A Comparison of Machine Learning Algorithms for
Fall Detection using Wearable Sensors
Nicolas Zurbuchen, Pascal Bruegger
Institute of Complex Systems (iCoSys)
School of Engineering and Architecture of Fribourg Switzerland
HES-SO University of Applied Sciences and Arts Western Switzerland
Centre for Health Technologies (CHT)
School of Electronics and Computer Science
University of Southampton
Southampton, United Kingdom
Abstract—The proportion of people 60 years old and above is
expected to double globally to reach 22% by 2050. This creates
societal challenges such as the increase of age-related illnesses
and the need for caregivers. Falls are a major threat for the
elderly, often causing serious injuries especially when the fallen
person stays on the ground for a long time without assistance.
This paper presents the development of a Fall Detection
System (FDS) using an accelerometer combined with a gyroscope
worn at the waist. Data come from SisFall, a publicly available
dataset containing records of Activities of Daily Living and
falls. We compared ﬁve Machine Learning algorithms. We ﬁrst
applied preprocessing and a feature extraction stage before
using ﬁve Machine Learning algorithms, allowing us to compare
them. Ensemble learning algorithms such as Random Forest and
Gradient Boosting have the best performance, with a Sensitivity
and Speciﬁcity both close to 99%.
Index Terms—fall detection; wearable sensors; sampling rate;
data preprocessing, feature extraction, machine learning
Falls are one of the leading causes of death among the
elderly . Every year, 28% to 35% of the elderly fall at
least once and this rate increases with age . Falls can have
severe physical, psychological and even social consequences.
They can also heavily affect the independent quality of living.
They can result in bruises and swellings, as well as fractures
and traumas . A signiﬁcant risk is the long-lie. This happens
when an elderly person remains on the ground for a long
duration without being able to call for help. It is associated
with death within the next few months following the accident
. It also affects the elderly’s self-conﬁdence who may
develop the fear of falling syndrome. It leads to anxiety when
performing Activities of Daily Living (ADLs) and can lead to
subsequent falls .
Therefore, the elderly must continuously be monitored to
ensure their safety. Families organize visits but these can be
inconvenient and even insufﬁcient. Hiring caregivers or mov-
ing into nursing homes are sometimes not affordable options.
Recent progresses in technology have enabled the development
of Assisted-Living Systems (ALSs) . They can assist the
elderly and provide a safer environment through constant
monitoring while relieving caregivers’ workload. However,
ALSs create other challenges such as privacy concerns and
acceptability issues that need to be addressed .
Fall Detection Systems (FDSs) are part of ALSs. Their
goals are to identify falls and notify caregivers so that they
can intervene as fast as possible. However, fall recognition
is challenging from a computational perspective. Falls can be
deﬁned as “the rapid changes from the upright/sitting position
to the reclining or almost lengthened position, but it is not
a controlled movement” . There is a higher acceleration
during falls. Another challenge is that falls can happen in
innumerable scenarios. They may occur anywhere at any time
. Their starting and ending body posture as well as their
direction (e.g. forward, backward) may vary . Hence, FDSs
must cover the whole living area. Their reliability must be
high while minimizing false alarms, all the while respecting
the elderly’s privacy.
In this paper, we developed a reliable FDS by the mean of
wearable sensors (accelerometer and gyroscope) and various
Machine Learning (ML) algorithms. The goal is to compare
lazy,eager and ensemble learning algorithms and assess their
results. We implemented ﬁve algorithms and tested them in
the same setup.
The rest of this paper is organized as follows. In section
II, we discuss existing FDSs and highlight their distinctive
features. Section III covers the employed methodology. Sec-
tion IV presents and discusses the obtained results. Finally, we
conclude with a comment on future work in section V.
II. RE LATE D WO RK
Scientists have employed various approaches to implement
FDSs over the past years. They have been classiﬁed as
presented in Fig. 1. Each of them has its strengths and
weaknesses. We focus on wearable technologies since we use
this approach. Nevertheless, several survey studies , 
reported the other methods in more depth.
A. Choice of sensors and sampling rate
Several types of sensors including accelerometers, gyro-
scopes, magnetometers, and tilt sensors have been used to
detect falls. Based on the fall characteristics, most studies, such
as –, employed only acceleration measurements. Only
Bourke and Lyons  used a single biaxial gyroscope and
measured changes in angular velocity, angular acceleration,
and body angle. Wang et al.  employed a heart rate monitor
and discovered that the heart rate increases by 22% after
a fall in people over 40 years old. This demonstrates that
physiological data can be used in such a system. Across these
papers, the sensors’ sampling rate varied within a range from
20 to 1000 Hz. This variation is not small, one having 50
times more samples than the other, seemingly arbitrary.
B. Sensing position
The sensor placement highly affects the detection perfor-
mance. Previous studies ,  demonstrated that better
results are achieved when sensors are placed along the longitu-
dinal axis of the body (e.g. head, chest, waist) when compared
to other placements (e.g. thigh, wrist). The movement of this
axis during a fall is more consistent and steady. However,
this requires to wear a dedicated device on uncommon body
parts which consequently creates inconveniences. For this
reason, other studies , ,  used commodities (e.g.
smartphones carried by the thigh, smartwatches worn on the
wrist). These usually do not disturb the users since they already
wear them. However, people tend to take these devices off
when they are at home which makes the FDS useless. Another
method is to combine various sensing positions. ¨
Ozdemir et al.
 developed a system consisting of six wearable devices that
are all used together. The problem is that the elderly already
have acceptability issues with one device, let alone six.
There are two categories of algorithms: threshold-based and
ML-based. Threshold algorithms simply deﬁne limit values,
outside of which, a fall is detected.. They have often been
sufﬁcient but they tend to produce false alarms especially with
fall-like activities such as sitting abruptly . To compensate,
these studies ,  added simple posture and pattern
Fig. 1. Classiﬁcation of Fall Detection System approaches.
Fig. 2. General architecture of Fall Detection Systems.
recognition algorithms that detect changes in body posture and
level of activity. This improves the detection’s robustness while
keeping a low computational complexity. However, it may still
fail during speciﬁc falls and ADLs. For example, Sucerquia et
al.  used a threshold-based classiﬁcation over their dataset
SisFall, achieving 96% accuracy.
ML algorithms automatically learn patterns based on data,
and very commonly include feature extraction. They require
more computational power and are complex to optimize but
produce improved results. Most of the studies such as ,
 employed a supervised learning technique. Common algo-
rithms are k-Nearest Neighbor , Support Vector Machine
 and Artiﬁcial Neural Network , . Yuwono et al.
 used unsupervised learning which works with clusters.
This is a compelling solution because it does not require
labelled data. Deep Learning algorithms are nowadays very
popular and achieve promising results in various ﬁelds. Musci
et al.  employed Recurrent Neural Networks to detect falls.
They used a publicly available dataset  and outperformed
the paper’s results.
D. Strengths and weaknesses
Wearable technologies have several advantages. They are
relatively inexpensive, can operate anywhere and require little
computational power all of it with minimal intrusion compared
to other approaches, such as environmental monitoring .
They can also identify the wearer and get precise measure-
ments. However, they may create discomfort due to their
size and intrusiveness. The main disadvantage is their human
dependency. These sensors must have enough battery and be
worn to work properly. Furthermore, the elderly may have a
cognitive impairment and thus, may forget to wear the sensor.
Our FDS is based on a common pipeline (Fig. 2) which has
been seen in the literature . This pipeline is a common
practice when working with ML algorithms. We ﬁrst acquire
raw data using various sensors and convert them into discrete
values. We then preprocess the raw data to remove measuring
errors which can badly affect the performance. Afterwards,
we construct and extract meaningful information in a vector.
Finally, we train and evaluate our ML algorithm to distinguish
falls from ADLs.
DETAI LS O F THE ACTIVITIES OF DA ILY LIV ING A ND FA LLS C ON TAINE D
IN T HE SisFall DATASET .
Walking slowly 100
Walking quickly 100
Jogging slowly 100
Jogging quickly 100
Walking upstairs and downstairs slowly 25
Walking upstairs and downstairs quickly 25
Slowly sit and get up in a half-height chair 12
Quickly sit and get up in a half-height chair 12
Slowly sit and get up in a low-height chair 12
Quickly sit and get up in a low-height chair 12
Sitting, trying to get up, and collapse into a chair 12
Sitting, lying slowly, wait a moment, and sit again 12
Sitting, lying quickly, wait a moment, and sit again 12
Changing position while lying (back-lateral-back) 12
Standing, slowly bending at knees, and getting up 12
Standing, slowly bending w/o knees, and getting up 12
Standing, get into and get out of a car 25
Stumble while walking 12
Gently jump without falling (to reach a high object) 12
Fall forward while walking, caused by a slip 15
Fall backward while walking, caused by a slip 15
Lateral fall while walking, caused by a slip 15
Fall forward while walking, caused by a trip 15
Fall forward while jogging, caused by a trip 15
Vertical fall while walking, caused by fainting 15
Fall while walking with damping, caused by fainting 15
Fall forward when trying to get up 15
Lateral fall when trying to get up 15
Fall forward when trying to sit down 15
Fall backward when trying to sit down 15
Lateral fall when trying to sit down 15
Fall forward while sitting, caused by fainting 15
Fall backward while sitting, caused by fainting 15
Lateral fall while sitting, caused by fainting 15
We used a publicly available dataset named SisFall . We
selected this dataset over others because of its high quality.
We assessed this quality with various criteria, namely the
size of the dataset and the diversity of subjects in terms of
age, gender, weight, and height. We also took into account
the number of falls and ADLs performed by each subject. In
the SisFall dataset, two triaxial accelerometers (ADXL345 and
MMA8451Q) and a triaxial gyroscope (ITG3200) were used
at a sampling rate of 200 Hz. We decided not to use the data of
the second accelerometer (MMA8451Q) because usual setups
only have a single accelerometer.
Twenty-three young people (19 to 30 years old) performed
15 types of falls and 19 types of ADLs including fall-like
activities. Fifteen elderly people (60 to 75 years old) also
performed the same ADLs for more authenticity. There were
ﬁve trials per activity except for the walking and jogging
activities, each of which had only one trial (See Table I).
Hence, SisFall contains a total of 4505 records including 2707
ADLs and 1798 falls, making it unbalanced. A total of 38
people including 19 women and 19 men participated. Table I
lists the falls and ADLs and their duration.
B. Data preprocessing
The SisFall dataset required minimal preprocessing. We
started by equalizing the duration of each record, by equally
cutting (top and tail in equal measure) reducing the length to
10 seconds. We chose 10 seconds to remove outliers induced
by the fall experiment, whilst preserving the fall within each
Regarding the two walking and two jogging activities, which
only have one trial (Table I), we extracted 5 times 10 seconds
for each record. We did this to have the same number of trials
per activity. We selected ﬁve windows with no overlap along
each record as follows:
1) From 5 to 15 seconds
2) From 25 to 35 seconds
3) From 45 to 55 seconds
4) From 65 to 75 seconds
5) From 85 to 95 seconds
C. Feature extraction
We then extracted meaningful information from the pre-
processed data. This process helps extracting information that
better characterize each activity. A common practice, when
working with time series, is to extract time and frequency
domains features . A time-domain feature extracted widely
in the literature , ,  is the norm of a sample (1).
We calculated the norm of acceleration and rotation mea-
sures. However, this feature alone is not sufﬁcient to allow a
robust fall detection. In the case of fall-like activities (e.g.
faster movement), this feature would probably be mislead-
ing. Thus, we also extracted time-domain features such as
the variance,standard deviation,mean,median,maximum,
minimum,delta,25th centile, and 75th centile. Additionally,
we extracted frequency-domain features, using a Fast Fourier
Transform (FFT) and we extracted two features: the power
spectral density and the power spectral entropy.
Table II summarizes the selected features. We extracted
them for each axis of each sensor (3 axes, 2 sensors) but also
for each sensor norm. This results in a feature vector of 88
features per record.
Finally, we normalized the extracted features to rescale the
data to a common scale. This gives more inﬂuence to data
with small values which can be neglected depending on the
employed algorithm. In this work, we used the common min-
max normalization (2) which scales the values between 0 and
min max[0,1] →x0=x−min(x)
D. Classiﬁcation algorithm
We selected ﬁve different types of ML algorithms among
the most widely used ones. The following list shortly describes
LIST OF EXTRACTED TIME AND FREQUENCY DOMAINS FEATURES.
Standard deviation (STD) Time
Delta (peak-to-peak) Time
25th Centile Time
75th Centile Time
Power Spectral Density (PSD) Frequency
Power Spectral Entropy (PSE) Frequency
•k-Nearest Neighbour (KNN) is a simple and popular yet
effective algorithm. Data are classiﬁed by a majority
vote with the class most represented among its k-closest
neighbours. It is a lazy learning type of algorithm because
almost no work is done until a prediction. KNN has
previously been used for fall detection and produced
promising results .
•Support Vector Machine (SVM) is also commonly used
in various tasks. It tries to ﬁnd the best hyperplane
which maximises the margins between each class. It
is an eager learning algorithm because it works a lot
during the training stage, building a model, in this case,
a hyperplane. It has often been used to detect falls either
with wearable,  where it has also produced promising
results, or video-based systems.
•Decision Tree (DT) is tree shaped. Each node is a decision
leading to a more precise category of a data for classiﬁ-
cation. It is also an eager learning algorithm because the
tree is constructed during the training process. DTs are
easy to interpret but may produce overﬁtting.
•Random Forest (RF), as its name suggests, uses multiple
DTs in parallel. Each DT is trained on a subset of data and
their results are then merged to determine the most likely
class. This allows reducing the overﬁtting generated by
DTs. It is an ensemble learning type of algorithm because
it uses multiple other algorithms.
•Gradient Boosting (GB) also uses multiple DTs but this
time in a sequence. Each DT learns iteratively on the
errors made by its predecessor. It is also an ensemble
learning kind of algorithm. It can perform better than RF
but potentially has overﬁtting issues.
The performance evaluation of our FDS under the selected
classiﬁers was done using k-folds cross-validation. This re-
quired splitting the dataset into ksets. k−1sets are used
as training and 1 as testing. The process is repeated ktimes
with a different set as the test one. Given that FDSs must be
able to detect falls for new people (e.g. unseen data), the test
set should not contain people data that the algorithm has been
We chose a value of k= 5. This creates a training set
of 80% and a test set of 20%. We ﬁltered the SisFall to
only keep subjects that performed all activities. Thus, we
removed all data created by elderly people because they have
not performed simulated falls. We also removed three young
people data due to missing records. This leaves us with
data of 20 subjects that perfectly ﬁt the split in ﬁve folds.
Consequently, we have 3400 records consisting of 1900 ADLs
and 1500 falls, making the whole data more balanced.
During the evaluation of ML algorithms, each prediction
falls in one of the following categories:
•True Negative (TN): Correct classiﬁcation of a negative
condition, meaning a reject.
•False Positive (FP): Incorrect classiﬁcation of a negative
condition, meaning a false alarm.
•False Negative (FN): Incorrect classiﬁcation of a positive
condition, meaning a missed.
•True Positive (TP): Correct classiﬁcation of a positive
condition, meaning a hit.
Each prediction is added to the count of its category which
allows then to calculate various metrics such as the accuracy. A
usual representation of these categories is a confusion matrix.
In fall detection, two metrics are especially important:
Sensitivity (SE) (3) and the Speciﬁcity (SP) (4) . The SE
(or recall) corresponds to how many relevant elements are
actually selected. This is basically the detection probability
meaning how many falls have actually been detected. The SP
corresponds to how many non-relevant elements are selected.
It means how many non-falls are actually non-falls.
Sensitivity =T P
T P +F N (3)
Specif icity =T N
T N +FP (4)
Additionally, we calculated the accuracy and the Area Under
the Receiver Operating Characteristics Curve (AUROC). The
AUROC is used to evaluate classiﬁers performance which is
used in pattern recognition and ML . In simple terms,
an AUROC close to the value of one is indicative of a
well-performing algorithm, with high true-positive and true-
negative rates consistently.
IV. RES ULT S AN D DISCUSSION
Table III presents the results of the evaluation of our Fall
Detection System (FDS) under the selected ﬁve Machine
Learning (ML) algorithms, showing that we successfully de-
veloped a reliable FDS. Both the Sensitivity (SE) and Speci-
ﬁcity (SP) surpassed 98% with the and the Gradient Boosting
algorithm, outperforming results reported by . In general,
ensemble learning algorithms achieved better performance
than others. This is because they use multiple ML algorithms,
though the improvement in performance is at the expense of
more resources. Support Vector Machine (SVM) had more
difﬁculties to distinguish the activities. However, by tuning
some hyperparameters, its results would likely improve. In an
FDS, it is desirable to detect every fall to ensure the elderly’s
safety (i.e. a perfect SE). We could then use a threshold to
improve the SE, even though it would reduce the SP and raise
more false alarms.
These results were unexpected especially prior optimization.
We infer that Activities of Daily Living and falls in the SisFall
dataset are discriminating by default, similar to . Thus,
any algorithm can perform very well. However, in real-life
conditions, the SE and SP would very likely drop because
of the falls heterogeneity as highlighted by Krupitzer et al.
. The difﬁculty of obtaining real falls data is the main
shortcoming in FDS studies, given that it is challenging to
capture them in realistic settings with the elderly.
FALL D ET ECT IO N RES ULTS O F TH E MACHI NE LEARNING ALGORITHMS.
KNN 97.26 99.31 98.41 99.45
SVM 87.93 93.78 91.20 96.43
DT 96.60 97.26 96.97 96.93
RF 98.00 98.94 98.52 99.90
GB 98.06 99.21 98.70 99.93
V. CONCLUSIONS AND FU TU RE W OR K
In this paper, we successfully developed a Fall Detection
System (FDS) using wearable technologies. Our results are an
improvement over those reported by  and , with a ﬁnal
Sensitivity and Speciﬁcity over 98%. The system is reliable
as we were able to test it on a large dataset containing several
thousands of Activities of Daily Living (ADLs) and falls. We
obtained these results using various Machine Learning (ML)
algorithms which we were able to compare. We observed that
ensemble learning algorithms perform better than lazy or eager
There is scope for future work. With the high computa-
tion resources available nowadays, it would be interesting to
explore Deep Learning (DL) algorithms. There is a study
 using Recurrent Neural Networks but there are other
algorithms available such as Convolutional Neural Networks
with the advantage of automatic feature extraction from time
series . This reduces the number of steps to implement
and removes the question of how many and which features
are needed to be extracted. The SisFall dataset allows plenty
of experiments. However, the lack of falls data availability in
realistic settings is a common challenge in FDS studies, which
also affected our study.
 L. Z. Rubenstein, “Falls in older people: epidemiology, risk factors
and strategies for prevention,” Age and Ageing, vol. 35, pp. ii37-ii41,
 World Health Organisation, “WHO global report on falls prevention in
older age,” 2008
 S. Sadigh, A. Reimers, R. Andersson, and L. Laﬂamme, “Falls and
Fall-Related Injuries Among the Elderly: A Survey of Residential-Care
Facilities in a Swedish Municipality,” Journal of Community Health,
vol. 29, pp. 129-140, April 2004
 D. Wild, and U. S. Nayak, and B. Isaacs, “How dangerous are falls in
old people at home?,” Br Med J (Clin Res Ed), vol. 282, pp. 266-268,
 P. Rashidi, and A. Mihailidis, “A Survey on Ambient-Assisted Living
Tools for Older Adults,” IEEE Journal of Biomedical and Health
Informatics, vol. 17, pp. 579-590, May 2013
 H. Hawley-Hague, E. Boulton, A. Hall, K. Pfeiffer, and C. Todd, “Older
adults’ perceptions of technologies aimed at falls prevention, detection
or monitoring: A systematic review,” International Journal of Medical
Informatics, vol. 83, pp. 416-426, June 2014
 N Noury, A. Fleury, P. Rumeau, A.K. Bourke, G. O. Laighin, V. Rialle,
and J. E. Lundy, “Fall detection - Principles and Methods,” 2007 29th
Annual International Conference of the IEEE Engineering in Medicine
and Biology Society, pp. 1663-1666, August 2007
 M. Mubashir, L. Shao, and L. Seed, “A survey on fall detection:
Principles and approaches,” Neurocomputing, vol. 100, pp. 144-152,
 X. Yu, “Approaches and principles of fall detection for elderly and
patient,” HealthCom 2008 - 10th International Conference on e-health
Networking, Applications and Services, pp. 42-47, July 2008
 S. Abbate, M. Avvenuti, F. Bonatesta, G. Cola, P. Corsini, and A.
Vecchio, “A smartphone-based fall detection system,” Pervasive and
Mobile Computing, vol. 8, pp. 883-899, December 2012
 A. M. Chan, N. Selvaraj, N. Ferdosi, and R. Narasimhan, “Wireless
patch sensor for remote monitoring of heart rate, respiration, activity,
and falls,” 2013 35th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society (EMBC), pp. 6115-6118,
 M. Kangas, A. Konttila, P. Lindgren, I. Winblad, and T. J¨
parison of low-complexity fall detection algorithms for body attached
accelerometers,” Gait & Posture, vol. 28, pp. 285-291, August 2008
 M. Yuwono, B. D. Moulton, S. W. Su, B. G. Celler, and H. T. Nguyen,
“Unsupervised machine-learning method for improving the performance
of ambulatory fall-detection systems,” BioMedical Engineering OnLine,
vol. 11, pp. 9, February 2012
 A. K. Bourke, and G. M. Lyons, “A threshold-based fall-detection
algorithm using a bi-axial gyroscope sensor,” Medical Engineering &
Physics, vol. 30, pp. 84-90, January 2008
 J. Wang, Z. Zhang, B. Li, S. Lee, and R. S. Sherratt, “An enhanced fall
detection system for elderly person monitoring using consumer home
networks,” IEEE Transactions on Consumer Electronics, vol. 60, pp.
23-29, February 2014
 C. Krupitzer, T. Sztyler, J. Edinger, M. Breitbach, H. Stuckenschmidt,
and C. Becker, “Hips Do Lie! A Position-Aware Mobile Fall Detection
System,” 2018 IEEE International Conference on Pervasive Computing
and Communications (PerCom), pp. 1-10, March 2018
 E. Casilari, and M. A. Oviedo-Jim´
enez, “Automatic Fall Detection Sys-
tem Based on the Combined Use of a Smartphone and a Smartwatch,”
PLOS ONE, vol. 10, pp. 1-11, November 2015
 T. Vilarinho, B. Farshchian, D. G. Bajer, O. H. Dahl, I. Egge, S. S.
Hegdal, A. Lønes, J. N. Slettevold, and S. M. Weggersen, “A Com-
bined Smartphone and Smartwatch Fall Detection System,” 2015 IEEE
International Conference on Computer and Information Technology;
Ubiquitous Computing and Communications; Dependable, Autonomic
and Secure Computing; Pervasive Intelligence and Computing, pp. 1443-
1448, October 2015
 A. T. ¨
Ozdemir, and B. Barshan, “Detecting Falls with Wearable Sensors
Using Machine Learning Techniques,” Sensors, vol. 14, pp. 10691-
10708, June 2014
 J. Y. Hwang, J.M. Kang, and Y. W. Jang, and H. C. Kim, “Development
of novel algorithm and real-time monitoring ambulatory system using
Bluetooth module for fall detection in the elderly,” The 26th Annual
International Conference of the IEEE Engineering in Medicine and
Biology Society, pp. 2204-2207, September 2004
 A. Sucerquia, J. D. L´
opez, and J. F. Vargas-Bonilla, “SisFall: A Fall and
Movement Dataset,” Sensors, vol. 17, January 2017
 T. Fawcett, “An introduction to ROC analysis,” Pattern Recognition
Letters, vol. 27, pp. 861-874, June 2006
 M. Musci, D. De Martini, N. Blago, T. Facchinetti, M. and Piastra,
“Online Fall Detection using Recurrent Neural Networks,” arXiv preprint
arXiv:1804.04976, April 2018
 F. J. Ord´
nez, and D. Roggen, “Deep Convolutional and LSTM Recur-
rent Neural Networks for Multimodal Wearable Activity Recognition,”
Sensors, vol. 16, pp. 115, January 2016