Fusion of Machine Learning and Threshold-Based Approaches

for Fall Detection in Healthcare Using Inertial Sensors

Ya Wang1 a
, Peiman Alipour Sarvari2 b
and Djamel Khadraoui2 c
1 Faculty of Science, Technology and Medicine, University of Luxembourg, Esch sur Alzette, Luxembourg
2 IT for Innovative Services, Luxembourg University of Science and Technology, Esch sur Alzette, Luxembourg

Keywords: Wearable Fall Detection, Feature Extraction, Threshold, Machine Learning, Inertial Sensors.

Abstract: In the healthcare sector, specifically for elderly care, accurate and efficient fall detection is crucial. We present
an advanced fall detection methodology tailored for wearable systems. Our approach blends threshold-based
screening with machine learning models like Support Vector Machine, K-Nearest Neighbors, Decision Tree,
Random Forest, and XGBoost. Utilizing 65 features extracted from the gyroscope and accelerometer data from
Inertial Measurement Units, our method addresses the class imbalance often found between Activities of Daily
Living and actual fall events. Threshold-based pre-screening serves to mitigate the class imbalance of the fall
dataset, making the subsequent machine-learning classification more effective. Validation on two open-source
IMU datasets, Sisfall and FallAllD, achieving high accuracy rates of 99.55%, 99.68% (wrist), 99.76% (waist),
and 99.52% (neck), shows our model surpassing existing solutions in detection accuracy. Furthermore, our
strategic feature extraction not only enhances the model’s performance but also allows for a fourfold reduction
by using the 15 most important features in data transmission without sacrificing accuracy. These findings
underscore the efficiency and potential of our methodology, indicating that wearables can indeed be powerful
tools for high-precision fall detection with minimal data overhead.

1 INTRODUCTION have seen a significant uptick, underscoring the field’s

growing importance (Tanwar et al., 2022).
According to the World Health Organization, falls ac- Fall detection systems (FDSs) primarily fall into
count for approximately 600,000 global deaths each two categories: Context-Aware Systems (CAS) and
year, ranking second among unintentional injury- wearable FDSs. CAS systems utilize sensors like mi-
related deaths (WHO, 2023). Alarmingly, 75% of crophones, cameras, and radars placed in a prede-
these fatalities occur in adults over the age of 65 fined area surrounding the individual. However, the
(Vaishya and Vaish, 2020). With the global popula- need for customization, high installation and mainte-
tion aging at an unprecedented rate (WHO, 2022), im- nance costs, and limited coverage areas restrict their
mediate assistance following falls is vital to minimize applicability outside controlled environments such as
medical complications. In fact, prolonged periods of nursing homes. Contrarily, wearable FDSs use in-
immobility after a fall, often lasting over an hour, have ertial measurement units (IMUs) directly attached to
been shown to increase the risk of mortality and lead the individual, allowing for location-independent fall
to severe health issues such as dehydration and pneu- detection. These systems offer numerous advantages
monia (Fleming and Brayne, 2008). including cost-efficiency, easier installation, privacy
Given these concerns, there’s been a surge in inter- preservation, and simpler design and configuration
est in cost-effective Fall Detection Systems (FDSs). (Hashim et al., 2020). This feature, combined with
Telecare and remote biosignal monitoring offer in- the widespread use of smartwatches and sports bands,
novative pathways for these systems. Since 2010, enhances the feasibility and accessibility of wearable
both research articles and patents in automatic FDSs FDSs, making them particularly suitable for urban ar-
eas with reliable mobile connectivity.”
a https://orcid.org/0000-0002-4542-1074
b https://orcid.org/0000-0003-1235-2102
c https://orcid.org/0000-0003-1054-1612

573
Wang, Y., Sarvari, P. and Khadraoui, D.
Fusion of Machine Learning and Threshold-Based Approaches for Fall Detection in Healthcare Using Inertial Sensors.
DOI: 10.5220/0012250500003657
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 1, pages 573-582
ISBN: 978-989-758-688-0; ISSN: 2184-4305
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

2 RELATED WORK fall-like activities, adding a layer of complexity to the

detection problem (Martinez-Villaseñor and Ponce,
Algorithms of Wearable FDS can be broadly clas- 2020). Despite their promising results, machine
sified into three types: (i) threshold-based approach learning-based approaches do face a bottleneck in
(ii) machine learning-based approach and (iii) deep feature selection. The process of identifying the most
learning-based approach. Table 1 summarizes rep- relevant features for fall detection is non-trivial and
resentative studies of the above three categories of can affect the algorithm’s overall performance.
wearable sensor-based fall detection.
Deep Learning-Based Approach. The advent of
Threshold-Based Approach. Threshold-based powerful computational hardware has propelled the
methods hinge on predefined threshold values to de- utilization of deep learning algorithms in fall detec-
tect falls by comparing specific sensor data—namely, tion (Yu et al., 2020). These algorithms automati-
acceleration (Saadeh et al., 2019), angular velocity cally identify important features from raw sensor data,
(Bourke and Lyons, 2008), and body angle (Sorvala eliminating the need for manual feature engineering.
et al., 2012). It operates on the premise that falls Remarkable performance metrics, such as a sensitiv-
exhibit distinct differences in body position and ve- ity of 99.3% and specificity of 91.86% using ResNet
locity compared to activities of daily living (ADLs). architecture, have been reported (Zhang et al., 2021).
The system’s effectiveness largely depends on the Moreover, the ConvLSTM model was shown to excel
accuracy of these preset values. For example, de in both sensitivity and specificity, achieving 99.32%
Sousa et al. (de Sousa et al., 2021) introduced a low- and 99.01% (Yu et al., 2022).
power wearable system for fall detection that utilized To address this challenge of fall detection and
a threshold-based approach and achieved a sensitivity enhance computational efficiency, we introduce a
of 92.6% and specificity of 97.7%. Jung et al. (Jung two-tiered hybrid algorithm that integrates threshold-
et al., 2020) employed thresholds based on the sum based and machine-learning methods for wearable
vector magnitude of acceleration, the sum vector Fall detection systems.
magnitude of angular velocity, and the vertical angle.
They reported a sensitivity of 100% and specificity
of 97.54%. However, the approach has limitations 3 METHODOLOGY
in terms of generalizing across different settings
and populations, causing decreased performance, Fall datasets often exhibit a class imbalance between
particularly in specificity, when tested on complex Activities of Daily Living (ADL) and genuine fall
datasets like SisFall (Wang et al., 2020)(Sucerquia events. This imbalance skews the performance of
et al., 2017). models trained on such datasets, particularly affect-
ing their ability to accurately identify falls, which
Machine Learning-Based Approach. Machine are the minority class. To address this challenge and
learning techniques offer a flexible and adaptive to enhance computational efficiency, we introduce a
alternative to threshold-based fall detection methods, two-tiered hybrid algorithm that integrates threshold-
often yielding improved performance in diverse based and machine-learning methods. The flow chart
scenarios. For example, Giuffrida et al. used a in Figure 1 illustrates the overall process of the fusion
Support Vector Machine (SVM) model trained on a approach.
curated set of features, which significantly optimized
the system’s parameters (Giuffrida et al., 2019). In 3.1 Data Processing
a similar vein, Yu et al. adopted a Hidden Markov
Model (HMM) for fall detection that circumvented 3.1.1 Dataset
the need for manual feature selection altogether.
Their approach processed raw acceleration data and In this study, the open source datasets, SisFall and
achieved an impressive sensitivity of 99.2% and FallAllD were utilized to validate the effectiveness of
specificity of 99.0% (Yu et al., 2017). To assess three different fall detection approaches, after analyz-
the general efficacy of machine learning in this ing the various fall detection datasets.
domain, Martinez-Villaseor et al. compared four
key machine learning classifiers: Random Forest
SisFall Dataset. SisFall (Sucerquia et al., 2017)
(RF), SVM, Multilayer Perceptron (MLP), and
collected data from the IMU attached to the waist.
k-Nearest Neighbors (KNN). These algorithms were
The IMU includes sensors such as accelerometers,
evaluated for their ability to differentiate falls from

574
 Fusion of Machine Learning and Threshold-Based Approaches for Fall Detection in Healthcare Using Inertial Sensors

Low Pass
A x ,A y ,A z
Raw Data ? x,? y,? z filter 10Hz

window around Fall

peak value Data
Data
Groundtruth
Processing
Sliding window ADL
with fixed length Data

Figure 2: Body accelerator and angular velocity measure-

ment system.
SMVmax>CT1
NO ADL
SMAmax>CT2
3.1.2 Low-Pass Filtering
Yes
The raw data collected from the Inertial Measurement
Feature
Unit (IMU) may contain electronic noise or other
Extraction types of artifacts that can affect the accuracy and reli-
ability of the measurements. To mitigate these distur-
Normalization
bances, a filtering process is often applied to the data
Feature Matrix (Yu et al., 2022)(Jung et al., 2020) (Shi et al., 2020).
In this research, we utilized a fourth-order low-pass
Butterworth filter with a 2-pass digital implementa-
ML-based Feature Machine Learning
Classifier Selection Models tion to remove noise and artifacts from the accelera-
tor data and angular velocity data. The chosen cut-off
frequency for the filter was set at 10Hz. This value
Fall/ADLs was selected because the relevant frequency spectra
information
of human motion typically fall within the range of 0
Figure 1: Flow chart of the designed fall detection ap- to 10Hz (Winter, 2009).
proach.
3.1.3 Data Refinement and Segmentation
gyroscopes, and magnetometers. It consists of 19
ADLs and 15 fall types performed by 23 young adults, The six-axis inertial signals from the IMU sensor
15 ADL types performed by 14 healthy and indepen- were defined as Ax ,Ay ,Az ,ωx ,ωy and ωz . During a fall
dent participants over 62 years old, and data from one event, as the body makes contact with the ground, it
participant of 60 years old that performed all ADLs typically causes sudden and pronounced peaks in ac-
and falls. The dataset includes motion data captured celeration. Figure 3 illustrates a representative dia-
by an inertial measurement unit (IMU) placed on the gram displaying the 3-axis acceleration and angular
waist at a sampling rate of 200Hz. The SisFall dataset velocity data of a forward fall event starting from the
provides a substantial number of fall and ADL trials, initial state position. Such peaks, captured by the Sig-
making it suitable for evaluating fall detection algo- nal Magnitude Vector (SMV), serve as the primary
rithms. indicators for fall events. Mathematically, for each
measurement during a particular fall instance, SMV
is defined as:
FallAllD Dataset. The FallAllD dataset, proposed q
by (Saleh et al., 2020), utilizes an IMU placed on the SMVi = Ax 2i + Ay 2i + Az 2i (1)
neck, chest, and waist to measure movement during
experiments. The dataset was obtained from 15 sub-
jects (eight males and seven females), defined as con- Segmentation Strategy Using Signal Magnitude
taining 35 falls and 44 ADL types. The waist and Vector (SMV):
neck acquired ADL and fall data for 14 and 12 sub- 1. Fall Events: Fall instances are pinpointed by
jects, respectively, whereas the wrist sensor acquired observing the 2-second window surrounding the
ADL and fall data for 13 and 9 subjects. peak SMV values within the recorded data.
Figure 2 illustrates the coordinates of the ac- Ground truth information is used to validate and
celerometer and the angular velocity measurements in label these samples as falls. The definition of the
our research. i
time of peak SMV (tSMV max ) for the i th fall event
is defined as:

575
BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

phases of a fall, including the free-fall and impact pe-

riods, happen within an even narrower range of 0.5-
0.85 seconds (Huynh et al., 2013). Eduardo Casilari
et al. fine-tuned this by proposing a 2-second window,
centered around the peak of the fall signal, capturing
the most relevant features of a fall event (Casilari and
Silva, 2022). In our analysis, we adopt this 2-second
observation window for optimal fall detection.

3.1.5 Data Split

After the data processing, the whole data set is split

into the training dataset and test dataset with a ratio
of 75/25. The flow chart in Figure 4 illustrates the
training process and testing process.
Figure 3: Fall accident from a static posture. (tW : length of
the time window for the fall segment).
A x ,A y ,A z Data
Raw Data ? x,? y,? z Processing

i
tSMV max = arg max{SMVi (t) : SMVi (t) ∈ ith fall }
t
Data Split

(2)
2. ADL Events: To diversify the ADLs dataset, a Training data Testing data

sliding window method is employed. This mecha-

nism uses a consistent window of 2 seconds, pro- CT1
Threshold Thrshold
gressing at steps of 0.2 seconds across the data based SMVmax >CT1
`
SMA max >CT2
based
CT2
timestamps. This method captures data fragments
and categorizes them as non-fall events, ensur-
Trained
ing a more encompassing representation of typical ML-based
Classifier
Feature
Extraction
Machine Learning
Model Training
ML-based
Classifier
Model
movements.
However, there are instances where high-intensity ac- Fall/ADLs
information
tivities, such as rapid walking or jumping, can pro-
duce acceleration patterns similar to falls, posing Figure 4: Flow chart of the train process and test process.
challenges in accurate detection. To address this co-
nundrum, gyroscope data is incorporated to assess
and determine the subject’s posture. For every data 3.2 Threshold-Based Methods
point during the i-th fall event, the Signal Magni-
tude Vector of Angular Velocity (SMA) is leveraged The first tier acts as an initial filter using threshold-
to measure alterations in angular velocity. This metric based criteria to swiftly differentiate potential fall
is articulated as: events from routine activities. This approach allows
q for rapid processing, screening out most ADL in-
SMAi = ωx 2i + ωy 2i + ωz 2i (3) stances and forwarding only suspected fall events to
the second tier for detailed analysis. Specifically,
3.1.4 Adaptive Time-Window Size the algorithm employs two thresholds calculated from
sensor measurements of the training data: Signal
Fall detection algorithms commonly utilize temporal Magnitude Vector (SMV) and Signal Magnitude Vec-
windows to analyze inertial signals where a fall event tor of Angular Velocity (SMA). These thresholds are
might occur. These windows typically span between defined as follows:
0.2 and 2 seconds. Optimal window durations for fall
and human activity detection have been subject to re- CT 1 = min{SMVmax : SMVmax ∈ fall training data}
search. Banos et al. recommended a 1-2 second win- (4)
dow for general human activities, balancing recogni- where SMVmax = max{SMV (t) : t ∈ tW }
tion speed with accuracy (Banos et al., 2014). The CT 2 = min{SMAmax : SMAmax ∈ fall training data}
intrinsic dynamics of a fall, characterized by abrupt (5)
and unexpected movements, usually occur within a where SMAmax = max{SMA(t) : t ∈ tW }
1-3 second timeframe (Yu, 2008). The most critical

576
 Fusion of Machine Learning and Threshold-Based Approaches for Fall Detection in Healthcare Using Inertial Sensors

Within the predefined time window tW , if both changes in the forward and sideward attitudes of in-
SMVmax and SMAmax exceed their respective thresh- dividuals during a fall.
olds, the data is forwarded to the second tier; other- |Ax |
wise, it is disregarded. Roll = arccos q (8)
Ax 2 + Ay 2
3.3 Machine Learning-Based Approach The attitude angle change during the fall is defined
as:
The second tier capitalizes on machine learning algo-
NAng = Pitch + Roll (9)
rithms to meticulously classify the dataset that’s been
pre-screened by the initial tier. This dual-stage ap- The selected statistic features are analytically de-
proach amplifies the algorithm’s precision in detect- fined as follows:
ing falls, while also boosting computational speed.
Maximum (Peak) of Feature S. This feature rep-
3.3.1 Feature Extraction resents the peak or maximum value of a specific
data feature (S) during the fall window. It serves as
Before applying machine learning classifiers, we fo- a meaningful descriptor of the force of the impact
cus on feature extraction to accurately represent the against the ground. Mathematically, it can be defined
underlying data patterns. We compute a set of eight as:
statistical features that encapsulate information from
accelerometer and gyroscope readings. These fea- Smax = max{S(t) : t ∈ tW } (10)
tures include metrics such as angular velocity, Signal
Magnitude Vector (SMV), and Signal Magnitude Vec-
Minimum of Feature S. This feature indicates the
tor of Angular Velocity (SMA). For a detailed sum-
minimum value achieved by the data feature (S) dur-
mary, refer to Table 1. These statistical features are
ing the fall window. It is a key element in describing
universally accepted in the domains of Human Ac-
the fall.
tivity Recognition (HAR) and Fall Detection Systems
Smin = min{S(t) : t ∈ tW } (11)
(FDS) (Sucerquia et al., 2017; Giuffrida et al., 2019;
Martinez-Villaseñor and Ponce, 2020; Casilari and
Silva, 2022). Mean of Feature S. The mean provides informa-
We denote the human acclivity feature derived tion about the average body motion intensity during
from the raw data of the IMU sensor by S, which is the fall. It is computed as the average of the feature
defined as values over the observation window (tW ) containing
NW feature samples.
S = [Ax , Ay , Az , ωx , ωy , ωz , SMV, SMA, NAng ], (6)
1
µS = ∑S (12)
where Ax , Ay , Az represent the accelerometer read- NW t∈tW
ings along the X, Y, and Z axes, respectively, and
ωx , ωy , ωz represent the angular velocity readings where NW is defined as:
along the X, Y, and Z axes, respectively, SMV and T
SMA denote the Signal Magnitude Vector and Signal Nw = 2[ fs ] + 1 (13)
2
Magnitude Vector of Angular Velocity, respectively,
NAng represent the attitude change during the fall. Standard Deviation of Feature S. This feature de-
The attitude angle change during the fall NAng is scribes the variability of the feature (S) during the ob-
defined as follows: servation window. It is calculated as the square root
of the average squared deviation from the mean.
Pitch Angle. The pitch angle represents the forward s
angle of the sensor during a fall. 1
σS = ∑ (S − µS )2 S
NW t∈t
(14)
W
|Az |
Pitch = arccos q (7)
Az 2 + Ay 2 Skewness of Feature S. Skewness characterizes the
symmetry of the distribution of feature values. It in-
Roll Angle. The roll angle represents the sideward dicates whether the distribution is skewed to the left
angle of the sensor during a fall. By incorporat- or right.
1
ing these pitch and roll angles, we can capture the γS = 3 ∑ (S − µS )3
σS NW t∈t
(15)
W

577
BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

Table 1: Descriptive statistics of 65 features.

Statistic Equation Acc Ang SMV SMA NAng
Axmax ωxmax
Maximum Smax = max{S(t) : t ∈ tW } Ay max ωy max SMVmax SMAmax NAng max
Az max ωz max
Axmin ωxmin
Minimum Smin = min{S(t) : t ∈ tW } Ay min ωy min SMVmin SMAmin NAng min
Az min ωz min
µA x µω x
1 µA y µω y
Mean µS = NW ∑ S µSMV µSMA µNAng
t∈tW
µA z µω z
σAx σωx
Standard
r
1
Deviation
σS = NW ∑ (S − µS )2 S σAy σωy σSMV σSMA σNAng
t∈tW σAz σωz
γA x γω x
1
Skewness γS = σS 3 NW ∑ (S − µS )3 γA y γω y γSMV γSMA γNAng
t∈tW
γA z γω z
Kurt Ax Kurt ωx
Kurtosis Kurt S = E[( (S−µS) 4
σS ) ]
Kurt Ay Kurt ωy Kurt SMV Kurt SMA Kurt NAng
Kurt Az Kurt ωz
RAx Rωx
valley-to-peak
RS = Smax − Smin RAy Rωy RSMV RSMA NAN
range
RAz Rωz
valley-to-peak
TS = tSmax − tSmin NAN NAN TSMV TSMA NAN
time
1 Ax , Ay , Az represent the accelerometer readings along the X, Y, and Z axes, respectively,
2 ωx , ωy , ωz represent the angular velocity readings along the X, Y, and Z axes,
3 SMV and SMA denote the Signal Magnitude Vector and Signal Magnitude Vector of Angular Velocity

Kurtosis of Feature S. Kurtosis measures the based on their ability to accurately distinguish be-
tailedness (frequency of outliers) of the distribution tween fall and non-fall events.
of feature values. Our dual-layer approach, incorporating these se-
(S − µS ) 4 lected features, is rigorously tested on the Sisfall and
Kurt S = E[( ) ] (16) FallAllD datasets. By applying the chosen statisti-
σS
cal features to both categories, we aim to critically
assess and validate the performance of our machine
Valley-to-Peak Range (Rs ). This feature represents learning-based fall detection model.
the value of the interval between the minimum (Sm in)
and maximum (Sm ax) of the feature S. It is calculated
as:
RS = Smax − Smin (17) 4 RESULTS AND DISCUSSION

Valley-to-Peak Time (Ts ). This feature indicates 4.1 Performance Evaluation Metrics
the duration of the interval between the minimum
Various methods have been developed to evaluate the
(Smin ) and maximum (Smax ) of the feature S.
performance of different classifiers. These methods
TS = tSmax − tSmin (18) rely on the outcomes obtained from the classifiers,
which are represented in the form of a confusion ma-
3.3.2 Model Implementation trix (Figure 10). The confusion matrix provides a vi-
sual representation of the classifier’s performance, in-
With the features extracted, we proceed to apply cluding true positives, true negatives, false positives,
machine learning models for classification. In this and false negatives.
study, we explore multiple classifiers including Sup- • True positive (TP): The ADL events have been
port Vector Machines (SVM), K-Nearest Neighbors correctly classified.
(KNN), Decision Trees, Random Forests, and XG-
Boost. These classifiers are trained and evaluated • True negative (TN): The fall events have been cor-
rectly detected.

578
 Fusion of Machine Learning and Threshold-Based Approaches for Fall Detection in Healthcare Using Inertial Sensors

• False positive (FP): Fall events that have not been FallAllD
𝜇𝜔𝑧
detected.
𝑇𝑆𝑀𝑉
• False negative (FN): A false alarm situation oc- 𝜎𝑆𝑀𝑉
curs. 𝜎𝐴𝑧
𝜇𝜔𝑥
One commonly used method to assess classifier 𝛾𝑆𝑀𝑉
performance is accuracy, which calculates the pro- 𝑅𝑆𝑀𝑉
portion of correctly classified samples overall. How- 𝛾𝑁𝐴𝑛𝑔
ever, accuracy has certain limitations, such as being 𝑅𝑆𝑀𝐴
susceptible to the influence of large abnormal data 𝑆𝑀𝑉𝑚𝑎𝑥
𝜇𝑁𝐴𝑛𝑔
and potentially misleading results in class-imbalanced
𝜇𝜔𝑦
training data scenarios. To address these limitations,
𝐴𝑍 𝑚𝑎𝑥
alternative evaluation methods are selected to evalu-
𝑅𝐴𝑧
ate class-imbalanced classifiers effectively. For fall 𝐾𝑢𝑟𝑡𝑆𝑀𝑉
detection tasks with imbalanced classes, sensitivity,
specificity, F-score, and receiver operating character-
istic (ROC) are commonly utilized to assess the clas- (a) Shap values for FallAllD.
sifier’s ability to differentiate falls from a large num- SisFall
ber of activities of daily living (ADL) events. 𝑇𝑆𝑀𝐴
𝑅𝑆𝑀𝑉
TP+TN
Accurancy = (19) 𝑆𝑀𝐴𝑚𝑖𝑛
T P + T N + FP + FN 𝑇𝑆𝑀𝑉
Sensitivity measures the proportion of correctly 𝑅𝐴𝑧
identified positive samples. Specificity, on the other 𝑅𝐴𝑥
𝑁𝐴𝑛𝑔
hand, measures the proportion of correctly identified 𝑚𝑖𝑛
𝐾𝑢𝑟𝑡𝑆𝑀𝑉
negative samples. 𝑅𝜔𝑥
TP 𝐾𝑢𝑟𝑡𝐴𝑧
Sensitivity = (20) 𝐾𝑢𝑟𝑡𝑆𝑀𝐴
T P + FN
𝐴𝑧 𝑚𝑖𝑛
TN 𝐾𝑢𝑟𝑡𝜔𝑥
Speci f icity = (21)
T N + FP 𝜔𝑦
𝑚𝑎𝑥
The F1 is a robust evaluation metric that balances 𝛾𝑆𝑀𝑉
sensitivity and specificity. In large-scale datasets, sen-
sitivity and specificity often have a trade-off relation-
ship, and the F-score provides a comprehensive mea- (b) Shap values for SisFall.
sure of their discrimination. The parameter β in the Figure 5: The Shap values of the top 15 features of the
F-score equation allows adjusting the weight between dataset (a) FallAllD; (b) SisFall.
sensitivity and specificity. Setting β to 0.5 assigns a
higher weight to specificity, which is crucial in fall consistently ranked among the top ten most important
detection as it reflects the detection of all fall signals features when both datasets were integrated, empha-
in the data. sizing their critical role in accurately detecting falls
Sensitivity × Speci f icity and Activities of Daily Living (ADLs). These find-
F1 = (1 + β2 ) 2 (22) ings are corroborated by Table 2, which offers an
β × (Sensitivity + Speci f icity) exhaustive evaluation of different machine learning
In summary, sensitivity, specificity, and F-score classifiers based on the feature sets. In which, fea-
provide comprehensive measures for evaluating the ture (1) utilizes all 65 extracted features and features
performance of fall detection classifiers, particu- (2) focuses on the 15 most important features as deter-
larly in scenarios with imbalanced class distributions. mined by feature integration. The model’s robustness
These metrics address the limitations of accuracy and was evident from its high accuracy rates across vary-
offer a more nuanced assessment of the classifier’s ing types of falls and ADLs. The results show that
ability to distinguish falls from ADL events. the model is robust to the input of different types of
falls/ADLs and achieved superior performance. The
4.2 Results and Discussion FallAllD dataset is collated from 3 different locations
with limited data size, it has been shown that the po-
Our analyses underscored the varying importance of sition of the sensors also plays an important role in
features across the two datasets—Sisfall and Fal- the fall detection models. And we can also see that
lAllD. Notably, features like TSMV , TSMA , and RSMV

579
BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

Table 2: Fall detection results for Sisfall and FallAllD: the Table 3: Comparison of results between the proposed and
unit is %, feature (1) utilizes all 65 extracted features, and previous research models, the unit is %.
feature (2) focuses on the 15 most important features. Algorithem Sensitivity Specificity Accuracy
Dataset Model Sensitivity Specificity Accuracy F1 Yu et al. ResNet10 97.91 72.89 96.22
SisFall1 SVM 99.47 99.35 99.45 98.22 (2020) SMOTE 99.17 89.98 97.54
KNN 98.06 98.91 98.96 94.27 Sisfall CDL-Fall 99.33 91.86 97.52
DT 98.88 99.38 99.28 98.45 Santoyo et al.
RF 99.44 98.75 99.54 99.03 (2022) CNN 85.97 96.79 NAN
XGB 99.62 98.81 99.55 99.14 FallAllD waist
SisFall2 SVM 98.67 99.03 98.62 96.43 Jeong et al.
KNN 97.61 98.63 98.05 86.44 (2023) LightGBM 91.04 96.38 94.86
DT 97.29 99.23 98.70 97.98 FallAllD wrist
RF 99.01 97.87 99.26 98.88 Proposed Fusion
99.62 98.81 99.55
XGB 98.79 98.33 99.16 98.92 Sisfall method
FallAllD1 SVM 99.36 1 99.84 99.68 Proposed Fusion
99.62 99.47 99.68
(wrist) KNN 98.71 99.58 99.37 98.71 FallAllD wrist method
DT 98.87 99.61 99.55 98.82 Proposed Fusion
99.34 98.75 99.76
FallAllD waist method
RF 98.89 99.88 98.95 99.24
Proposed Fusion
XGB 99.62 99.47 99.68 99.47 98.24 99.47 99.52
FallAllD neck method
FallAllD2 SVM 98.27 99.40 99.37 97.55
(Wrist) KNN 97.67 97.44 97.49 94.02 Note: In this comparison, while the same open-source datasets are
DT 98.88 99.38 99.18 98.9 used, the training and testing datasets for the fall detection model
RF 96.44 98.75 99.66 99.4 differ due to variations in data processing methods, like filter fre-
XGB 98.62 98.51 99.55 98.7 quency and sample window size. Hence, the results should be
FallAllD1 SVM 99.77 99.03 99.34 99.27 viewed as indicative rather than conclusive.
(waist) KNN 98.06 98.91 98.96 98.69
chine learning approach benchmarked against Sis-
DT 97.88 99.02 99.18 99.22
RF 99.34 98.75 99.76 99.07
fall and FallAllD datasets. Employing a suite of
XGB 99.32 98.31 99.35 99.23 65 rigorously selected statistical features (as shown
FallAllD2 SVM 98.63 97.51 96.74 98.58 in table1) extracted from inertial sensors, the study
(waist) KNN 96.85 95.81 97.39 97.41 leveraged tree-based ensemble models to achieve un-
DT 96.86 99.84 97.68 97.88 precedented accuracy rates: 99.55%, 99.68% (wrist),
RF 98.62 98.80 99.44 98.47 99.76% (waist), and 99.52% (neck) across the exam-
XGB 97.82 98.82 98.84 98.64
ined datasets. This level of performance substantially
FallAllD1 SVM 96.53 98.94 99.15 91.25
(neck) KNN 89.06 98.91 95.96 87.10
outperforms existing benchmarks documented in the
DT 98.28 99.38 99.12 98.32 scholarly literature.
RF 98.24 99.47 99.52 99.40 SHAP value analysis was instrumental in distill-
XGB 98.88 99.46 99.47 99.51 ing the feature set down to the top 15 most influen-
FallAllD2 SVM 95.81 99.03 99.35 89.55 tial features. Comparative analysis indicated that the
(neck) KNN 79.66 91.34 86.90 86.79 reduced feature set incurred a statistically insignifi-
DT 97.48 99.38 98.18 97.44
cant diminution in performance metrics—less than a
RF 97.44 98.75 98.66 97.96
XGB 98.42 98.51 98.55 98.73
1% deviation relative to the exhaustive feature set.
The hybrid model architecture, ingeniously combin-
the tree-based models show superior and robust per- ing threshold-based and machine learning algorithms,
formance in different datasets. facilitates minimal data transference from the wear-
In Table 3, we present a comparative analysis able device to the computational node while sustain-
of fall detection capabilities between our proposed ing high fidelity in fall detection outcomes.
model and existing models, all evaluated using the While the current study’s accomplishments are
same dataset. Our model’s robustness and enhanced manifold, it is imperative to acknowledge its limita-
performance are evident; it consistently identifies a tions. The absence of real fall data in the utilized
range of falls and Activities of Daily Living (ADLs) datasets denotes an opportunity for future work to fur-
with remarkable accuracy. ther validate the model’s performance under invalid
conditions. In light of the latter, future research en-
deavors will be directed toward the integration of this
5 CONCLUSIONS validated model architecture into wearable technol-
ogy platforms, emphasizing the necessity of feature
In this study, we proposed a wearable fall detec- selection optimization for real-time fall detection.
tion model that combines the threshold and ma-

580
 Fusion of Machine Learning and Threshold-Based Approaches for Fall Detection in Healthcare Using Inertial Sensors

ACKNOWLEDGEMENTS Martinez-Villaseñor, L. and Ponce, H. (2020). Design and

analysis for fall detection system simplification. JoVE
This work was adapted and extended from the mas- (Journal of Visualized Experiments), (158):e60361.
ter’s thesis titled ’Analysis and Comparison of Dif- Saadeh, W., Butt, S. A., and Altaf, M. A. B. (2019). A
patient-specific single sensor iot-based wearable fall
ferent Types of Algorithms for Fall Detection in Fall
prediction and detection system. In IEEE transactions
Alerting Systems’ completed at the University of on neural systems and rehabilitation engineering, vol-
Luxembourg, supported by the European Active and ume 27, pages 995–1003. IEEE.
Assisted Living 2021(AAL) Programme, the Luxem- Saleh, M., Abbas, M., and Le Jeannes, R. B. (2020). Fal-
bourg National Research Fund (FNR), and the Lux- lalld: An open dataset of human falls and activities
embourg Institute of Science and Technology (LIST). of daily living for classical and deep learning applica-
This research is part of the AGAPE project, with the tions. IEEE Sensors Journal, 21(2):1849–1858.
code AAL-2021-8-124-CP, and titled ’ADVANCING Shi, J., Chen, D., and Wang, M. (2020). Pre-impact
INCLUSIVE HEALTH & CARE SOLUTIONS FOR fall detection with cnn-based class activation mapping
method. Sensors, 20(17):4750.
AGEING WELL IN THE NEW DECADE.
Sorvala, A., Alasaarela, E., Sorvoja, H., and Myllylä, R.
(2012). A two-threshold fall detection algorithm for
reducing false alarms. In 2012 6th International Sym-
REFERENCES posium on Medical Information and Communication
Technology (ISMICT), pages 1–4. IEEE.
Banos, O., Galvez, J.-M., Damas, M., Pomares, H., and Ro- Sucerquia, A., López, J. D., and Vargas-Bonilla, J. F.
jas, I. (2014). Window size impact in human activity (2017). Sisfall: A fall and movement dataset.
recognition. Sensors, 14(4):6474–6499. 17(1):198.
Bourke, A. K. and Lyons, G. M. (2008). A threshold- Tanwar, R., Nandal, N., Zamani, M., and Manaf, A. A.
based fall-detection algorithm using a bi-axial gyro- (2022). Pathway of trends and technologies in fall
scope sensor. In Medical engineering & physics, vol- detection: a systematic review. In Healthcare, vol-
ume 30, pages 84–90. Elsevier. ume 10, page 172. Multidisciplinary Digital Publish-
ing Institute.
Casilari, E. and Silva, C. A. (2022). An analytical com-
parison of datasets of real-world and simulated falls Vaishya, R. and Vaish, A. (2020). Falls in older adults are
intended for the evaluation of wearable fall alerting serious. In Indian journal of orthopaedics, volume 54,
systems. Measurement, 202:111843. pages 69–74. Springer.
de Sousa, F. A. S. F., Escriba, C., Bravo, E. G. A., Brossa, Wang, X., Ellul, J., and Azzopardi, G. (2020). Elderly fall
V., Fourniols, J., and Rossi, C. (2021). Wearable pre- detection systems: A literature survey. 7:71.
impact fall detection system based on 3d accelerom- WHO (2022). Ageing and health. WHO, N/A, n/a edition.
eter and subject’s height. In IEEE Sensors Journal, WHO (2023). World health statistics 2023: monitoring
volume 22, pages 1738–1745. IEEE. health for the SDGs, sustainable development goals.
Fleming, J. and Brayne, C. (2008). Inability to get up af- WHO, N/A, n/a edition.
ter falling, subsequent time on floor, and summoning Winter, D. A. (2009). Biomechanics and motor control of
help: prospective cohort study in people over 90. In human movement. John Wiley & Sons.
Bmj, volume 337. British Medical Journal Publishing Yu, S., Chen, H., and Brown, R. A. (2017). Hidden markov
Group. model-based fall detection with motion sensor orien-
Giuffrida, D., Benetti, G., Martini, D. D., and Facchinetti, T. tation calibration: A case for real-life home monitor-
(2019). Fall detection with supervised machine learn- ing. IEEE Journal of Biomedical and Health Infor-
ing using wearable sensors. In 2019 IEEE 17th In- matics, 22(6):1847–1853.
ternational Conference on Industrial Informatics (IN- Yu, X. (2008). Approaches and principles of fall detection
DIN), volume 1, pages 253–259. IEEE. for elderly and patient. In HealthCom 2008-10th In-
Hashim, H. A., Mohammed, S. L., and Gharghan, S. K. ternational Conference on e-health Networking, Ap-
(2020). Accurate fall detection for patients with plications and Services, pages 42–47. IEEE.
parkinson’s disease based on a data event algorithm Yu, X., Koo, B., Jang, J., Kim, Y., and Xiong, S. (2022). A
and wireless sensor nodes. In Measurement, volume comprehensive comparison of accuracy and practical-
156, page 107573. Elsevier. ity of different types of algorithms for pre-impact fall
Huynh, Q. T., Nguyen, U. D., Tran, S. V., Nabili, A., and detection using both young and old adults. Measure-
Tran, B. Q. (2013). Fall detection system using com- ment, 201:111785.
bination accelerometer and gyroscope. In Proc. of the Yu, X., Qiu, H., and Xiong, S. (2020). A novel hybrid deep
Second Int. l Conf. on Advances in Electronic Devices neural network to predict pre-impact fall for older
and Circuits (EDC 2013). people based on wearable inertial sensors. Frontiers
Jung, H., Koo, B., Kim, J., Kim, T., Nam, Y., and Kim, in Bioengineering and Biotechnology, 8:63.
Y. (2020). Enhanced algorithm for the detection of Zhang, J., Li, J., and Wang, W. (2021). A class-imbalanced
preimpact fall for wearable airbags. 20(5):1277. deep learning fall detection algorithm using wearable
sensors. Sensors, 21(19):6511.

581
BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

APPENDIX
𝑺𝑴𝑽𝒎𝒂𝒙 of SisFall dataset ADLs FALLs 𝑺𝑴𝑨𝒎𝒂𝒙 of SisFall dataset ADLs FALLs

𝐶𝑇1: min{𝑆𝑀V𝑚𝑎𝑥 : 𝑆𝑀𝑉𝑚𝑎𝑥 𝜖𝐹𝑎𝑙𝑙𝑠} 𝐶𝑇2: min{𝑆𝑀𝐴𝑚𝑎𝑥 : 𝑆𝑀𝐴𝑚𝑎𝑥 𝜖𝐹𝑎𝑙𝑙𝑠}

(a) Boxplot of SMVmax for SisFall dataset (b) Boxplot of SMAmax for SisFall dataset

𝑺𝑴𝑽𝒎𝒂𝒙 of FallAllD Waist dataset ADLs FALLs

𝐶𝑇1: min{𝑆𝑀V𝑚𝑎𝑥 : 𝑆𝑀𝑉𝑚𝑎𝑥 𝜖𝐹𝑎𝑙𝑙𝑠}

(c) Boxplot of SMVmax for FallAllD Waist dataset (d) Boxplot of SMAmax for FallAllD Waist dataset
𝑺𝑴𝑽𝒎𝒂𝒙 of FallAllD Neck dataset ADLs FALLs

𝐶𝑇1: min{𝑆𝑀V𝑚𝑎𝑥 : 𝑆𝑀𝑉𝑚𝑎𝑥 𝜖𝐹𝑎𝑙𝑙𝑠}

(e) Boxplot of SMVmax for FallAllD Neck dataset (f) Boxplot of SMAmax for FallAllD Neck dataset
𝑺𝑴𝑽𝒎𝒂𝒙 of FallAllD Wrist dataset ADLs FALLs
𝑺𝑴𝑨𝒎𝒂𝒙 of FallAllD Wrist dataset ADLs FALLs

𝐶𝑇1: min{𝑆𝑀V𝑚𝑎𝑥 : 𝑆𝑀𝑉𝑚𝑎𝑥 𝜖𝐹𝑎𝑙𝑙𝑠}

𝐶𝑇2: min{𝑆𝑀𝐴𝑚𝑎𝑥 : 𝑆𝑀𝐴𝑚𝑎𝑥 𝜖𝐹𝑎𝑙𝑙𝑠}

(g) Boxplot of SMVmax for FallAllD Wrist dataset (h) Boxplot of SMAmax for FallAllD Wrist dataset

Figure 6: The boxplots of statistical features of the dataset provide a visual representation of their distribution and the thresh-
olds.

582