar1
1
Affiliation not available
Posted on 19 Mar 2024 — CC-BY-SA 4 — https://doi.org/10.36227/techrxiv.171084921.16728034/v1 — e-Prints posted on TechRxiv are preliminary reports that are not peer reviewed. They should not b...

March 19, 2024

Abstract
Over the past years, Fall Detection System (FDS) has undergone extensive research to improve living risk, especially for the
elderly who are vulnerable to these fall events. Devices employing sensors are crucial components of FDS in achieving high
accuracy and sensitivity. This article overviews different sensor modalities, such as ambient-based and vision-based systems, as
well as commonly used wearable devices for fall detection, along with the associated data processing algorithms. The critical
elements of fall detection, such as architectures and algorithms for processing sensor data, machine learning and deep learning
methodologies, and validation of FDS performance, are considered. The article also delves into safety aspects and presents
technical challenges and concerns in FDS for researchers in the field to identify areas requiring further improvement. Finally,
future research opportunities to improve fall detection for widespread use are outlined.

1
 Fall Detection System using Wearable Sensor
Devices and Machine Learning: A Review
Nur Izdihar Muhd Amir, Rudzidatul Akmam Dziyauddin, Senior Member, IEEE, Norliza
Mohamed, Senior Member, IEEE, Nor Syahidatul Nadiah Ismail, Member, IEEE, Hazilah
Mad Kaidi, Senior Member, IEEE, Norulhusna Ahmad, Senior Member, IEEE and Mohd Azri
Mohd Izhar, Member, IEEE

Abstract— Over the past years, Fall Detection System (FDS) has undergone extensive research to improve living
risk, especially for the elderly who are vulnerable to these fall events. Devices employing sensors are crucial
components of FDS in achieving high accuracy and sensitivity. This article overviews different sensor modalities,
such as ambient-based and vision-based systems, as well as commonly used wearable devices for fall detection,
along with the associated data processing algorithms. The critical elements of fall detection, such as architectures
and algorithms for processing sensor data, machine learning and deep learning methodologies, and validation of
FDS performance, are considered. The article also delves into safety aspects and presents technical challenges
and concerns in FDS for researchers in the field to identify areas requiring further improvement. Finally, future
research opportunities to improve fall detection for widespread use are outlined.

Index Terms— Fall Detection system (FDS), human activity recognition, wearable devices, machine learning and
deep learning, sensor data processing

health-related issues and accidents that may lead to bodily

I. INTRODUCTION instability, ultimately resulting in hazardous falls. Falls in

W ITH the advancements in medical technologies and elderly can exacerbate serious injuries compared to younger
healthcare systems, researchers are actively exploring people [4]. Statistics show that more than 25% of people aged
innovative approaches to benefit the current population. One over 65 years old experience falls every year and this figure
such avenue involves the development of fall detection systems grows to 42% for those over 70 [5]. Especially for those without
(FDS), aimed at mitigating the consequences of such a caretaker could turn this event into fatality. Hence, the
unfortunate events [1]. FDS utilises sensor(s) and processing monitored safety conditions allow them to go about their daily
solutions to discern incidents of falls. Hence, various types of routine comfortably. Therefore, the gradual increase of elderly
FDS have been introduced throughout the research, particularly who live alone could significantly benefit from this technology.
the interest in this paper is the wearable-type detection devices, The remainder of this article is organized as follows. Section
leveraging one or several sensor modalities, including cameras II overviews a generic fall detection architecture of the existing
and Inertial Measurement unit (IMU). By adeptly processing FDS. Next, Section III dwells deeper into sensors used for fall
raw sensor data, these devices can effectively detect and detection. Subsequently, Section IV details state-of-the-art
distinguish fall incidents from Activities of Daily Living methods used to infer information. processing and learning
(ADLs), such as walking, running, and others. Every sensor has algorithms in FDS, Section V addresses the emergent
demonstrated its strength and constraints, and the signal validation domain, discussing the norms, performance metrics,
response varies on the physical features of a subject and how and the monitoring of abnormal situations especially
the fall incurred (i.e. right, left, front or back). Therefore, the considering machine learning. Finally, Section VI concludes
processing of sensor data is the primary key to deriving the future opportunities for FDS.
information on the movement of the subject including falls. The
aspects involve 1) processing architectures considering types of II. GENERIC FALL DETECTION SYSTEM ARCHITECTURE
sensors and their placement; 2) processing algorithms for Fig. 1. illustrates the taxonomy of FDS practised in the
ADLs, such as the movement of subject and classification, fall research, it comprises types of FDS which cover the elements,
detection and fall alertness; 3) integration of physics-/model- sensors and its components, and common learning algorithms
/data-driven approaches into data processing; 4) performance employed in FDS. Meanwhile, Fig. 2. presents a general
metrics and validation approaches for a high level of architecture of FDS. The system establishes once a fall occurs
application; and 5) wearable device standards and safety. In to the person being monitored, and the fall detection method
this paper, our objectives are to determine the key challenges can either be ambient-based or wearable-based sensors. The
in sensor processing for fall detection, review the advances data from sensors are transferred to the local workstation [6] or
toward addressing them, and cloud services [7] for processing and learning purposes before
identify the existing gaps in achieving high-level detection deploying the result through the gateway to notify the
particularly in real-time. authorities of an emergency.
The FDS work appears to provide an advantage to the elderly In the state-of-the-art, there are three types of fall detection
while substantively extending support to vulnerable individuals approaches (ambient, vision, and wearable) and two types of
across diverse age groups. Owing to the congenial living space, approaches (ambient and wearable). In [8], the author classified
the number of elderly who are living alone also increases fall detectors into three categories: environmental sensing-
accordingly [2, 3]. Nevertheless, individuals in this based systems, wearable sensor-based systems and vision-
demographic are vulnerable to adverse events, including based systems. While [9] categorises hybrid sensing in their
 FDS Threshold-based and ML-based algorithms. However, over
time, threshold-based algorithms encounter challenges when
Type Algorithm attempting to adapt to new types of falls, leading to reduced
Thresh Machine
accuracy in fall detection [12]. ML methods have emerged as a
Ambient Wearable
hold Learning solution to overcome the limitations of threshold-based
algorithms. These methods are considered more sophisticated
Placement Sensor Dataset Model
approaches in addressing the challenges of fall detection,
Kalman Quertion offering improved accuracy and adaptability. Each ML
Wrist Waist filter technique has its learning-based approach, and factors such as
Chest Magneto Fall feature selection, sensor feasibility, sensor placement, dataset,
meter Data
and parameter setting, need to be considered when comparing
Gyros Accelero
Classical ANN different techniques. However, DL techniques have largely
cope meter
overcome these factors and achieved outstanding performance.
Camera Random
The availability of datasets plays an important role in the
-based
Infrared
Forest
K-NN CNN RNN
effectiveness of ML algorithms because having a large training
Pressure Audio Logistic LSTM GRU
dataset can help in achieving optimal performance while a
Regression
SVM small dataset may result in overfitting. However, by addressing
this concern, researchers in [22] proposed that the scarcity of
Fig. 1. A taxonomy of the review on FDS using Sensors and Machine data for training fall detection models can be alleviated through
Learning
the application of appropriate techniques, such as stacking in
third approach which utilises a combination of wearable and DL.
ambiance-based sensors. In contrast, [10] concluded that fall
detectors can be broadly categorised into two types: context- III. SENSOR FOR FALL DETECTION
aware systems and wearable devices. The former combines
A. Ambient-based and vision-based
ambient-based and vision-based detection to become one
category, while the latter remains as a means of detecting falls Ambient-based systems work on input from sensors placed
through any device attached to the user physically in the environment which usually install sensors in the room
In wearable, the most commonly used sensors for various where the user lives or is monitored. The sensors employed are
FDS are the accelerometer and gyroscope [1]. In addition to cameras, pressure sensors, microphones or a famous but less
these two sensors, the magnetometer has been mentioned as preferred ambient sensor; the motion sensor as illustrated in
another kinematic attribute in wearable FDS [8], and [11] has Fig. 3. Some works demonstrate the viability of vision-based
added an inclinometer to the list. There are also other sensors [23] approaches by employing infrared cameras [24] and Red
introduced in wearable FDS such as heart rate variability Green Blue-Depth (RGB-D) cameras [25]. Most static Red
(HRV), electrocardiogram (ECG), and pulse oximetry (SPO2). Green Blue (RGB) cameras are less intrusive, in the sense that
Typically, wearables for fall detection are commonly no sensor needs to be attached to or worn by the user, and it run
positioned on the waist, chest, or thigh. However, there are on wired-power sources; hence, power supply is not an issue
alternative placements such as the forehead [12], ear [13], neck [26]. However, the main challenge of vision-based detection is
[14], shoulder [15], back [16, 17], wrist [18], ankle [19], or foot the extent of information that cameras may record, including
[20], although these locations are generally less favoured for personal details, physical features, and visuals of personal
fall detection purposes. Wrist-worn devices such as spaces.
smartwatches are particularly popular and have motivated the Recent work by Khan et al. [27] integrated an unsupervised
development of FDS. Additionally, the positioning of sensors acoustic sensor to collect the sound of footsteps to distinguish
can impact the accuracy of fall detection [21]. Therefore, falls from ADLs. The same concept by [4] explained very-low-
researchers claim that the ideal placement of sensors is along resolution thermal sensors located on two horizontal planes
the longitudinal axis of the body to ensure steady readings of along the floor, for classifying falls. Improvement in [28]
the sensor [8]. proposed a fusion between the force and accelerometer sensors
Two main types of algorithms are introduced in FDS: concealed under the tile, which produces satisfactory results.

Fig. 2. The general architecture of FDS in the research.

 aspects including the power consumption by the interfacing
sensors is necessary. [32] claimed that, according to the
datasheet provided, a gyroscope consumes an average of
between 6 and 10 times more current than an accelerometer.
Nevertheless, by using DL, the issue is more about the number
of samples rather than the number of interconnected sensors in
the wearable devices, the most common sensors are
accelerometers and gyroscopes [1]. Besides these two sensors,
the magnetometer is mentioned as another kinematic attribute
in wearable FDS [8], and Pannurat et al. [11] added an
inclinometer to the list. There are also other sensors introduced
in wearable FDS such as heart rate variability (HRV) [36, 37] ,
electrocardiogram (ECG) [38, 39], and pulse oximetry (SPO2)
[40]. These features could assist the doctor in logging the health
conditions of the elderly easily. Nevertheless, any additional
Fig. 3. Illustration of ambient sensors used in the research.
sensors depend on the application's purpose. Wearables
The main benefit of ambient-based method requires no practice continuous monitoring and exhibit low-power
particular equipment to be worn by the user or monitored consumption, thus, simple sensors like the mentioned three
subject, making them perfect for seniors who dislike having sensors: accelerometer, gyroscope and magnetometer are the
something foreign attached to their body, like jewellery or a favourites in FDS work.
hand-watch. However, the acquisition sensors are likely to be Falls must be recognised in a non-intrusive manner. For
more complex in computation process to increase effectiveness example, a wearable sensor that is too heavy or inconvenient
as it has no significant constraints on sources and energy supply for the subject, may not be a popular solution [8].
[1]. The characteristics of the wearable devices including their
In addition, these kinds of solutions are limited to their placements are detailed as follows:
deployment area [1], and not suitable in some situations. For 1) Processor
instance, the user leaves home often or moves to other places The main component of the wearable fall detection device is
rather than staying in the monitored room. Furthermore, the the processor features, the capability of high-performance
cost of sensors installation, wiring and sensor allocation processor is crucial to perform real-time computation. Early
planning [29], are anticipatingly expensive which may make stage by [41] proposed a threshold-based pre-impact fall
them economically unfeasible for common people. Another detection device employing an Intel PXA255 processor, where
significant point is that it may be difficult to install these the actual processing takes place. Another, work [12] integrated
devices in public health systems because installing camera TI MSP430 MCU for the head-worn device. In a more
technology and using advanced computer vision is intrusive sophisticated method, data are analysed through ML
[30] and they would not only gather information from the target approaches. Such as in [42], a sensor platform called
patients but also from other individuals, which could SHIMMER [43] is made up of a TI MSP430 MCU baseboard,
potentially compromise their privacy. Since these issues arise and the processing occurs at a remote workstation applying
in ambient-based FDS, wearables hold the potential to SVM. While [44] chose DT for fall events classification due to
overcome this problem as discussed next, its low computational complexity and easy implementation on
embedded systems.
B. Wearable-based
DL in wearable devices is still considered new due to the
Common sensors used in fall detection devices are high complexity of the algorithms and memory constraints of
accelerometers, gyroscopes, magnetometers, and sometimes the device. Nevertheless, researchers gradually venturing into
barometers. Accelerometers have been the most commonly this approach centred on the feasibility study of embedding DL
used sensors, both as standalone wearable devices and as part into specific features of the MicroController Unit (MCU).
of larger systems [31]. In [32], a single ADXL 345 Work by [22] is considered the pioneer in this work, the
accelerometer with a threshold-based fall detector achieved up classifier proposed based on the RNN model with underlying
to 99.4% accuracy on the largest freely available dataset with relatively simple Long Short-Term Memory (LSTM) blocks
four stages of the methodology. Authors in [33] interfaced three suitable for wearable devices. However, the specifications of
sensors; accelerometer, gyroscope, and magnetometer in their the device used such as processor type, memory storage, and
system; this sensor fusion integrates several ML algorithms for computational speed are not stated in their work. Therefore, the
obtaining the best comparison. These methods could implementation process is quite vague for further reference.
potentially compensate for the erratic readings of the wrist's Authors in [45] tested 16 RNN architectures before being
movement where the sensor data were collected. The effort embedded in wearable devices. The best two of the architecture
seems worthwhile as the system yielded reliable output models were selected for future implementation in their work
parameters through the use of SVM. The barometer added to in [1] where they chose two SensorTile® devices developed by
improve the current detection methods, operates by providing STMicroelectronics for the integration and performance
pressure readings when a person falls. It can be attached to the analysis of the trained models. SensorTile [46] (STEVAL-
user’s garment or fixed on the floor [34]. The approach in the STLCS01V1) with a built-in 3D accelerometer and 3D
head-worn device proves useful for detecting falls even if the gyroscope modules. This system also provides lower economic
acceleration reading is relatively small [12]. costs compared to context-aware systems. As for
For the wearable embedding DL approach, most researchers disadvantages, these devices need to be worn by the user and
prefer to use a single sensor to reduce power consumption [35]. must be charged periodically. The power consumption of
To increase the device’s battery, careful consideration in all wearables depends on the device configuration, type of sensors,
and communication technology used. Work [32] highlights this for low-power and low-current applications (current usage is
issue and has assured good results regarding FDS battery 17.4 mA for transmission and 18.8 mA for reception) and may
lifetime in their work demonstration. As stated in TABLE I, the also be turned off by the MSP430 for low-power operation.
only differences between those selected devices are the Chipcon CC2420 also available in Zigbee protocols. Zigbee
significant amount of their flash storage and STM32F411RE was used in [50] to detect falls from wrist-worn sensors and
offering higher processing performance. works well throughout the study. Zigbee has 100m radius of the
In [47], working on investigating microcontroller embedding working range and is a low power device [26]. A famous
RNN for real-time detection. Meanwhile, authors by [48] work communication module adapting the Zigbee protocol is known
on RNN’s architecture design for embedding into a wearable as Xbee. It has a reputational performance in a wireless
device. Both works manage to compute well-founded communication system. Furthermore, the ease of setting the
outcomes. module Xbee with its designated programming interface
2) Communication Medium software: XCTU, makes this device favourable among
In wearable devices, the communication medium is typically electronics hardware inventors
wireless transmission. Justification of selected protocols such 3) Sampling Frequency
as Zigbee, Bluetooth, or Wireless Fidelity (WiFi) depends on In wireless data transmission devices, sampling rate plays an
the system application. The SHIMMER platform in [43] used important role in contributing to the effectiveness of the
Roving Networks™ RN-41 Class 2 Bluetooth® module to detection other than powerful detection algorithms. An
communicate via an integrated 2.4 GHz antenna. The variety of experiment by [52] tabulates performance analysis comparing
wireless communication modules in these devices 1 Hz to 200 Hz sampling rates. It stated that, after achieving 50
complements the number of functional features available on the Hz, the performance no longer considerably improves. They
system. Bluetooth was considered a data transmission medium also conclude that higher sampling rates do not necessarily
due to the feasibility of integration of the module and the low improve performances. 50 Hz to 100 Hz is well enough to
power consumption featured [1]. However, Bluetooth only has produce good results with any method while maintaining a
10m radius of the working transmission range, which is reasonable computing cost.
inconvenient for long-distance detection applications.
On the other hand, nowadays, devices are widely available in IV. PROCESSING AND LEARNING ALGORITHMS
WiFi-enable types [49]. Researchers in [10] introduce an In the following, we overview the processing and learning
Internet of Things (IoT) -based device-type invariant wearable algorithms based on the fall detection
FDS that allows any client devices with IoT ability such as
smartphone, NodeMcu, Raspberry Pi Model B3+, and Arduino A. Thresholds
integrated WiFi modules to interface with necessary fall In the early era of emerging FDS, two leading types of FDS
detection modules and send data to the server. The ability of algorithms were threshold-based and ML, researchers used
WiFi to handle multiple connections at once brings out the simple threshold-based algorithms where such algorithms vary
drawback of WiFi which is high energy consumption. Even from Quaternion, using sum acceleration and angle information
though, WiFi has a wide operating range of up to 100m, [53] to Kalman filter with preprocessing stage [32]. These
integrating WiFi with energy-constraint devices is seemingly threshold-based algorithms have shown an acceptable high
unfit due to the impractical of correlating high-power demand performance in terms of detection effectiveness and low
by WiFi module embedded with a low-power sensor or device. computational complexity. Furthermore, threshold-based
Meanwhile, [43] used a radio transceiver, Chipcon CC2420, algorithms are typically designed to minimise computational
and gigaAnt 2.4 GHz Rufa™ antenna. The CC2420 is designed overhead. However, the threshold values may vary with the
TABLE I
placement of sensors and individual activity patterns. In [21],
PROCESSORS’ SPECIFICATION IN WEARABLE DEVICES the authors state that the thresholds and positioning of sensors
impact the accuracy of fall detection. A series of observations
Ref. Device Processor Storage Speed
were made by adjusting both the thresholds of acceleration for
Arm®
[44] STM32F103 Cortex 1 MB flash memory 72 MHz
fall detection and the placement of sensors to improve the
®-M3 system’s performance. Sensor placement on the waist resulted
SensorTile in fewer false positives than on the shoulder and foot [54].
1 MB flash memory
[47] STM32L476JG
128 KB SRAM Nevertheless, over time, threshold-based algorithms present
Y
difficulties adapting to new types of falls and user
STM32 characteristics. Thus, this system becomes less favourable for
Nucleo-64 professional use [12]. Moreover, threshold settings may be
NUCLEO- 1MB flash memory different across subjects because of differences in their ADL
128 KB SRAM
L476RG patterns. A wrong threshold could lead to lower accuracies in
STM32L476R Arm® fall detection [55]. Therefore, the limitation is then overcome
G Cortex
[1] ™-M4 80 MHz
by ML methods, which are considered more sophisticated
STM32
32-bit approaches to solving this problem. To compare threshold-
Nucleo-64 MCU 512 KB flash
based and ML-based mechanisms, extensive research on the
NUCLEO- efficiencies of various ML techniques for fall detection has
memory
F411RE 128KB SRAM been conducted [33]. The work concludes ML yielded much
STM32F411R better results than the threshold-based as detailed in the
E following subsection.
STEVAL-
1 MB flash memory B. Machine Learning
[48] STLCS01V1 128 KB SRAM
STM32L476JG The alternative, ML-based approaches use supervised or
unsupervised algorithms on large datasets to train classifiers, The authors successfully obtained the overall result by
hence, building the ability to recognize a fall [56]. The combining AdaBoost with the SVM. The AdaBoost classifier
inference model can also detect falls for other input data. Most is a strong classifier composed of many weak classifiers. The
of the studies employed supervised learning technique SVM-AdaBoost classifier result gives an accuracy of 99%, a
algorithms, commonly Support Vector Machine (SVM) [57], 3% improvement compared with SVM.
and k-Nearest Neighbour (k-NN) [58]. Comparison between 2) K-Nearest Neighbor (k-NN)
ML algorithms is also among the favourite techniques to Quadros et al. [33] developed FDS and presented two
identify the best performance algorithms [50]. The study in [33] different approaches. The first was related to the threshold-
considers the effectiveness of different algorithms: k-NN, based method with Madgwick’s Decomposition (TBM-MD).
linear discriminant analysis (LDA), logistic regression (LR) Then, five different ML methods were evaluated. Additionally,
and classic decision tree (DT). The work shows that k-NN as a feature extraction stage for fall detection, a comprehensive
achieves the best results. study of the spatial orientation and movement breakdown in
Different ML models have different learning-based vertical and non-vertical components is presented. Eventually,
techniques; factors to consider may include feature sets, sensors k-NN achieved the best performance among other ML
employed, sensors’ placement, algorithms applied, the dataset, algorithms although SVM obtained the least false alarms
performance parameters, and so on. DL, the enhancement of during fall detector evaluation.
ML more or less overcome these factors in their outstanding Other than detecting falls, [29] also aims to avoid the risk
performance [59]. of future falls by recognising multiple fall patterns. The
The three ML algorithms listed in TABLE II. are examples differences between these two are, that the former detects all
of well-known algorithms that are frequently used in FDS as kinds of fall patterns into binary output (fall or normal), while
well as remarkable mentions of other ML algorithms. SVM is the latter specifically classifies patterns of falls into their
regarded as the prominent employed ML in FDS [60], followed respective fall-types (fall forward, fall backwards and so
by the k-NN and DT, thus, filling the selection for this forth). SVM, k-NN and RF are proposed for the performance
presentation. To preserve relevance, the reference period spans investigation. The highest accuracy in fall detection is the k-
from the years 2018 through 2022 and there is yet any paper in NN classifier, whereas, in the case of falling activity
2023 by the time of writing. recognition, it achieved the highest accuracy of 96.82% using
1) Support Vector Machine (SVM) RF.
The authors in [57] aim to employ a low computational-cost 3) Decision Tree (DT)
and highly accurate machine learning-based fall detection Work in [44] implementing DT in their system, data obtained
algorithm. They propose new features based on the first and solely from a triaxial gyroscope. Authors claimed that DT is
second-order moments, extracting 12 new features with an favoured as the classifier because of its low algorithm
SVM algorithm. Eventually, both works proposed to improve complexity and easy implementation on embedded systems. In
the detection performance with an excellent result and [63] applying the IoT technology in their system where
overcome the drawbacks and limitations of demanding detection occurs on cloud computing instead of on physical
computation and resource constraints allocation. A hybrid of workstations, this method is tolerably faster in giving output
SVM, the Multiple Kernel-Support Vector Machine (MKL- compared to common offline detection on the workstation. The
SVM) used to classify complex fall-like events was deployed result for each type of fall and activity detected shows an
in [61], helped by threshold-based methods (TBM) beforehand average of 91.67% accuracy and 93.75% precision. The work
to distinguish most of the ADL data. The MKL-SVM is an showed an FDS architecture design that combines big data
effective ML tool that allows the optimal combination of techniques to improve a DT algorithm continuously. Initially,
multiple kernels to build a single kernel SVM. The technique the algorithm was trained with a subset of activities from the
helps reduce false alarms while maintaining low computation SisFall dataset to classify three different fall classes and ADL.
costs on average, resulting in lesser power consumption. It was later, tested with data obtained from empirical
Meanwhile, in [62], the authors presented a benchmark experiments and achieved good results
database to evaluate the performance of detection algorithms. 4) Other Machine Learnings
Results show that each algorithm exhibits different In [64], four different ML algorithms were analysed using
performance ratings when combined with different features. two combined datasets: k-nearest neighbors (k-NN), artificial
TABLE II
NOTABLE ML USED IN FDS
Machine Learning
Ref Contribution
SVM k-NN DT Other
LR, Various evaluation methods; sensor types, device orientations,
[33] ✓ ✓ ✓
LDA algorithms.
[56] ✓ Propose new features
[61] ✓ Combined algorithm; TBM + MKL-SVM
[29] ✓ ✓ RF Using 10-Fold CV for performance parameter
[44] ✓ Extract three features from the dataset
[63] ✓ Combine big data for improvement of the algorithm
ANN, Employ multiple ML algorithms and datasets
[49] ✓ ✓
KFD
QSVM Two combined datasets, propose a new feature set
[62] ✓ , EBT,
ANN
RF, Compare five ML algorithms
[58] ✓ ✓ ✓
GB
neural network (ANN), quadratic SVM (QSVM) and While work in [67] solely tests CNN on various datasets and
ensembled bagged tree (EBT). The main contribution to this out of 14 repositories, the SisFall dataset achieved a prominent
research area is the proposal of new features obtained from accuracy, sensitivity and specificity of 99.22%, 98.64%, and
accelerometer and angular velocity data that lead final accuracy 99.63%, respectively.
of 100%. EBT achieved the best overall performance of 94.1% In [68], CNNs are trained on different datasets of optical
in raw acceleration data input detection and 97.9% after extract flow images, which aids the network in detecting various
features application. Work in [50] also performs multiple ML actions. Then, from action recognition through fall detection,
comparisons, data input from sensors fusion between transfer learning is used. The experiment was carried out on
accelerometer and gyroscope to multiply the successful rate. three separate datasets, with an accuracy rate of above 95% in
Gradient Boosting (GB) achieves the best result with an each case. Harmonisation of fall datasets was performed in [18]
accuracy of 98.7%. where the majority of data was collected on the wrist but also
Such work in [64],[58] and [62] compare several ML included data from the waist and other positions. The work
algorithms which include ANN in their list, the ANN demonstrated that small-scale CNN for embedded systems
supposedly contains one hidden layer which could be part of could achieve reasonable performance with the designated
DL depending on the complexity of the data, however in these dataset.
works, the ANN applied, assumed as the simple and In wearable FDS by [68] explore the application of CNN to
lightweight version. Thus, computationally equivalent to other a stream of sensor data collected from Body Sensor Networks
common ML algorithms since it cannot outperform them. ML (BSN). The data were converted to images before being fed into
algorithm works better with large datasets. The dataset must CNN, and the final accuracy result shows 92.3%. The data is
undergo feature extraction and additional preprocessing then released to the public and known as Kinect Activity
methods before entering the ML algorithm for classification. Recognition Dataset (KARD). The proposed Multi-source
Some of the feature extraction even requires manual extraction CNN Ensemble (MCNNE) performs better than the normal
by using certain techniques. In a different case in DL, the user CNN structure. The system takes data from the smart insole
only needs to provide datasets to the processing framework. incorporating an array of four pressure sensors, a three-axis
The applied algorithm will perform the whole learning pro-cess acceleration sensor, and a three-axis gyro sensor as the input of
which eases the user's work. However, from all the ML the base classifier independently.
algorithms presented, no specific ML algorithm is superior to Concurring with the statement, authors [69] introduced DL in
others. A finding by [62] shows that different algorithm's their FDS as it can learn the representations needed for
detection accuracies, specificities, and sensitivities results can detection or classification from raw data without costly human
vary based on the combined features of the dataset used. The involvement. However, DL-based methods need to transmit
statement indicates that the algorithm’s performance changes data to the server for processing because it is difficult to be
depending on the features presented or extracted being fed into deployed directly on embedded devices. Although this solution
that algorithm. Therefore, the decision on choosing ML to be may bring some delays, in such scenarios as nursing homes, it
deployed corresponds to the aim, target user, specific data and facilitates the unified monitoring of the elderly's movement
system application. status in the park to ensure timely rescue in case of falls.
However, the method in [70] still needs a large amount of data
C. Deep Neural Network
to detect a fall, which will consume a lot of network bandwidth
Recently, deep neural networks have risen to the top of the in an IoT system. proposed a fall detection algorithm based on
scientific community's list of most promising and discussed CNN-LSTM combined network, which uses neural networks to
research paradigms. These approaches have outperformed automatically extract features and classify them, avoiding the
many traditional algorithms in various domains, ranging from cumbersome preprocessing process.
common computer vision tasks like human reidentification and Nait et al [71] conducted a comparative study of three DL
object recognition to sequence-to-sequence language model methods namely CNN, LSTM and a combination of
translation [26]. ML decides data based on settings done these two that is ConvLSTM to identify the ability of DL to
manually by humans, while DL methods are capable of automatically derive features from raw accelerometer data that
automatically identifying the characteristics needed for assess fall risk. Results show that the DL methods in a single-
detection or classification from raw data without extensive task learning mode outperform in recognizing the risk of falls.
human intervention [65]. In addition, the use of DL techniques The various DL models are achieving outstanding performance
for fall detection with wearable devices has been an area of across several disciplines, such as in health care systems, high-
recent interest [8]. TABLE III presents details of DL algorithms dimensional numerical data, forecasting weather data, image
used in previous works. and video processing, speech recognition and natural language
1) Convolutional Neural Network processing applications.
Work in [66] proposed discrete data within the window’s However, the application of CNN is commonly favoured in
method and continuous time-series data classifiers. The vision-based detection systems which rely on image processing
continuous time series model performed better performance techniques on the video frames or images captured by cameras
than the discrete data-based model when classified as a binary around the region of interest (RoI) instead of performing
class and vice versa for multiclass. Nine classifiers were parameter monitoring of the subjects. One reported
proposed with feature extraction; feature calculation and disadvantage of this method is that it is vulnerable to mistakes
infinite latent feature selection (ILFS); CNN comes out at the caused by variations in ambient lighting. Therefore, wearable
top, beating LSTM; however, preprocessing is computationally sensors which usually produce sequential data output are more
heavy for memory constraint devices and an unnecessary step suitable to be processed by recurrence processor algorithm to
for the DL algorithm. Overall, this study’s continuous time- predict and detect fall occurrence based on the continuous
series data-based fall showed a higher classification rate than parameter of the data. This is where the recurrent neural
previous research with 99.8% accuracy of the CNN algorithm. network comes in handy to match the wearable sensor
 TABLE III
COMPARISON BETWEEN ML AND DL FROM THE LITERATURES
Compared Algorithm
Ref Machine Learning Deep Learning
ML DL
LSTM-Acc, LSTM- Produces few false positives and true Produced only a single false positive
[80] Acc + SVM-Depth
Acc Rot negative result
Restricted to learning from a small set
Automatically learn subtle features from
[79] SVM, Naïve Bayes RNN (GRU) of manually specified extracted
the raw accelerometer
features.
Need to manually extract features for The neural network automatically extracts
[69] SVM CNN-LSTM
training features and classifies them
MLP, DAE, CAE,
[83] XGB, k-NN, SVM, DT Human design features are needed Extract features automatically
CNN
Naïve Bayes, k-NN,
Significant in performance from Best DL showed more than 99.8%
[66] SVM, RF, QDA, ANN, LSTM, CNN
different features selection accuracy with fewer FNs and FPs
NN-DTW
Hard to implement on an embedded Improves the specificity, that is it
[48] SVM RNN (LSTM)
system minimizes false positives in fall detection
All ML compared, unable to Capture non-linearity relationship
[84] HMM, RF, SVM RNN (LSTM) between the input and output layers
outperform a single DL
efficiently
RNN (LSTM & Require an extensive feature extraction Higher accuracy and efficiency in fall
[7] SVM, k-NN
GRU) stage detection.

characteristic. The CNN [68] captures spatial correlation in of memory occupation, computational load, and power
data, while the RNNs' recurrence captures temporal consumption.
correlations. A study in [8] mentioned that often ML works better with
2) Recurrent Neural Network data fusion of multiple sensors, and the technique is quite
Early embedding of RNN in wearable FDS shows the results inconvenient for a system with computational constraints.
obtained outperformed the statistical classifier proposed in However, in [42] attempt to distinguish falls from ADLs using
SisFall: C9 indicator [72]. The most remarkable contribution of only a single accelerometer data applied to multiple ML
this work is the temporal labelling in the SisFall dataset used algorithms for comparison, different algorithms score highest
for model inferencing [22]. The label provided consists of two in different performance parameters. Thus, choosing the best
types, binary-label out-put which consists of fall and ADL, and algorithm will depend on the system's concern. In [76], the
three-label output: Alert, fall, and ADL. The difference authors demonstrate the advantages of combined deep
between these two types of labelling is the ‘Alert’ train system architecture based on convolutional and LSTM recurrent units.
to identify pre-fall conditions, enabling researchers to choose The framework outperforms by competing for deep non-
suitable labels that cater to their system. While RNN proposed recurrent networks on the challenge dataset by 4% on average.
in [47] addresses the feasibility of implementing the DL The authors stated that depending solely on overall accuracy
algorithm in the embedded system by presenting a few general measurement is not an appropriate method. Thus, they added
formulas regarding memory occupation, battery consumption, F1-score for both datasets to be 0.930 for OPPORTUNITY and
and computational load. [73] specifically address these 0.958 for Skoda. In [77], works on parameter optimisation
challenges by proposing SaveMeNow.Ai, a new wearable experiment for LSTM obtains the best regularization rate of
device for fall detection. The work preliminarily created a new 0.015 at 128 hidden layer nodes for best accuracy with the best
dataset by merging four datasets available online, then tested computation time. The work also contributes to resourceful
different classification algorithms and found DT as the most outcomes for reference in terms of other parameters testing
satisfactory algorithm. with LSTM.
Research work [74] first proposes LSTM blocks of RNN RNN is different from classical ANN in terms of the hidden
for online detection for wearable devices. Later on, [45] state feature, which is important for remembering some
modified the RNN architecture into 16 versions and adapted information about a sequence of data. After producing the
the Gated Recurrent Neural Network (GRNN) of LSTM and output, it is copied and sent back (backpropagate) into the
GRU, thorough analysis of the feasibility of embedding RNN recurrent network. For making a decision, it considers the
into wearable devices conducted, except the embedding was current input and the output that it has learned from the
performed in their next work in [1]. The extension continues previous input. In traditional neural networks, all the inputs and
implementing the proposed RNNs architecture, except, the outputs are independent. Previous layers are not considered for
best four were selected to identify the simplest yet reliable the next layers’ input; thus, they do not memorize the previous
algorithm suitable for wearable implementation. The final outputs.
architecture consists of a batch normalization layer receiving Among the prominent DL methods mentioned in FDS are
the input, an RNN layer and a dense end layer. Meanwhile, Convolutional Neural Networks (CNN) [8] which often being
[75] has implemented two different versions of RNN integrated with image processing techniques, and Recurrent
architectures into three different FPGAs and validated their Neural Networks (RNNs) [78], the best approach for
work with [22]. The proposed architecture outperforms the processing time series signals [79]. The work in [80] even
SensorTile implementation particularly, the processing time compared traditional ML (SVM, Naive Bayes) techniques with
is about one order of magnitude lower, and three orders of RNN. As a result, RNN shows superior fall detection
magnitude reduce the power consumption. Researchers in performance without large training datasets.
[47] also refer to [22] for RNNs architecture however focus
on the technical outcome of the embedding hardware in terms
D. Machine Learning vs Deep Learning dataset for training and validation using an inference model.
Comparative discussion between conventional ML and DL However, as mentioned before, the classical ML algorithm is
has been progressively researched. Several findings less favoured because the dataset has to undergo several stages
highlighting the opposition between these two methods are of preprocessing and feature extraction before applying ML for
shown briefly in TABLE III. final classification. Furthermore, most works integrate binary-
Among earlier collation between these two approaches is label output in their system to reduce resource usage.
apparent in [81], the authors proposed an RNN model, named Meanwhile, DL, an advanced version of ML can directly
LSTM-Acc and a variant LSTM-Acc Rot, to compare results classify output without initial preprocessing conditions. The
against extant research, Acc + SVM-Depth [82] and UFT [83], algorithms could be computationally complex yet less hassle
the URFD data set was used. The LSTM-Acc Rot obtained the for the developer. Moreover, researchers worked on optimizing
best performance with an accuracy of 98.57%. Then in [80], the architecture models to reduce the complexity without
they compare DL against two ML algorithms, and the GRU compromising the quality of detection output [47].
generally outperforms them across three datasets tested. The Work in [80] performed offline algorithm evaluation using
authors added, that the DL model has a greater ability to three datasets and conducted real-time detection as final
generalise to new users when forecasting falls, which is a validation to obtain the best-performing algorithm. The DL
crucial attribute for any model to be effective in the real world. eventually outperformed all ML algorithms in both offline and
Work in [70] claims that integrating DL: CNN-LSTM in the online tests. Besides that, the authors mentioned DL could use
classification system can avoid the hassle of manual feature raw data without any preprocessing and even achieve the best
extraction by the conventional ML. Moreover, despite the small result over traditional ML. The same goes for [84] who also
data volume, the proposed method outperforms ML with higher mentioned that feature extraction needs to be performed for
detection accuracy. Wang, Gaojing, et al. [84] introduced conventional ML in their work while for neural networks,
several DLs in terms of lightweight neural networks to features can be extracted automatically. All DL algorithm
implement them in wearable devices feasibly. The DLs were results show higher accuracy than 99.5%, which was even
compared with conventional ML with similar parameter higher than the best accuracy from the conventional ML.
settings and achieved the best result from lightweight CNN. DL However, the main concern of the author is more on
models developed by [7] manage to fulfil the resource- implementation, the final CNN model came out with 1.2KB of
constrained requirement with acceptable high-accuracy results storage size which considered suitable to be implemented in
to be deployed in the fog node. wearable devices.
Nine ML algorithms were proposed to detect the prior impact Nevertheless, algorithm performance depends on both
of falls, the authors tested on discreet and continuous data types datasets and parameters. Some works focus on their dataset
and still managed to get the best performance from the DL adaptability with the algorithm applied, while others, work on
algorithm [66]. As predicted, LSTM achieved the best parameter adjustment to get the best results. Works by [70] and
performance among the ML; Hidden Markov Model (HMM), [87] specifically mentioned their parameters in detail however
Random Forest (RF), and SVM. The method proposed by [85] they do not justify their choices. Other works used the
namely FDS, uses discrete wavelet transform to denoise data. parameter by defaults adopted from previous works [52]. There
Then the RNN model is utilised to classify human motions and is not much common basis among the different fall-detection
identify the fall status automatically. It also achieved the best studies. Every re-search group takes their own decisions on the
performance overall, thus resulting in robustness to the random types, quantities, and configurations of the sensors, the
noise derived from the environment. participants, the assessed and compared algorithms, the size of
On the contrary, authors in [86] experiments carried out on the datasets and the method used to gather the data, as well as
the SmartFall dataset. The experimental results show that the the performance metrics. As a result, the data offered as
RF algorithm outperforms a single deep LSTM model and other performance measurements in various articles are not
different ensemble techniques. It is identified that the DL comparable.
overpowered by ML was computationally complex and
resources. Despite being known as more complex in terms of V. VALIDATION METHODOLOGIES
computation, work in [48] has proven that, with the proper AND SAFETY CONSIDERATIONS
algorithmic design, the RNN is feasible to be embedded in A. Safety Standards and Guidelines for Detection
wearable hardware devices. On the other hand, implementing
Wearable technology standardisation is still in its early
ML is said to be challenging due to a large number of features
stages. In 2021, IEC TC 124 released IEC 63203 [88], which
and the ensuring requirements in terms of memory and
defines terminology often used in connection with wearable
processing power
Nonetheless, datasets are crucial in ML algorithms; the large technology, the contents of the standard comprise wearable
training dataset is preferred to obtain optimal performance as a electronic gadgets and technologies, electronic textiles, near-
body wearable electronics, on-body wearable electronics, and
small dataset may lead to overfitting. Despite the concern
in-body wearable electronics.
mentioned, researchers in [22], with the right techniques such
Since wearables and connected body sensor devices ae
as stacking in DL models architecture, could mitigate the
scarcity of data for training a fall detection model. Furthermore, considerably used in the Internet of Things (IoT), wireless body
once trained offline, neural network-based models, require are networks enable data sharing in smart environments such as
smart homes, and smart life. Upon specific digital healthcare,
fewer computational resources during real-time classification
wireless connectivity between the sensing nodes and the edge
tasks compared to traditional ML methods which rely on
computing device requires a standardized communication
feature extraction and processing. This shows that DL models
interface and protocols. Thus, the present IEC 63203-801-1
for low-power device implementation are achievable and have
since been current attention by researchers worldwide [30]. [89] describes the following physical layer (PHY)
Applying ML in fall detection requires a reliable mass specifications, such as packets formats, modulation; and
forward error correction. On the other hand, IEC 63203-801-2 Prediction Outcomes
[90] specifies low complexity medium access control (MAC) Positive Negative
for SmartBAN, such as channel structure; MAC frame format,
and MAC functions. True False
The shape of the sensor is also among the sought-after aspect Positive Positive Negative TP + FN
in wearable devices; other than being small in size and (TP) (FN)
convenient, proper design more or less affect the user’s Actual
experience; the current shape of rectangular FDS tends to hurt Value
the wearer with their pointy-edge design when accidentally
pressured onto it [10]. IEC 60601 [91] specifies a process for a False True
Negative Positive Negative FP + TN
manufacturer to analyse, specify, design, verify and validate (FP) (TN)
usability, as it relates to basic safety and critical performance
of medical electrical equipment. Data processing in wearable
devices is preferable due to the privacy features offered Fig. 5. Component of Confusion Matrix.
compared to sending data to the cloud server for processing,
which information breaches could occur amidst transmission receiver operating characteristic curve (AUC-ROC), among
[75]. The embedding of ML in wearable devices allow others. To evaluate the accuracy of these performances,
researchers to overcome latency issue in term of detection time sensitivity and specificity which describe the capability of the
and protect the privacy of the data from leaking to other third proposed system to identify falls and differentiate them from
parties during data transmission to the nearest computational ADLs are measured based on the equation as follows:
resource, such as wearable sensor do not perform detection
directly on the device will require it to transmit fall data to the 𝑇𝑃 (1)
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = × 100
workstation to perform the task. Therefore, before developing 𝑇𝑃 + 𝐹𝑁
and embedding ML algorithm into the device, knowing the
device’s specifications is crucial to select the suitable wearable 𝑇𝑁 (2)
to be employed in FDS. 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = × 100
𝑇𝑁 + 𝐹𝑃
B. Performance Evaluation and Metrics
Regarding fall detection, the explored methods include 𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 + 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 (V)
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
threshold-based, conventional ML, and DL. The latter two are 2
prominent, owing to their robustness [26]. Nevertheless, most
researchers acknowledge that DL able to outperform ML in sort While various metrics are used in this field, not all are
of categories including latency, output reliability, data necessary for evaluating the FDS's performance; their selection
adaptation and others. An illustration of the processing stages depends on research objectives. Such that the work by [57]
between ML and DL in Fig. 4. shows one of the advantages of intends to make the algorithm embeddable in wearable devices,
DL over ML. Relatively, DL possess the ability to skip the Therefore, they propose cost-effective ML-based algorithms.
preprocessing stage to produce output instantly [80, 84]. Besides the famous trio, the evaluation method is anticipatedly
Common evaluation metrics employed in FDS research related to resource utilisation and optimisation which includes
include the renowned trio: sensitivity/recall (the proportion of false alarms and computational complexity. The complexity
true positives among all actual positive instances), specificity calculation uses floating-point operations (FLOP) and achieves
(the proportion of true negatives among all negative instances), 99.9% accuracy at a computational cost of less than 500
and accuracy (the proportion of correctly classified instances). floating-points per second.
These performance metrics are correlated. By using the values Another attempt by [47] also performed embedding on a
from the confusion matrix as shown in Fig. 5., various wearable device. The feasibility of LSTM implementation is
performance metrics can be calculated to evaluate the model's evaluated with three abstract metrics presented about
performance. computation power, memory occupancy and battery duration.
Other notable metrics include precision (the proportion of The metrics are general as they rely on parameters describing
true positives among all positive predictions), F-score (the the values specific to wearable’s MCU which is structurally
harmonic means of precision and recall), and the area under the invariant across multiple MCUs. The implement method
achieves real-time compliance and up to 130 operation hours
without charging.
A work by [86] wanted to predict pre-fall situations by
proposing stacking-based ensemble learning of DL models
namely LSTM with various configurations. They also included
conventional ML algorithms to predict the risk of falling in the
elderly. By integrating a public dataset, the work addresses the
challenge with limited training instances and imbalance class
labels exist in the dataset. Hence, F1-score, ROC AUC and
geometric means (G-mean) are included in the evaluation
metrics. The latter two commonly address the importance of
both classes and the former emphasises how both false negative
and positive could be costly for the system. The work obtains
RF outperforming single deep LSTM model and different
Fig. 4. Machine Learning vs Deep Learning. ensemble techniques.
 The feasibility of integrating multiple technologies in FDS positioned their sensors at eight different locations on the body
also become a common interest these days. In corresponding and concluded that sensor placement on the body trunk gives
with the emergence of IoT, device-type invariants with the best result.
embedded accelerometers are authorised for fall detection [10]. Placing the sensor on the chest is less obtrusive compared to
The analysis was performed at the server in real-time with a other favourable places such as the waist. Both the waist and
constant of 190 ms per cycle from the second cycle onwards. chest hold almost the same status of steady sensor reading [96],
The first cycle took relatively longer or about 790ms probably sometimes the differences in sensitivity and accuracy among
due to the initialisation process before the system could the different sensor locations are relatively low and negligible
familiarise itself with the network. Nevertheless, response time [93]. Nevertheless, the chest is supposedly a suitable placement
depends on variables such as connection speed, network for the FDS sensor in terms of compromising convenience and
conditions and others. For IoT-enabled devices, power detection accuracy, because people tend to take these devices
consumption also varies according to the device's features. off when they feel sudden uneasiness or before going to sleep,
The heterogeneity of evaluation metrics was sidetracked as making the FDS ineffective [52]. Therefore, proper placement
researchers focused more on parameter attenuation with of sensor is crucial in FDS. Other than the ease of placing it
combined detection methods, such as Threshold-DL (STM- onto the garment, it is favoured to be less disturbing when
LSTM) ) [77], and CNN-LSTM [92]. The former proposed five performing certain movements such as bending down, sitting
types of parameters, while the latter presented 14 combined DL down and climbing stairs. Moreover, Several studies have
models, both evaluated using common evaluation metrics. mutually agreed that arms and legs are not suitable parts of the
Another notable metric, algorithm complexity refers to the body to carry a fall detection device since they are typically
measure of resources required by an algorithm to solve a correlated with higher accelerations than the actual condition
problem. The two main factors that determine the complexity [97].
of an algorithm are time and space. To analyse the complexity Due to the mass reference of SisFall dataset [72] which
of ML used, Big-O notation is used to determine how well an performs data acquisition with sensors located around the
algorithm will perform in practical scenarios such as resource- waist, researchers gradually diverted their focus to algorithms
constrained environments of wearable devices and to make and hardware feasibility rather than paying attention to the
comparisons between different algorithms. The metric is location of sensors. Fig. 6. illustrates several placements of
mentioned briefly in [47]. TABLE IV summarise work by order detection sensors on the user. According to [75], wearable
and preference of metrics evaluation. devices of FDSs are made with particular care in mind; they
have to be compact and lightweight to provide comfort to the
C. Sensor Placement
wearer. The authors added that the combinations of
The next essential aspect of obtaining reliable sensor miniaturised electronic components supposedly could be worn
readings for fall detection is the sensor placement; the selection by a user beneath, with, or on top of garments, containing both
of location is unique to every researcher; it depends on the sensors and a processing unit [47]. These details in fall
system's purpose and justification. There are several options for detection devices are supported with well-performed
the placement of wearable devices on the human body, the most processing and learning algorithms are reviewed in the next
common places mentioned in the research are the head, chest, section.
waist, wrist, thigh, and ankle [93]. Based on [1], favour goes to
waist placement as it is close to the body’s centre of gravity D. Data Considerations and ML systems
[94] . Some claim chest or head is better [95], while authors in Over the past years, human activity recognition researchers
[50] mentioned better results are achieved when sensors are have constructed databases of different human activities in
placed along the longitudinal axis of a body such as a waist and many scenarios. These databases can be the validation for the
a chest compared with other placements. In [31], they identification of daily human activities. However, they are
TABLE IV usually insufficient because some databases contain only a few
SUMMARY OF EVALUATION METRICS PREFERENCE types of fall data and cannot represent complex fall situations.
Evaluation Metric TABLE V. present the publicly available datasets and notable
Work Dataset Year Sensitivity, details of these databases. Choosing the right dataset for model
Specificity, Others training and validation is critical.
Accuracy Moreover, reliable datasets dictate how well ML algorithms
[56] ✓ FLOP perform, complimented with proper parameter setting. Most of
Wearable
embeddable SisFall 2019 Memory the existing datasets are based on simulated fall data and
occupancy,
ML [47] ✓
algorithm FLOP, power
consumption
DL; forecast fall and
F1-score, G-
ensemble learning; Smart-
2020 ✓ mean, ROC
forecast risk of falling Fall
AUC
[85]
Responsive
IoT-based device tFall 2020 ✓ time, power
invariants FDS [10] consumption
URFall
Propose topologies of Detection
2021 ✓ -
a multimodal CNN and
[91] UPFall

Propose combined Own 2022 ✓ -

detection method [76]
Fig. 6. Range of sensors placements
.
 TABLE V most complete among the others. In fact, including the largest
LIST OF FALL DATASETS USED IN THE LITERATURE, THE NUMBER OF
amount of data in both number and heterogeneity. Besides,
SUBJECTS INVOLVED, ACTIVITIES CONDUCTED AND DEVICE(S) USED IN THEIR
RESPECTIVE WORKS Mobifall [103], tFall [104], DLR [105], and Project Gravity
Number of
[106], none of these datasets includes elderly people which is
Dataset
Subjects
Activities Sensing Device the favourite subject in FDS. Therefore, working with complex
URFD NS 70 Custom algorithms, and valid and reliable databases plays an important
DLR 16 6 Smartphone role in producing a high-quality system.
tFall 10 8 Smartphone
MobiFall 24 13 Smartphone E. Ethical Testing
UniMiB
SHAR
30 17 Smartphone The latest technical developments in wearable devices open
SisFall 38 34 Custom the field to various applications. Parameters such as blood
UMAFall 17 11 Custom pressure, pulse rate, and temperature can be tracked easily and
SmartFall 7 NS Smartphone conjointly paving the way for healthcare applications which
NS = Not Stated focus on intelligent and wearable sensors. In context data
although real fall data exist [98], they do not make these acquisition by wearable sensors, automatic fall detection has
datasets publicly available. Therefore, the system only provides recently attracted the scientific community’s attention [107].
insight into the real-world situation through performance As mentioned in section A above, data processing on-board is
results. Several databases of human behaviours containing fall more practical to reduce data breaching compared to sending
incidents are proposed in [99]. Most of these datasets include data to the cloud server for processing.
virtual tasks re-enacted by young volunteers, which may be Nevertheless, the advantage of cloud computing could be
difficult to convert into a real-world scenario [64]. utilised to overcome overhead at the resource-constraint device
Authors in [81] deal with image data from the UR Fall without compromising privacy. Two-level privacy protection
Detection (URFD) dataset in their evaluation method; could be implemented; personal data from the sensor is stored
uniquely, the dataset contains both acceleration reading and and processed locally on the smartphone and only the
vision-based data providing extracted features of depth images. parameters to the server [108]. Despite any mishandling of the
The measurements from the accelerometer and the Kinect cloud server's administration, no data will be exposed.
camera have been temporally synchronized at 30Hz of Being designed for a specific purpose, several criteria and
sampling rate, so that information from all sensors is available feature designs of the wearable device are taken into account,
at each time step. Work in [100] claimed that the UniMiB such as the type of processor used, sensors available, and
SHAR (University of Milano Bicocca Smartphone-based medium of communication for flawless data transmission. In
Human Activity Recognition) dataset was difficult to analyse. addition, the sensor placement on the wearer needs to be
Patients of a wide variety of ages provided the dataset, which considered as well, and proper placement leads to a stable
included 11,770 activity data of 17 different sorts. average sensor reading, thus giving accurate and reliable output
Nevertheless, the reviewed system achieves a detection rate of for the system [21, 109]. Moreover, as metioned in Section A
91.5% with the applied CNN model, using 5-fold cross- above, the sensor's shape is a crucial consideration. Apart from
validation. the advantages of small size and convenience, the design
Work in [101] experiments their developed CNN model on significantly impacts the user's experience [10].
three different datasets (URFD, Multicam, and FDD) and Therefore, informed consent should be initiated in the
obtained the best performance from URFD with an accuracy of process of data acquisition, participants should provide
99% and sensitivity of 100%. The authors however added that informed consent before being involved in testing, and they
the method might only perform well for single-person fall should be informed about the purpose, risks, and benefits of the
detection, although multi-person fall detection is also crucial. system. Even better, a physician is available at the experiment
Meanwhile, two different datasets, namely Smartwatch and venue to monitor the process and to attend the participant if any
Notch are combined to make the SmartFall database. It was misconduct occurs [72].
used in [102] to test their generalised developed system, with It is important to practice transparency in the system's
an additional cumulative sum algorithm to improve operation. Data used for fall detection should be addressed and
performance. The system produces 96.79% accuracy. easily understandable to users, caregivers, and healthcare
The developed system by [67] decided to test up to 14 providers. Such that public dataset for fall detection is available
datasets including the notable MobiAct, Sisfall, Mobifall, freely on the website [74]. Most of the data suited with brief
UniMiB SHAR, UP-Fall, and others by analysing the data descriptions and anthropometric characteristics, real names and
characteristics and suitable features for fall detection. identification numbers are prohibited and against the privacy
Performance metrics using the proposed architecture all over protocol.
the 14 datasets are measured. Good results are obtained with F. Commercialised Fall Detection System
certain data repositories when it is hyper-parameterised. TABLE VI Shows some commercially available FDS, they
However, global evaluation with the other repositories are selected based on great values offered by the company such
emphasises the challenge of extrapolating the network as the system features, feasibility and pricing.
architecture configured and optimised for a specific dataset to
Medical Guardian provides five alternatives to choose from
other testbeds. The SisFall dataset achieved a prominent in terms of detection devices and connection types. The
accuracy, sensitivity, and specificity of 99.22%, 98.64%, and packages offered are comprehensive, whether for active
99.63% respectively. lifestyle, a homebody or a mix of both. However, the system
Based on the current update, numerous works approved the comes with a monthly subscription fee although users cancel at
SisFall dataset, and certain works [47] claimed that SisFall is any time. Bay Alarm Medical system comes with a GPS-
an adequate choice for testing and validating their FDS.
integrated feature and a tracking app at affordable pricing.
Moreover, authors [22] also added that SisFall is deemed the
 TABLE VI VI. FUTURE OUTLOOK
NOTABLE COMMERCIALIZED FDS
The current FDS system focuses solely either on fall
FDS Device type Connection type detection or fall prevention. As the name states, fall prevention
Medical Guardian
considers both fall data and pre-fall data readings before issuing
Necklace Pendant / Bracelet Landline / cellular
Bay Alarm an alert, while fall detection only takes fall data into its system.
Medical
Smartwatch Landline / cellular Fall recognition can be described as recognising a specific type
Lively Smartphone Cellular of fall. Moreover, fall recognition can be useful for medical
Alexa Together Compactible third-party WiFi assistants to mount the right response to handle possible
LifeFone pendant Landline / cellular complications regarding specific types of falls. However, to
Alert1 Wristwatch/Pendant Cellular recognise specific patterns of action, the concern is on the
complexity of the processing algorithm and massive dataset of
Lively comes with cell phones with built-in medical alert, also different fall patterns. The complex algorithm exploits high
at an affordable price. The only con is available only one resource consumption because more features must be extracted.
standalone medical device option. Alexa Together provides Hence, most researchers prefer distinguishing falls from the
excellent for those with mobility or vision issues but is only ADLs system to complicated fall recognition.
accessible by any Alexa-enabled device with monthly fees. Not Either ambient-based or wearable-based FDS as explained in
to mention its short range of roaming area from the connected Section III, both types cannot run from the privacy matter.
device. In this case, LifeFone provides a wide range yet solely Since detection requires close contact with the subject in terms
indoor coverage up to a 1,300-foot area. While other features of wearables and the subject’s presence in the area of interest
and equipment options same as other systems. Alert1 value its for vision/ambient FDS, issues on confidentiality are always
customers with discounted price option for bulk buying. top-notch. The initial solution for the issue is to reduce the
Extensive add-on features and equipment options [110]. Since degree of privacy intrusion; for example, researchers resort to
all of these FDS are officially in the market, details on the using Kinect [15] instead of using surveillance cameras to
sensor used and other specifications are confidential to the detect falls. It solely captures skeletal and depth image data,
public. with blurred information utilised to characterise human action.
G. Monitoring Against Abnormal Situations Meanwhile, for wearable devices for FDS, there is no apparent
discussion on answering privacy matters probably due to less
Monitoring against abnormal situations involves
intrusive methods that only involve inertial and/or motion
implementing various techniques and practices to ensure that
detection compared to ambient sensors.
the system can accurately detect and respond to falls while
Recently, the application of wearable devices in FDS has
minimising false alarms. Some key strategies to monitor and
slowly shifted their concern from the design of devices to the
address abnormal situations are;
feasibility of implementing ML computational programs into
i) Operational Design Development - The detection method or
the device itself. This is due to many requirements that need to
ML function used within a system is reliable and can produce
be attended to when embedding high computational algorithms
accurate outputs. However, despite the advancements in ML
specifically ML and DL as discussed in Section IV, into
techniques, it is crucial to acknowledge that these functions can
portable memory-constrained devices. Such as the work in [47]
still produce incorrect outputs due to errors in generalisation.
abstracts a series of formulae for calculating the memory,
ML model's ability to perform well on new, and unseen data is
processing power, and power consumption requirements for
based on what it has learned from the training data.
embedding a generic DL architecture on a microcontroller. The
ii) Uncertainty - Uncertainty can also be measured using
work mentioned the type of sensors and processor of
redundancy and majority voting. Redundancy aims to enhance
microcontroller used, but since the ML algorithm is trained on
the reliability of the system by reducing the likelihood of errors
a publicly available fall dataset, researchers focus on the
or false alarms. Whereas, a majority voting approach means
feasibility of embedding the algorithm into the device, instead
that the system makes a final decision based on the majority
of being concerned about the placement of the device on the
agreement among these components. This can contribute to a
user, moreover, there is no visual example on the final product
more robust and reliable detection system.
of the device.
iii) Regular System Maintenance -Ensure regular maintenance
Nevertheless, as years go by, researchers gradually pay less
of the FDS to check sensor functionality, system integrity, and
attention to this matter, instead, focusing on developing ML,
software updates. Address any issues promptly to maintain the
particularly DL, which has more computational complexity
system's effectiveness.
than ML to be embedded into wearable devices. ML techniques
iv) Environmental Context Awareness- Consider the
as discussed in can detect falls to a reasonable accuracy level.
environment in which the system is deployed. Abnormal
However, a wearable system consisting of a device such as a
situations may arise due to environmental factors like dim
wristband alone is insufficient to meet the requirements of a
lighting or obstacles in the way. Design the system to be
comprehensive FDS. In a study by [8], the authors observed
context-aware and capable of adapting to such situations.
that the systems are capable of generating alerts on detecting
Apart from technical challenges, monitors face a significant
falls, but they fall short in the ability to activate adequate alerts
implementation-related obstacle concerning imbalanced data
while minimizing the cost and power requirements. Therefore,
for “unknowns” or input outside the Operational Design
we highlighted works that provide the device specifications and
Domain (ODD). Many of the presented techniques here use in-
characteristics they used when integrating with the ML
distribution data in a self-supervised fashion and use Out-of-
algorithm for fall detection.
Distribution (OoD) data for calibrating the boundary. However,
The existing literature is still lacking in terms of wearable
to perform such a calibration and test the applicability, one only
specifications used in FDS that run embedded real-time
has “known unknowns” collected before the development.
detection algorithms. Several issues must be handled to embed
ML algorithms with FDS, such as memory resources,
computational power and power consumption. Other than that, [7] Sarabia-Jácome, D., et al., Highly-efficient fog-based deep learning AAL
fall detection system. Internet of Things, 2020. 11: p. 100185.
the placement of sensors as highlighted in Section V also plays
[8] Ramachandran, A. and A. Karuppiah, A survey on recent advances in
a vital role in determining the performance of FDS. The study wearable fall detection systems. BioMed research international, 2020.
on the sensor’s placement-base fall detection performance is 2020.
conducted by [94], the work investigates six different parts of [9] Singh, A., et al., Sensor technologies for fall detection systems: A review.
IEEE Sensors Journal, 2020. 20(13): p. 6889-6919.
the sensor location with several ML algorithms. The authors
[10] Nooruddin, S., M.M. Islam, and F.A. Sharna, An IoT-based device-type
claimed that their work is the most comprehensive study invariant fall detection system. Internet of Things, 2020. 9: p. 100130.
regarding this issue. [11] Pannurat, N., S. Thiemjarus, and E. Nantajeewarawat, Automatic fall
The ideal FDS will be an all-time running device with a monitoring: A review. Sensors, 2014. 14(7): p. 12900-12936.
[12] Cola, G., et al. Fall detection using a head-worn barometer. in
powerful processor and low latency. However, since the
International Conference on Wireless Mobile Communication and
beginning of technology evolution, the challenge is the power Healthcare. 2016. Springer.
usage has always been the main concern of every system, [13] Lindemann, U., et al., Evaluation of a fall detector based on
especially for wearable devices. By integrating high resources accelerometers: A pilot study. Medical and Biological engineering and
computing, 2005. 43(5): p. 548-551.
algorithms, into battery-powered compact device with limited
[14] Martinez-Villaseñor, L. and H. Ponce, Design and analysis for fall
memory, run-time for the device might be reduced thus ruining detection system simplification. JoVE (Journal of Visualized
the ideal criteria for a fall detection device working system. Experiments), 2020(158): p. e60361.
Since FDS performs detection using the algorithm applied in [15] Yao, L., W. Min, and K. Lu, A new approach to fall detection based on
the human torso motion model. Applied Sciences, 2017. 7(10): p. 993.
the system, the algorithm model produced was set according to
[16] Shi, G., et al. Fall detection system based on inertial mems sensors:
the dataset trained. Therefore, since datasets are limited to the Analysis design and realization. in 2015 IEEE International Conference
current amount of data collected, algorithms would require a on Cyber Technology in Automation, Control, and Intelligent Systems
regular update of data to par with the evolution of thousands of (CYBER). 2015. IEEE.
[17] Albert, M.V., et al., Fall classification by machine learning using mobile
fall actions made by humans. Another challenge is to update
phones. PloS one, 2012. 7(5): p. e36556.
and revise the existing programming and concurrently support [18] Kraft, D., K. Srinivasan, and G. Bieber, Deep learning based fall
big data, which is the way for the system to detect highly detection algorithms for embedded systems, smartwatches, and IoT
accurate fall detection. devices using accelerometers. Technologies, 2020. 8(4): p. 72.
[19] Martínez-Villaseñor, L., et al., UP-fall detection dataset: A multimodal
Existing fall detection datasets are simulated datasets.
approach. Sensors, 2019. 19(9): p. 1988.
Essentially, a fall is an occasional dangerous human action. A [20] Montanini, L., et al., A footwear-based methodology for fall detection.
real fall is a kind of unpredictable action. Thus, it is hard to IEEE Sensors Journal, 2017. 18(3): p. 1233-1242.
collect datasets associated with real data. Researchers find a [21] Mao, A., et al., Highly portable, sensor-based system for human fall
monitoring. Sensors, 2017. 17(9): p. 2096.
way to overcome this by selecting stuntmen or volunteers to
[22] Musci, M., et al., Online fall detection using recurrent neural networks.
wear fall detection devices for a long time to record all arXiv preprint arXiv:1804.04976, 2018.
activities. Nevertheless, most of the activities recorded are [23] Núñez-Marcos, A., G. Azkune, and I. Arganda-Carreras, Vision-based
ADLs, and they would only be fortunate to collect any fall data fall detection with convolutional neural networks. Wireless
communications and mobile computing, 2017. 2017.
if only some of the volunteers fell while wearing the device.
[24] Mastorakis, G. and D. Makris, Fall detection system using Kinect’s
Since it is classified as real fall data, the occurrence of this infrared sensor. Journal of Real-Time Image Processing, 2014. 9(4): p.
situation is less than 20% of the time. 635-646.
It's also worth noting that there is not an integrated system [25] Cai, Z., et al., RGB-D datasets using microsoft kinect or similar sensors:
a survey. Multimedia Tools and Applications, 2017. 76(3): p. 4313-
that considers fall risk when detecting falls and issuing alerts
4355.
and camera-assisted observations. Furthermore, while existing [26] Wang, X., J. Ellul, and G. Azzopardi, Elderly fall detection systems: A
systems primarily focus on fall detection, there is potential for literature survey. Front. Robot. AI, 2020. 7: p. 71.
developing an FDS that incorporates closed-loop feedback— [27] Khan, M.S., et al., An unsupervised acoustic fall detection system using
source separation for sound interference suppression. Signal
that is, one that learns from a subject's fall patterns/history and
processing, 2015. 110: p. 199-210.
changes in physiological parameters at the time of fall and [28] Daher, M., et al., Elder tracking and fall detection system using smart
trains the fall detection algorithm based on these factors, tiles. IEEE Sensors Journal, 2016. 17(2): p. 469-479.
allowing for accurate profiling. This, eventually, enabled the [29] Hussain, F., et al., Activity-aware fall detection and recognition based
on wearable sensors. IEEE Sensors Journal, 2019. 19(12): p. 4528-4536.
system to issue alerts before the fall happened before the input
[30] Ngu, A.H., et al. Personalized Fall Detection System. in 2020 IEEE
of the pre-fall pattern. International Conference on Pervasive Computing and Communications
Workshops (PerCom Workshops). 2020. IEEE.
[31] Pannurat, N., S. Thiemjarus, and E. Nantajeewarawat, A hybrid temporal
REFERENCES reasoning framework for fall monitoring. IEEE Sensors Journal, 2017.
[1] Luna-Perejón, F., M.J. Domínguez-Morales, and A. Civit-Balcells, 17(6): p. 1749-1759.
Wearable fall detector using recurrent neural networks. Sensors, 2019. [32] Sucerquia, A., J.D. Lopez, and J.F. Vargas-Bonilla, Real-Life/Real-Time
19(22): p. 4885. Elderly Fall Detection with a Triaxial Accelerometer. Sensors (Basel),
[2] United Nations Department of Economic and Social Affairs, P.D., World 2018. 18(4).
Population Ageing 2020 Highlights: Living arrangements of older [33] de Quadros, T., A.E. Lazzaretti, and F.K. Schneider, A movement
persons. 2020. decomposition and machine learning-based fall detection system using
[3] Jahanjoo, A., M. Naderan, and M.J. Rashti, Detection and multi-class wrist wearable device. IEEE Sensors Journal, 2018. 18(12): p. 5082-
classification of falling in elderly people by deep belief network 5089.
algorithms. Journal of Ambient Intelligence and Humanized Computing, [34] Sun, Y., et al. Wearable fall detection system using barometric pressure
2020: p. 1-21. sensors and machine learning. in SENSORCOMM 2019: The Thirteenth
[4] Taramasco, C., et al., A novel monitoring system for fall detection in International Conference on Sensor Technologies and Applications.
older people. Ieee Access, 2018. 6: p. 43563-43574. 2019.
[5] Organization, W.H., WHO global report on falls prevention in older age. [35] Nooruddin, S., et al., Sensor-based fall detection systems: a review.
2008, World Health Organization. Journal of Ambient Intelligence and Humanized Computing, 2022: p. 1-
[6] Amir, N.I.M., et al. Real-time Threshold-Based Fall Detection System 17.
Using Wearable IoT. in 2022 4th International Conference on Smart [36] Shahiduzzaman, M., Fall detection by accelerometer and heart rate
Sensors and Application (ICSSA). 2022. IEEE. variability measurement. Global Journal of Computer Science and
Technology, 2016.
[37] Ramachandran, A., et al., Machine learning algorithms for fall detection [61] Shahzad, A. and K. Kim, FallDroid: An automated smart-phone-based
using kinematic and heart rate parameters-a comprehensive analysis. fall detection system using multiple kernel learning. IEEE Transactions
IAES International Journal of Artificial Intelligence, 2020. 9(4): p. 772. on Industrial Informatics, 2018. 15(1): p. 35-44.
[38] Butt, F.S., et al., Fall detection from electrocardiogram (ecg) signals [62] Liu, Z., et al., A benchmark database and baseline evaluation for fall
and classification by deep transfer learning. Information, 2021. 12(2): detection based on wearable sensors for the Internet of medical things
p. 63. platform. IEEE Access, 2018. 6: p. 51286-51296.
[39] Nadeem, A., A. Mehmood, and K. Rizwan, A dataset build using [63] Yacchirema, D., et al., Fall detection system for elderly people using IoT
wearable inertial measurement and ECG sensors for activity and big data. Procedia computer science, 2018. 130: p. 603-610.
recognition, fall detection and basic heart anomaly detection system. [64] Chelli, A. and M. Pätzold, A machine learning approach for fall
Data in brief, 2019. 27: p. 104717. detection and daily living activity recognition. IEEE Access, 2019. 7: p.
[40] Bernadus, T.F., L.B. Subekti, and Y. Bandung. IoT-based fall detection 38670-38687.
and heart rate monitoring system for elderly care. in 2019 International [65] LeCun, Y., Y. Bengio, and G. Hinton, Deep learning. nature, 2015.
Conference on ICT for Smart Society (ICISS). 2019. IEEE. 521(7553): p. 436-444.
[41] Nyan, M., F.E. Tay, and E. Murugasu, A wearable system for pre-impact [66] Kim, T.H., et al., Machine learning-based pre-impact fall detection
fall detection. Journal of biomechanics, 2008. 41(16): p. 3475-3481. model to discriminate various types of fall. Journal of biomechanical
[42] Hossain, F., et al. A direction-sensitive fall detection system using single engineering, 2019. 141(8).
3D accelerometer and learning classifier. in 2016 International [67] Fakhrulddin, A.H., X. Fei, and H. Li. Convolutional neural networks
Conference on Medical Engineering, Health Informatics and (CNN) based human fall detection on body sensor networks (BSN)
Technology (MediTec). 2016. IEEE. sensor data. in 2017 4th international conference on systems and
[43] Burns, A., et al., SHIMMER™–A wireless sensor platform for informatics (ICSAI). 2017. IEEE.
noninvasive biomedical research. IEEE Sensors Journal, 2010. 10(9): p. [68] Santos, G.L., et al., Accelerometer-based human fall detection using
1527-1534. convolutional neural networks. Sensors, 2019. 19(7): p. 1644.
[44] Zhao, S., et al. Recognition of human fall events based on single tri-axial [69] Xu, J., Z. He, and Y. Zhang. CNN-LSTM combined network for IoT
gyroscope. in 2018 IEEE 15th International conference on networking, enabled fall detection applications. in Journal of Physics: Conference
sensing and control (ICNSC). 2018. IEEE. Series. 2019. IOP Publishing.
[45] Luna-Perejon, F., et al. An automated fall detection system using [70] Nait Aicha, A., et al., Deep learning to predict falls in older adults based
recurrent neural networks. in Conference on Artificial Intelligence in on daily-life trunk accelerometry. Sensors, 2018. 18(5): p. 1654.
Medicine in Europe. 2019. Springer. [71] Sucerquia, A., J.D. López, and J.F. Vargas-Bonilla, SisFall: A fall and
[46] Musci, M. and M. Piastra, Recurrent Neural Networks Architectures for movement dataset. Sensors, 2017. 17(1): p. 198.
Accidental Fall Detection on Wearable Embedded Devices. Deep [72] Anceschi, E., et al., SaveMeNow. AI: a Machine Learning based
Learning for Biomedical Data Analysis: Techniques, Approaches, and wearable device for fall detection in a workplace, in Enabling AI
Applications, 2021: p. 81-98. Applications in Data Science. 2021, Springer. p. 493-514.
[47] Torti, E., et al., Embedding recurrent neural networks in wearable [73] Chevalier, G., Lstms for human activity recognition. 2016.
systems for real-time fall detection. Microprocessors and Microsystems, [74] Torti, E., et al. An Hardware Recurrent Neural Network for Wearable
2019. 71: p. 102895. Devices. in 2020 23rd Euromicro Conference on Digital System Design
[48] Musci, M., et al., Online Fall Detection using Recurrent Neural (DSD). 2020. IEEE.
Networks on Smart Wearable Devices. IEEE Transactions on Emerging [75] Ordóñez, F.J. and D. Roggen, Deep convolutional and lstm recurrent
Topics in Computing, 2020. neural networks for multimodal wearable activity recognition. Sensors,
[49] Wang, H., et al., RT-Fall: A real-time and contactless fall detection 2016. 16(1): p. 115.
system with commodity WiFi devices. IEEE Transactions on Mobile [76] Jeong, S.S., N.H. Kim, and Y.S. Yu, Fall Detection System Based on
Computing, 2016. 16(2): p. 511-526. Simple Threshold Method and Long Short-Term Memory: Comparison
[50] Hsieh, S.-L., et al. A wrist-worn fall detection system using with Hidden Markov Model and Extraction of Optimal Parameters.
accelerometers and gyroscopes. in Proceedings of the 11th IEEE Applied Sciences, 2022. 12(21): p. 11031.
International Conference on Networking, Sensing and Control. 2014. [77] Medsker, L.R. and L. Jain, Recurrent neural networks. Design and
IEEE. Applications, 2001. 5: p. 64-67.
[51] Zurbuchen, N., A. Wilde, and P. Bruegger, A machine learning multi- [78] Goodfellow, I., et al., Deep learning. Vol. 1. 2016: MIT press
class approach for fall detection systems based on wearable sensors Cambridge.
with a study on sampling rates selection. Sensors, 2021. 21(3): p. 938. [79] Mauldin, T.R., et al., SmartFall: A smartwatch-based fall detection
[52] Wu, F., et al., Development of a wearable-sensor-based fall detection system using deep learning. Sensors, 2018. 18(10): p. 3363.
system. International journal of telemedicine and applications, 2015. [80] Theodoridis, T., et al. Human fall detection from acceleration
2015. measurements using a recurrent neural network. in International
[53] Krupitzer, C., et al., Beyond position-awareness—Extending a self- Conference on Biomedical and Health Informatics. 2017. Springer.
adaptive fall detection system. Pervasive and Mobile Computing, 2019. [81] Kwolek, B. and M. Kepski, Human fall detection on embedded platform
58: p. 101026. using depth maps and wireless accelerometer. Computer methods and
[54] Aziz, O., et al., A comparison of accuracy of fall detection algorithms programs in biomedicine, 2014. 117(3): p. 489-501.
(threshold-based vs. machine learning) using waist-mounted tri-axial [82] Bourke, A.K., J. O’brien, and G.M. Lyons, Evaluation of a threshold-
accelerometer signals from a comprehensive set of falls and non-fall based tri-axial accelerometer fall detection algorithm. Gait & posture,
trials. Medical & biological engineering & computing, 2017. 55(1): p. 2007. 26(2): p. 194-199.
45-55. [83] Wang, G., et al., Elderly fall detection with an accelerometer using
[55] Singh, K., A. Rajput, and S. Sharma, Human fall detection using lightweight neural networks. Electronics, 2019. 8(11): p. 1354.
machine learning methods: A survey. International Journal of [84] Ding, J. and Y. Wang, A WiFi-based smart home fall detection system
Mathematical, Engineering and Management Sciences, 2020. 5(1): p. using recurrent neural network. IEEE Transactions on Consumer
161-180. Electronics, 2020. 66(4): p. 308-317.
[56] Saleh, M. and R.L.B. Jeannès, Elderly fall detection using wearable [85] Farsi, M., Application of ensemble RNN deep neural network to the fall
sensors: A low cost highly accurate algorithm. IEEE Sensors Journal, detection through IoT environment. Alexandria Engineering Journal,
2019. 19(8): p. 3156-3164. 2021. 60(1): p. 199-211.
[57] Vallabh, P., et al. Fall detection using machine learning algorithms. in [86] Luna Perejón, F., et al. Sampling frequency evaluation on recurrent
2016 24th international conference on software, telecommunications neural networks architectures for iot real-time fall detection devices. in
and computer networks (SoftCOM). 2016. IEEE. IJCCI 2019: 11th International Joint Conference on Computational
[58] Zurbuchen, N., P. Bruegger, and A. Wilde. A Comparison of Machine Intelligence (2019), pp. 536-541. 2019. ScitePress Digital Library.
Learning Algorithms for Fall Detection using Wearable Sensors. in [87] Han, H.R., Hybrid Fiber Materials according to the Manufacturing
2020 International Conference on Artificial Intelligence in Information Technology Methods and IOT Materials: A Systematic Review.
and Communication (ICAIIC). 2020. IEEE. Materials, 2023. 16(4): p. 1351.
[59] Islam, M.M., et al., Deep Learning Based Systems Developed for Fall [88] IEC 63203-801-1:2022, in Wearable electronic devices and
Detection: A Review. IEEE Access, 2020. 8: p. 166117-166137. technologies - Part 801-1: Smart body area network (SmartBAN) -
[60] Ren, L. and Y. Peng, Research of fall detection and fall prevention Enhanced ultra-low power physical layer. 2022, International
technologies: A systematic review. IEEE Access, 2019. 7: p. 77702- Electrotechnical Commission: IEC Webstore International
77722. Electrotechnical Commission. p. 28.
[89] IEC 63203-801-2:2022, in Wearable electronic devices and
technologies - Part 801-2: Smart body area network (SmartBAN) - Low
 complexity medium access control (MAC) for SmartBAN. 2022,
International Electrotechnical Commission: IEC Webstore International
Electrotechnical Commission. p. 77.
[90] IEC 60601-1:2023 SER Series. 2023, International Electrotechnical
Commission: IEC Webstore International Electrotechnical Commission.
p. 3763.
[91] Galvão, Y.M., et al., A multimodal approach using deep learning for fall
detection. Expert Systems with Applications, 2021. 168: p. 114226.
[92] Ntanasis, P., et al. Investigation of sensor placement for accurate fall
detection. in International Conference on Wireless Mobile
Communication and Healthcare. 2016. Springer.
[93] Özdemir, A.T., An analysis on sensor locations of the human body for
wearable fall detection devices: Principles and practice. Sensors, 2016.
16(8): p. 1161.
[94] Kangas, M., et al., Comparison of low-complexity fall detection
algorithms for body attached accelerometers. Gait & posture, 2008.
28(2): p. 285-291.
[95] Lipton, Z.C., J. Berkowitz, and C. Elkan, A critical review of recurrent
neural networks for sequence learning. arXiv preprint
arXiv:1506.00019, 2015.
[96] Kangas, M., et al. Determination of simple thresholds for accelerometry-
based parameters for fall detection. in 2007 29th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society.
2007. IEEE.
[97] Khojasteh, S.B., et al. Fall detection analysis using a real fall dataset.
in The 13th International Conference on Soft Computing Models in
Industrial and Environmental Applications. 2018. Springer.
[98] Bianchi, F., et al., Barometric pressure and triaxial accelerometry-based
falls event detection. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 2010. 18(6): p. 619-627.
[99] Zhang, Q. and S. Zhu. Real-time activity and fall risk detection for aging
population using deep learning. in 2018 9th IEEE Annual Ubiquitous
Computing, Electronics & Mobile Communication Conference
(UEMCON). 2018. IEEE.
[100] Lu, N., et al., Deep learning for fall detection: Three-dimensional CNN
combined with LSTM on video kinematic data. IEEE journal of
biomedical and health informatics, 2018. 23(1): p. 314-323.
[101] Tasoulis, S.K., et al. Deep learning and change detection for fall
recognition. in International Conference on Engineering Applications of
Neural Networks. 2019. Springer.
[102] Casilari, E., R. Lora-Rivera, and F. García-Lagos, A study on the
application of convolutional neural networks to fall detection evaluated
with multiple public datasets. Sensors, 2020. 20(5): p. 1466.
[103] Vavoulas, G., et al., The mobifall dataset: Fall detection and
classification with a smartphone. International Journal of Monitoring
and Surveillance Technologies Research (IJMSTR), 2014. 2(1): p. 44-
56.
[104] Medrano, C., et al., Detecting falls as novelties in acceleration patterns
acquired with smartphones. PloS one, 2014. 9(4): p. e94811.
[105] Frank, K., et al. Bayesian recognition of motion related activities with
inertial sensors. in Proceedings of the 12th ACM international
conference adjunct papers on Ubiquitous computing-Adjunct. 2010.
[106] Vilarinho, T., et al. A combined smartphone and smartwatch fall
detection system. in 2015 IEEE International Conference on Computer
and Information Technology; Ubiquitous Computing and
Communications; Dependable, Autonomic and Secure Computing;
Pervasive Intelligence and Computing. 2015. IEEE.
[107] Nizam, Y. and M.M.A. Jamil, Classification of Daily Life Activities for
Human Fall Detection: A Systematic Review of the Techniques and
Approaches, in Challenges and Trends in Multimodal Fall Detection for
Healthcare. 2020, Springer. p. 137-179.
[108] Qian, X., et al., Wearable computing with distributed deep learning
hierarchy: a study of fall detection. IEEE Sensors Journal, 2020. 20(16):
p. 9408-9416.
[109] Krupitzer, C., et al. Hips do lie! A position-aware mobile fall detection
system. in 2018 IEEE International Conference on Pervasive Computing
and Communications (PerCom). 2018. IEEE.
[110] Redding, M. Best Medical Alert Systems With Fall Detection for Seniors.
2023 [cited 2023 25 January]; Available from:
https://www.seniorliving.org/medical-alert-systems/best/fall-detection/.