Choose a topic

Abstract—Fall detection is a crucial issue in elderly care and Human fall prediction using wearable sensing and machine
injury prevention. This paper presents a deep learning approach learning is a promising area of research, but it is not without
for the detection and prediction of human falls using data limitations. Despite the advancements, one of the main chal-
from wearable sensors. The algorithm employs a combination
of Convolutional Neural Networks (CNN) and Long Short-Term lenges is obtaining a sufficient amount of high-quality data that
Memory (LSTM) to analyse and classify fall events from data can be used to train machine learning algorithms [6]. The data
collected by wearable sensors. The proposed model was trained must be representative of a diverse range of falls, including
and tested on two publicly available datasets, the MHealth and different types of falls, fall scenarios, and demographic groups.
SisFall datasets, and achieved high accuracy results of 98.91% This is particularly challenging for elderly populations, as falls
and 96.79% respectively. In addition to fall detection, the study
also performed time-series forecasting on the data to predict can vary greatly depending on factors such as physical health,
falls early, using the Exponential Smoothing method. The results mobility, and medication usage [4], [6]. Another limitation is
show that the mean time difference between the moment a fall the accuracy of the machine learning algorithms themselves.
starts and the moment the model predicts a fall is 1100-1250 Many algorithms are prone to false positive and false negative
ms, demonstrating the potential for early fall prediction using errors, which can lead to either false alarms or missed falls
wearable sensors.
Index Terms—Fall Detection, Fall Prediction, MHealth, Con-
[1]. The accuracy of these algorithms is also influenced by
vLSTM, Exponential Smoothing Forecast. the quality and type of data used for training, and the com-
plexity of the features and models used [1], [2]. Additionally,
I. I NTRODUCTION there is often a trade-off between accuracy and computational
complexity, as more complex algorithms tend to be more
Fall detection is an important issue in elderly care and injury computationally expensive and may not be feasible for use on
prevention. As per the statistics, 3 out of 10 people more than wearable devices with limited computational resources [3]. To
the age of 60 go through atleast one fall per year [1]. Falls can overcome these limitations, new strategies for fall detection
cause serious consequences if not attended to promptly, and require testing and training Machine Learning models on
early response time is crucial in reducing these consequences. datasets that have some common types of falls and ADLs.
This number increases with age and previous falls, leading Some researchers have analysed the patterns of how elderly
to a fear of falling again among a significant portion of the people fall to develop more effective fall detection systems
population [1], [2]. A delayed response time to a fall can [6].
lead to serious consequences, including increased risk of death Our study focuses on the publicly available MHealth dataset
and chronic conditions. There has been a growing interest in [7]. The MHEALTH dataset is rich enough to gather informa-
developing systems for fall detection in elderly people [2]. tion regarding human body motion and vital signs recordings.
These systems generate an alert in the event of a fall, to The data is collected from ten volunteers while performing
mitigate the risk and seek medical attention on immediate physical activities [8]. The dataset is collected using wearable
basis. Many elderly people are keen to adopt such system as sensors placed on different body parts and filtered through
they generate immediate alert, make them less dependent and a sampling rate of 50 Hz, providing a comprehensive and
reduce the fear of falling [2]. The increasing number of elderly diverse representation of human movements [7], [8]. Our study
individuals living alone and the associated risk of falls has leverages this dataset to predict human falls with improved
resulted in a growing demand for fall detection systems. These accuracy and timely response. The study proposes a deep
systems are designed to provide a sense of independence, learning approach that employs a combination of Convolu-
reduce the fear of falling, and improve the response time in tional Neural Networks (CNN) and Long Short-Term Memory
case of an emergency [2]–[4]. (LSTM) to analyse and classify fall events from the data [9].
There are two common methods for fall detection: (1) Furthermore Exponential Smoothing Forecasting is being used
ambient-based sensors and (2) wearable devices. Ambient- for predicting the fall 1100-1250 ms prior to its occurrence.
based systems, commonly referred to as cameras or devices This advanced prediction capability provides a crucial advan-
containing a camera do not provide an ideal solution for tage in detecting falls and reducing the consequences related
elderly people especially the ones who don’t like to stay in a to medical attention response time.
confined place. Also more than half of the elderly falls occur In comparison to existing fall detection methods, our study
in outside places such as in a park [3]–[5]. another method is utilizes multiple sensors placed on diverse body parts, pro-
to make use of Wearable devices for fall detection as they are viding a more comprehensive and accurate representation of
robust and powerful and provide support to the elderly people human movements. The use of the Exponential Smoothing
without any locale dependency [2]. Forecasting method and machine learning techniques, allows
for a more timely fall prediction and reduces the chances of events is followed: non-fall, impact before the fall occurs, and
false positives and negatives. Additionally, the study considers actual fall occurence [10]. The non-fall interval refers to the
the execution speed, intensity, and diversity of physical activ- time when the the subject is idle or carrying out ADls, while
ities, making it more generalizable to common activities of the impact before the fall occurs indicates the time when the
daily living [6]. Our proposed Human Fall Detection system person is transitioning from a stable state to an unstable or a
is depicted in Fig. 1 : vulnerable state, on the verge of turning into an actual fall.
Finally, the actual fall interval represents the time when the
person is in an unstanle state leading the body towards a fall
[10], [11]. These labelled intervals were used for training and
evaluating the deep learning algorithms, allowing for effective
and accurate fall detection and prediction.

Fig. 1: Proposed Human Fall Detection system

Our study contributes to the field of fall detection by provid-

ing a comprehensive and accurate fall prediction system. The
use of multiple sensors and advanced prediction techniques
addresses the limitations of existing fall detection methods and
provides a more reliable solution for detecting falls in elderly
individuals. This research provides a benchmark in the field
of fall and movement detection and to address two important
issues: (1) training with features of young people that are
intended for elderly people, and (2) setting up algorithms
for maximum accuracy rather than maximum sensitivity. By
utilizing the MHEALTH dataset and the Exponential Smooth-
ing Forecasting method, the study aims to achieve a high
level of accuracy in fall prediction, even predicting the fall
a significant amount of time before it actually occurs. This
research will provide a valuable contribution to the field and
pave the way for further development and improvement in fall
detection systems for elderly people. The results of this work
are expected to provide a foundation for future researchers to
build upon and help improve the current state of fall detection
technology.
II. DATA P RE - PROCESSING AND F EATURE E XTRACTION Fig. 2: Proposed Human Fall Detection system
The MHEALTH dataset was chosen as the main data source
for developing and assessing the effectiveness of the deep Data preprocessing is a critical step in the analysis of health
learning algorithms for detecting human falls and predicting data, particularly for inertial sensor data collected by wearable
them before the fall occurs. The dataset contains a diverse devices. In the case of MHEALTH dataset, preprocessing
set of activities of daily living (ADLs) and falls, which were aims to remove any artifacts or abnormal values that may
recorded using wearable inertial sensors (accelerometer and impact the accuracy of subsequent analyses. This involves
gyroscope) with a frequency of 50 Hz [7]. One of the main the use of filtering or screening techniques to clean and
reasons for selecting this dataset is its public availability, transform the data to meet specific requirements [12]. Features
which allows researchers to compare their results with others having data outside 98% confidence interval were dropped
in the field. The dataset contains 24 different types of falls, alongside the null values. Segmentation is performed to collect
12 ADLs, and 10 subjects [7], [8]. The details of the selected the health data into smaller segments or pieces. This step
falls and ADLs in Fig. 2. is particularly important in body-motion data, where sliding
To ensure consistency with previous studies, a similar window approaches are commonly used to partition the data
labelling criterion for the data associated with three classes of [13]. In the case of MHEALTH dataset, this module paved
way for splitting the data to improve analysis accuracy. Box [13]. The threshold for the triangle feature was set to 0.19. The
plots in Fig. 3 shows the filtered features with no outliers and algorithm was optimized to maximize the lead time, which is
missing values at all. the time between the detection of a fall and the actual impact.
Feature extraction is another crucial step in health data anal- The selected thresholds for the angular velocity, acceleration,
ysis. It involves extracting general or domain-specific features and triangle feature provided 100% accuracy in detecting falls
to provide a more tractable representation of the signals. Using before impact with a lead time of approximately 1100-1250
feature extraction, the a feature vector is prepared reflecting ms.
a reduced representations of the actual features available in IV. C ONV LSTM AND E XPONENTIAL S MOOTHING
the data [10], [12]. In the case of MHEALTH dataset, this F OREECAST
module allows for the extraction of relevant features that will
In our research, we used deep learning algorithms to de-
facilitate the accurate detection of falls and ADLs. 14 different
tect falls in the MHEALTH dataset using the peak of the
Selected features from the data are available in Fig.2 . Finally,
acceleration module as a reference for the observation window
classification involves the use of fall detection and prediction
of the detector. A convolutional neural network (CNN) along
algorithms to classify data and predict falls beforehand. Ex-
with long short-term memory (LSTM) network is applied on
tracted features are traversed to the classification algorithms
MHEALTH dataset. A combination of both of these algorithms
for the classification of data. The output of this step is the
has the ability to extract features from time series data, which
categorization of the health data into relevant categories, such
is crucial for detecting falls accurately [14]. The models were
as fall or non-fall, providing valuable insights for further
trained on a subset of the MHEALTH dataset and tested on the
analysis [11]–[13].
remaining portion to evaluate their performance. The ConvL-
III. FALL D ETECTION AND P REDICTION STM model is a deep learning architecture that combines both
Falls can result in sudden peaks of acceleration when the convolutional and recurrent layers to process sequential data
body makes contact with the ground, after a brief period of free [14]. The main idea behind this architecture is to extract spatial
fall during which acceleration rapidly decreases [5]. Moreover, features from the input data using convolutional layers and
the body orientation undergoes significant changes both before then use recurrent layers to model the temporal relationships
and after the fall, resulting in substantial modifications in the between these features [15]. In the proposed ConvLSTM
values of the acceleration’s three coordinates and the angular model, hyperparameter tuning was performed to determine
velocity in the three axes measured by the gyroscope. As a the optimal number of convolutional and LSTM layers, their
result, it is crucial to use the acceleration peak as the reference width, and the dropout probability.
The ConvLSTM model consists of four convolutional
to determine the observation window of the detector [12]. This
blocks, each of which includes convolution, batch normal-
method is widely used to identify falls and pinpoint the exact
ization, ReLU activation, and max pooling operations. The
moment of impact.
convolutional layers extract spatial features from the input
The data analysis algorithm employed for fall detection in
sensor data, and the max pooling layers reduce the spatial
the MHEALTH dataset involved the use of thresholds based
resolution of the feature maps, making the model more com-
on the brusque peaks of acceleration that occur during a fall
putationally efficient. Batch normalization is used to improve
event. The impact was defined as the collision of the hip with
the stability and convergence of the model during training
the ground. The selected threshold was based on optimization
[15], [16]. The two LSTM cells in the ConvLSTM model
and aimed to maximize the lead time for detection [5]. The
extract temporal dynamics of the feature maps by learning
detection of a fall event was based on several parameters
long-term dependencies in the time sequence. The LSTM cells
including the vertical angle and acceleration of the body.
are responsible for identifying useful features spread across
Specifically, a fall was detected when the vector sum of
the time-series data in a sequence. Dropout operations are
acceleration was less than 0.9 g, the angular velocity was
used to prevent overfitting by randomly dropping out some
greater than 30°/s, and the triangular feature was larger than
neurons during training [15]–[17]. A detailed ConvLSTM
0.19 [10]–[12].
model architecture is shown in Fig. 4:
The detection of a fall before impact is based on the sudden
changes in body motion characteristics. The algorithm first
detects the increase in angular velocity, which occurs when
the body starts to tilt and lose balance [13]. The threshold for
angular velocity was set to 30°/s. Next, the algorithm looks
for a decrease in acceleration, which indicates that the body is
falling. The threshold for acceleration was set to 0.9 g. Finally,
the algorithm detects the tilting characteristic of the body by
using the triangle feature. The triangle feature is a combination
of the acceleration values along the x and z axes and the
acceleration along the y axis. This feature is used to detect
the tilting of the body, which is a sign of an impending fall Fig. 4: Implemented ConvLSTM model architecture
 Fig. 3: Box-plots for outlier detection

The ConvLSTM model has various advantages over other of detecting falls before impact, which could be useful in
machine learning models, as it effectively captures both short- preventing injuries and providing timely assistance to those
term and long-term dependencies of the motion data. In who experience falls.
contrast to the traditional CNN, which solely extracts features
and employs them as inputs of a fully connected network for V. R ESULTS AND D ISCUSSIONS
classification, the ConvLSTM model can identify and model
long-term temporal relationships in the time series data using We further analyzed the performance of our models by
a memory cell [14]. While the LSTM model can also model computing various metrics, such as sensitivity, specificity,
temporal relationships in the data, it may consume more precision, and F1 score. These metrics are commonly used
time due to its complex structure. By combining CNN and in evaluating the accuracy of fall detection algorithms. We
LSTM layers, with CNN layers extracting features from the also used confusion matrices to visualize the performance
raw data and passing them to LSTM layers for identifying of our models and identify the types of errors made by
temporal relationships, the ConvLSTM model achieves faster the algorithms. Overall, the results show that our models
computation compared to the LSTM model [14], [18]. The achieved high accuracy in detecting falls, which indicates the
proposed ConvLSTM model is anticipated to outperform other potential of deep learning techniques in fall detection using
machine learning models in predicting different fall stages the MHEALTH dataset.
because of its ability to capture both short-term and long-term The proposed Conv+LSTM model achieved high accuracy
dependencies of the motion data. and f1-score on the MHEALTH dataset, with a test accuracy of
99.38%, a loss of 0.0178, and an f1-score of 0.99. the training
Further Exponential smoothing forecast method is used to results are also shown in Fig. 5. The trained model also showed
predict falls before impact. This algorithm is a simple yet promising results on the SisFall dataset, with a fall detection
effective time-series forecasting method that uses historical accuracy of 96.79%. This demonstrates the robustness and
data to make future predictions [19]. The MHEALTH dataset effectiveness of the Conv+LSTM algorithm in detecting falls
was preprocessed and features were extracted as described in different datasets. Table I presents the comparison of our
earlier. The acceleration module data was used for exponential work with 4 different sources [21]–[24]. Our implemented
smoothing forecast, with a focus on the vertical acceleration ConvLSTM algorithm outperforms every other state of the art
component. The algorithm works by smoothing the data to work with a significant margin.
eliminate noise and seasonality and extracting a trend that Additionally, the Exponential Smoothing forecast algorithm
represents the underlying pattern of the time series [19], [20]. was also implemented to predict falls before impact, achieving
The model is trained using a sliding window of sensor data, a mean absolute error of 0.100 when predicting falls between
with a window size of 256 samples and a step size of 64 1100-1250 ms before impact. This indicates that the algorithm
samples. We used the last 128 samples in each window as is able to accurately forecast falls before they occur, providing
the test set and the previous 128 samples as the training early warning for potential falls. The forecasting predictions
set. We used the mean absolute error (MAE) as the loss were done on the MHEALTH dataset, which further empha-
function, and we optimized the hyperparameters using grid sizes the adaptability of the algorithm in predicting falls using
search. This algorithm provides an efficient and effective way different datasets. These results suggest that the combination
 TABLE I: Comparison with state-of-the-art work.

IEEEA’21 JAIHC’21 SENSORS’21 SENSORS’20 This work

[21] [22] [23] [24]
Signals A, G A, G A A, G A, G, O
Dataset Custom Custom PhysioNet Custom MHEALTH
Subjects 11 10 – 20 10
Features 30 12 11 15 14
Classifier CNN SVM Random Forest ANN ConvLSTM
Accuracy 98.1% 86.67% 94% 97.0 % 99.38%
Sensitivity 97.98 93.33 97% 97.06% 98.7%
Specificity 96.93 90.0 96% 96.5% 99.1%
F1-Score 0.98 0.9 0.95 0.9 0.99

of Conv+LSTM alongside Exponential Smoothing forecast al- [3] M. M. Islam, O. Tayan, M. R. Islam, M. S. Islam, S. Nooruddin, M. N.
gorithms has great potential for fall detection and prevention in Kabir, and M. R. Islam, “Deep learning based systems developed for fall
detection: a review,” IEEE Access, vol. 8, pp. 166 117–166 137, 2020.
real-world applications. Further studies are needed to validate [4] X. Yu, H. Qiu, and S. Xiong, “A novel hybrid deep neural network
the performance of the algorithm on a larger and more diverse to predict pre-impact fall for older people based on wearable inertial
dataset. sensors,” Frontiers in bioengineering and biotechnology, vol. 8, p. 63,
2020.
[5] S. Ahn, J. Kim, B. Koo, and Y. Kim, “Evaluation of inertial sensor-
based pre-impact fall detection algorithms using public dataset,” Sensors,
vol. 19, no. 4, p. 774, 2019.
[6] A. Sucerquia, J. D. López, and J. F. Vargas-Bonilla, “Sisfall: A fall and
movement dataset,” Sensors, vol. 17, no. 1, p. 198, 2017.
[7] O. Banos, R. Garcia, J. A. Holgado-Terriza, M. Damas, H. Pomares,
I. Rojas, A. Saez, and C. Villalonga, “mhealthdroid: a novel framework
for agile development of mobile health applications,” in Ambient Assisted
Living and Daily Activities: 6th International Work-Conference, IWAAL
2014, Belfast, UK, December 2-5, 2014. Proceedings 6. Springer, 2014,
pp. 91–98.
[8] O. Banos, C. Villalonga, R. Garcia, A. Saez, M. Damas, J. A. Holgado-
Terriza, S. Lee, H. Pomares, and I. Rojas, “Design, implementation and
validation of a novel open framework for agile development of mobile
health applications,” Biomedical engineering online, vol. 14, no. 2, pp.
1–20, 2015.
[9] J. Xu, Z. He, and Y. Zhang, “Cnn-lstm combined network for iot enabled
fall detection applications,” in Journal of Physics: Conference Series,
vol. 1267, no. 1. IOP Publishing, 2019, p. 012044.
[10] X. Hu and X. Qu, “Pre-impact fall detection,” Biomedical engineering
online, vol. 15, pp. 1–16, 2016.
[11] M. Nyan, F. E. Tay, and M. Z. Mah, “Application of motion analysis
system in pre-impact fall detection,” Journal of biomechanics, vol. 41,
no. 10, pp. 2297–2304, 2008.
[12] M. A. Kutlay and S. Gagula-Palalic, “Application of machine learning
in healthcare: Analysis on mhealth dataset,” Southeast Europe Journal
of Soft Computing, vol. 4, no. 2, 2016.
[13] Y. Nizam, M. N. H. Mohd, and M. M. A. Jamil, “A study on human fall
detection systems: Daily activity classification and sensing techniques,”
International Journal of Integrated Engineering, vol. 8, no. 1, 2016.
[14] S. P. Singh, M. K. Sharma, A. Lay-Ekuakille, D. Gangwar, and S. Gupta,
“Deep convlstm with self-attention for human activity decoding using
wearable sensors,” IEEE Sensors Journal, vol. 21, no. 6, pp. 8575–8582,
2020.
[15] J. O’Halloran and E. Curry, “A comparison of deep learning models in
human activity recognition and behavioural prediction on the mhealth
dataset.” in AICS, 2019, pp. 212–223.
[16] N. Sheraz, “Comparative analysis of machine learning algorithms on
activity recognition from wearable sensors’ mhealth dataset supported
Fig. 5: Training and Validation Accuracy and Loss plots with a comprehensive process and development of an analysis tool,”
Ph.D. dissertation, University of Bradford, 2019.
[17] G. Sannino, I. De Falco, and G. De Pietro, “A supervised approach to
R EFERENCES automatically extract a set of rules to support fall detection in an mhealth
[1] E. Casilari, M. Álvarez-Marco, and F. Garcı́a-Lagos, “A study of the system,” Applied Soft Computing, vol. 34, pp. 205–216, 2015.
use of gyroscope measurements in wearable fall detection systems,” [18] M. I. Amara, A. Akkouche, E. Boutellaa, and H. Tayakout, “A smart-
Symmetry, vol. 12, no. 4, p. 649, 2020. phone application for fall detection using accelerometer and convlstm
[2] L. Tong, Q. Song, Y. Ge, and M. Liu, “Hmm-based human fall detection network,” in 2020 2nd International Workshop on Human-Centric Smart
and prediction method using tri-axial accelerometer,” IEEE Sensors Environments for Health and Well-being (IHSH). IEEE, 2021, pp. 92–
Journal, vol. 13, no. 5, pp. 1849–1856, 2013. 96.
[19] C.-T. Chu, H.-K. Chiang, C.-H. Chang, H.-W. Li, and T.-J. Chang,
“A exponential smoothing gray prediction fall detection signal analysis
in wearable device,” in 2017 6th International Symposium on Next
Generation Electronics (ISNE). IEEE, 2017, pp. 1–4.
[20] O. Aziz and S. N. Robinovitch, “An analysis of the accuracy of wearable
sensors for classifying the causes of falls in humans,” IEEE transactions
on neural systems and rehabilitation engineering, vol. 19, no. 6, pp.
670–676, 2011.
[21] S. Sultana, R. M. Noor, F. Afifi, and M. K. Hasan, “Human fall detection
system using wearable sensors and deep learning,” IEEE Access, vol. 9,
pp. 62 810–62 824, 2021.
[22] R. Mukherjee, A. Sinha, T. K. Debnath, and J. N. Bera, “Human fall
detection using smartphone sensors and machine learning: A review,”
Journal of Ambient Intelligence and Humanized Computing, pp. 1–29,
2021.
[23] M. Sohail, S.-U. Din, N. Javaid, and W. Iqbal, “A comparative study
of machine learning algorithms for fall detection using wrist and waist
accelerometers,” Sensors, vol. 21, no. 2, p. 564, 2021.
[24] Y. Xu, X. Ji, M. Zhao, and Z. Shang, “Human fall detection based on
multiple features extracted from wearable sensor data using machine
learning techniques,” Sensors, vol. 20, no. 5, p. 1339, 2020.