EAI Endorsed Transactions

on Scalable Information Systems Research Article

CURA: Real Time Artificial Intelligence and IoT

based Fall Detection Systems for patients suffering
from Dementia
Sanket Mishra1 , Bernard Ngangbam2 , Shritik Raj3 , Nihar Ranjan Pradhan4
1,2,3,4 School of Computer Science & Engineering, VIT-AP University, Amaravati, 522237, Andhra Pradesh, India

Abstract

According to the rising concern of the effects on the families due to dementia suffering patients, we aim to
provide caretakers a work-life balance in which monitoring can be done with much more ease and efficiency
in real time. This device can also be used in old age homes as well as hospitals which reduces the workload
of the caretakers and helps them to easily monitor the patients. We aim to contribute for the betterment of
the society and provide a virtual assistance for the patients suffering from dementia. The number of elderly
people living alone has been increasing all over the world. If dementia has been detected at an early stage, the
progress of disease can be slowed. The patients suffering from dementia are prone to falling quite frequently
so as to detect that and to alert their caretakers to take necessary actions. In this study, we proposed a system
in which we detect the real time state of the elderly people living alone by using the Machine Learning and
IoT (Internet of Things) technology. We installed sensors inside a finger strap which is attached to the person.
These sensors can detect the motions of the patient and predict their real time state to have a 24 by 7 support
to provide assistance to the patients.

Received on 12 July 2023; accepted on 01 September 2023; published on 25 September 2023

Keywords: Fall Detection, Accelerometer, Gyroscope, Sliding Window, Timeframe, Classification, Alert, False Alarms
Copyright © 2023 S. Mishra et al., licensed to EAI. This is an open access article distributed under the terms of the CC
BY-NC-SA 4.0, which permits copying, redistributing, remixing, transformation, and building upon the material in any
medium so long as the original work is properly cited.
doi:10.4108/eetpht.9.3967

1. Introduction people suffering from Dementia needs to tick the

elements of the device being lightweight, wearable,
In recent years, the world has been suffering from aging. reliable, non-invasive and skin comforting. A recent
There have been many studies estimating the number of study [1] in 2018, describes potential challenges of
elderly people living alone is increasing frighteningly. developing such wearable healthcare devices.
Moreover, patients who suffered from dementia also
would be increasing. In 2025, we estimated the number We aim to create a device that helps the family
of elderly people with dementia would be increased to members and the doctors of the elderly patients
7 million. suffering from dementia to be alerted in real time
Dementia is a disorder of memory and judgment about their state with the help of the Arduino pro chip
caused by dying brain cells. Therefore, patients named, Nicla SENSE ME. The chip has a BHI260AP
suffering from dementia are unable to live an ordinary motion sensor system with integrated AI, BMM150
social life. There is severe need to have a device which magnetometer, BMP390 pressure sensor, and BME688
could monitor the actions of such patients as to provide four in one gas sensor with AI integrated high linearity,
aid as soon as possible and to reduce the burden of the as well as high accuracy pressure, humidity, and
caretakers as much as possible. temperature sensors. The use of wearable device (i.e.,
The need for flexible electronic devices that can be Finger Splint) in order to monitor the day-to-day
comfortably worn by our target subject, i.e., elderly activities of a person for the health care is quite
of importance. The use of sensors like accelerometer
and gyroscope plays a vital role in this regard. For
∗ Corresponding author. Email: sanketmishra@live.com this purpose, human activities are divided into two

EAI Endorsed Transactions on

1 Scalable Information Systems
Online First
S. Mishra et al.

categories of static postures like standing, bending, being able to conveniently monitor remotely at home
sitting, and lying and dynamic transitions are the and alert the hospital unit in case of any emergency
movements and motions between the different static seamlessly via the cloud service.
postures. The accelerometers are used to detect static Another related work published in 2018 by H.
postures and gyroscopes are used to detect dynamic Nweke, T. Wah, M. Al-Garadi and U. Alo et al. [7]
transitions. We propose CURA (pronounced ‘koo-rah’ where they performed feature extraction inorder to
is the Latin word for ‘Cure’), a real-time fall detection reduce computation complexities for the proposed
system for patients suffering from dementia deep learning algorithms. We also adapted a similar
The main contribution of this paper is to present approach to reduce the computational complexity using
a unified solution for fall detection and emergency feature extraction using ts-fresh as described in (3.4).
alert. The device can monitor the patient’s activities at
home and unfamiliar surroundings as compared to the 3. Methodology
solutions that are limited to certain boundaries.
3.1. Setup Architecture
2. Related Work The architecture that we have setup for creating an
smart healthcare system for fall-detection, proposed in
In context of a fast aging population the use of this work is CURA (pronounced as ‘koo-rah’, meaning
advanced technology driven smart wearables need to be ‘Cure’ in Latin) as shown in Fig 1.
comfortably accepted by the subject in order to target
creating such devices. J. Li, Q. Ma, A. Chan and S. S.
Man et al. [2] took a survey of 146 samples from elderly
people over the age of 60 years to gain insights on the
acceptance and preference of older adults to practice
use of such devices.
Another similar study where D. Dias and J. P. Cunha
et al. [3] reviewed important aspects of these wearable
devices focusing on vital monitoring systems listing out
the state-of-art of it. They also discussed on extending
this to smart clothing wear for medical purposes and
finally targeted to learn the market trends and the Figure 1. A visual representation of the setup of CURA
future challenges of such devices.
There has been a lack of dataset for creating
health care devices such as fall-detection models using 3.2. Data Collection
sensors. E. Casilari, J. Santoyo and J. M. Garcia et al. [4] First, we gather the required data for through the
proposed a new dataset UMAFall where they acquired available sensors present on Nicla SENSE ME board
movement traces using multisensors, including the i.e., accelerometer (x, y, z) and gyrometer (x, y, z)
sensors used in this paper accelerometer and gyroscope, for detecting the movement pattern of the subject
which is used in any movement tracing related research. for creating TinyML [8] models. We collect data
Another work similar to CURA was proposed a long simulating the movements of a patient suffering from
time back in 2006, but highly relevant as a benchmark dementia and by recording the possible cases that
as C. Lin, M. Chiu, C. Hsiao, R. Lee and Y. Tsai et we need to detect, in this case especially focusing on
al. [5] proposed global positioning system (GPS) into fall/not fall detection mechanism and walk/not walk
healthcare related wearable devices, which is crucial in detection mechanism. A group of around 10 college
case of key events occuring. They provided for family students helped us in simulating the main events
members and volunteer caretakers to be able to identify (sit/sleep/walk/fall) by collecting data for 25 days at
the real-time positions of missing elderly people. irregular intervals throughout the day by trying to
Recently, in 2019 M. Al-khafajiy, T. Baker, C. simulate the events as close as possible to our target
Chalmers, M. Asim, H. Kolivand, M. Fahim, and A. subject suffering from dementia (slow and small-scale
Waraich et al. [6] proposed a remote health monitoring movements). Fig 2. and Fig 3. represents visualisation
system to monitor vitals via a mobile app portal of one such data instance recording of Fall and Walk
using Arduino UNO with alert sending mechanism to respectively. Similarly, we collected several test cases
hospitals using cloud services transferring data from within a span of 7 days, in small instances keeping
homes to hospitals through a cloud datacenter. The in mind that it should represent the simulation of our
proposed Smart Healthcare Monitoring System (SMHS) target subject a dementia patient.
used smart phones and wearable sensors to monitor Mainly two sensors for used for the purpose for
elderly people in real-time, with the functionality of CURA, namely gyroscope and accelerometer that is

EAI Endorsed Transactions on

2 Scalable Information Systems
Online First
 CURA: Real Time Artificial Intelligence and IoT based Fall Detection Systems for patients suffering from Dementia

Figure 2. Data visualization of Fall/Not-Fall signal instances

collected

embedded in the Arduino Board Nicla SENSE ME. The Figure 4. Labelling of Fall signal instances in EdgeML
Nicla SENSE ME Board was attached on a finger splint,
as shown in Fig . The subjects who contributed for the
data collection were all batch mates whose age range complexities of creating different machine learning
from 19-22 years. models for other use cases too such as Walk, Sit, Sleep,
Climbing, etc combined together which is not feasible
with limited computation power of our on-site localhost
computers used for this proposed work and thus be
able to focus on the target of fall detection using edge
computation [11]. However, for flexibility purposes
of CURA we have targeted two major classification,
i.e., Fall/Not-Fall classification and Walk/Not-Walk
classification as described in (3.5).

3.4. Feature Extraction and Sliding Window Size

Figure 3. Data visualization of Walk/Not-Walk signal instances
collected Initially, the data being collected is a time series data
in which the data points are plain readings from the
Due, to limited resources and computational power sensors of the board at a sampling rate of 15 Hz for each
of our localhost systems, we took a comparatively low sensor enabled.
sampling rate of 15 Hz for each sensor. Since, we have the data points in such a way that they
As we are dealing with fall-detection we labelled the are just the x, y and z values of accelerometer and gyro
recorded data instances into Fall class, indicating fall meter not meaning anything, we add a window size to
cases and Other class, indicating other events like Walk, define a certain action/pattern within a timeframe as
Sleep, Sit, Stand, etc. Our collected data comprises of required. This is based on the pattern of the readings
a total 33,858 data samples where 7,708 instances for at a certain window timeframe, so now the model can
Fall class, and 26,150 instances for Other class which learn the pattern/action rather than just the raw values
included, 7,565 instances for sitting, 7,292 instances for which would mean null. So, say for a window size of k =
sleeping, 7,572 instances for walk, and 3,721 instances 10, we will have the data points of the time series from
for climbing stairs. t = 1/15*10 s in a single sliding window.
Now, in order optimize this further we perform
feature extraction using ts-fresh [12] on each and every
3.3. Data Labelling
sliding window to extract the features contributing
The data collected is just raw values of x, y, z axes crucially to the classes Fall/Not-Fall or Walk/Not-Walk.
readings at a given sample rate. Using EdgeML [9] we Upon feature extraction using ts-fresh we get new
label the data as shown in Fig 4. into each class i.e., Fall features for each channel i.e., for acc_x we now have the
class and Other class (Sit/Sleep/Walk/Climbing Stairs) following,
to be able to visualize and train the pattern that exists ‘acc_x_sum_values’,
from the readings. These set of labelled data are then ‘acc_x_median’,
stored in the EdgeML [10] database and can be accessed ‘acc_x_mean’,
for training machine learning models to detect fall for ‘acc_x_length’,
CURA. ‘acc_x_standard_deviation’,
Data labelling is done such that we label data ‘acc_x_variance’,
points into Fall class and Other class inorder to reduce ‘acc_x_root_mean_square’,

EAI Endorsed Transactions on

3 Scalable Information Systems
Online First
S. Mishra et al.

‘acc_x_maximum’, The above test cases were all simulated and recorded
‘acc_x_absolute_maximum’, on with variances and permutations trying to imitate
‘acc_x_minimum’ the action of a probable subject as close as possible.
For, acc_x we get 10 such extracted features, that is
the inferred values from that particular sliding window 4. Model Creation
[13]. Similarly, we get 60 such total extracted features
for each timeframe for 6 sensor values (acc_x, acc_y,
4.1. Timeframe labelling for learning pattern
acc_z, gyro_x, gyro_y, gyro_z). As we are dealing here with a timeseries data, simply
training a model to learn the possible label i.e., Walk,
3.5. Target Classification No-Walk, Fall or No-Fall from the features, x y and
z readings of accelerometer and gyro meter would be
Since, the data collection is only limited to a very meaningless as there is no relation with the action done
small sample rate and very less battery span using before and the action coming after the present one. So,
very less computational power we have only two major to get the pattern which is to be learned by the model
working classifications. The Fall/No-Fall classification for the possible labels we need to be able to relate the
works such that in case the subject is remaining seated data points of the timeseries data.
or lying down for a rest, or is simply walking around at
their normal pace its target is to classify such patterns
as No-Fall. As soon as there is an accident and a loss of
muscle-memory coordination resulting in the falling of
the subject, the target is to classify a Fall scenario and
alert the caretakers.
Similarly, the Walk/No-Walk model works such that Figure 5. Model Architecture
the target for the pattern in which the subject is doing
anything other than walking is No-Walk. As soon as
We obtain this by using a sliding window. As shown
the subject starts, walking around, climbing the stairs
below, a sliding window now collects the datapoints of
the target then is Walk which can also be monitored
a particular timeframe and each has its own label i.e.,
simultaneously.
the action that is recorded at that particular timeframe
(Fall/No-Fall or Walk/No-Walk cases). If we have
3.6. Possible Cases datapoints for an instance of the subject walking for 15
secs and then sitting down for the next 15 s, and we
Sitting/Sleeping: In this case, the targets expected are have a sliding window of size 5 and a step size of 5, we
No-Walk and No-Fall. will have the following:
Standing Still: In this case, the targets expected are
No-Walk and No-Fall. Timeframe 1: Walk (0-4)
Walking: In this case, the targets expected are Walk Timeframe 2: Walk (5-9)
and No-Fall Timeframe 3: Walk (10-14)
Falling: In this case, the targets expected are Walk and Timeframe 4: No-Walk (15-19)
Fall. Timeframe 5: No-Walk (20-24)
Timeframe 6: No-Walk (25-29)
Here, we have Walk then fall as the fall is a result
of muscle-memory coordination loss resulting to go So, for a sliding window of size k, and a step size of
from walk phase to a fall. So, generally we can have m, and n samples at a certain sample rate, for example
three cases: we have,
a) Safe/static: Sitting/Sleeping/Standing does not 0123456789
require much attention from the caretakers as the 00000
subject and most particularly safe. _00000
b) Movement/Warning: In this case the subject is __00000
walking around and is not advised to do so in case of ___00000
no supervision and needs to be supervised immediately ____00000
to prevent any accident. _____00000
c) Falling/Alert: Here, the subject has most likely
undergone a fall or something that needs to be Here, we have n = 10 samples, window size k = 5
immediately looked upon by the caretaker as soon as and step size m = 1, therefore number of timeframes
possible. = 6. Now, as we have a pattern (at a timeframe) and

EAI Endorsed Transactions on

4 Scalable Information Systems
Online First
 CURA: Real Time Artificial Intelligence and IoT based Fall Detection Systems for patients suffering from Dementia

the particular label for the action being recorded at that above 90% but however, Decision Tree works the best as
instance we can now train the model using different it gives the output in lesser time, which is a key factor
algorithms for classifying the labels we need for this here as dealing with emergency cases, time is key.
purpose. Given, below Table 1. shows the accuracy (r-squared)
values of the 6 different classifiers at window size
4.2. Parallel Walk/No-Walk and Fall/No-Fall models (k=2,4,6).
Now, we have a two parallel model working architecture
Table 1. Accuracy scores (r-squared) on different window size (k)
that predicts the two main cases: Walk/No-Walk and
values
Fall/No-Fall, which leads us unto our 3 main possible
cases discussed earlier.
a) Safe/Static: 0/0 (No-Walk/No-Fall) Machine Learning Window Size k values
b) Movement/Warning: 1/0 (Walk/No-Fall) Models k=2 k=4 k=6
c) Falling/Alert: 1/1 (Walk/Fall) Logistic Regression 0.85 0.90 0.89
Random Forest 0.90 0.90 0.94
Depending on the output of the setup architecture Decision Tree 0.70 0.95 0.68
we obtain our required target predictor for CURA by Support Vector 0.90 0.95 0.94
combining the output of the two base models. For Naive Baye’s 0.85 0.85 0.94
learning the patterns, we used 6 different classifier K Neighbors 0.90 0.95 0.90
algorithms, namely Logistic Regression, Random
Forest, Decision Tree, Gaussian Naive Bayes, Support
Vector and K Neighbors Classifier. Since, the data was collected by simulation of an
As for learning each pattern, we require those elderly person suffering from dementia and not with
important features from the original raw data points of an actual subject there are some limitations that comes
the reading values of the used accelerometer and gyro with the simulation. Other limitations such as lower
meter we use ts-fresh for extracting the features of the battery power of the device, limited bluetooth range
timeseries data. of the Arduino Board, limited storage in the EdgeML
The problem here is that in the real-world scenario database, not being able to record a single data instance
when an emergency event has occurred, and then for any longer than around 30 minutes(there were
having to run all 6 models parallelly is computationally multiple takes recorded to increase data samples for
expensive and is not the best suited approach when learning), and low sampling rate of the sensor i.e., 15
time is a key factor in saving the subject’s life or Hz and low memory resources available for collecting
providing the needful attention immediately. larger amount of data and simulation of the possible
cases mentioned (3.6).
4.3. Optimal Model Selection Algorithm Optimising and eliminating these limitations in
We trained several models using different 6 machine future research works the model accuracy scores and
learning [14] [15] algorithms namely, Logistic Regres- the overall working can be improved.
sion [16], Random Forest [17], Decision Tree [18], Gaus-
sian Naive Baye’s, Support Vector, and K Neighbors 6. Conclusion and Future Work
Classifier [19]. So, we obtain different accuracies for
each learning algorithm. Using the Optimal Model In this research we have used cutting edge technology
Selection Algorithm approach, we select the most opti- to develop a fall-detection wearable in from of a
mal learning approach for specific test cases, thereby finger splint by using IoT sensors embedded in the
optimising on time and having to classify using only the Arduino Nicla SENSE ME board which works on edge
best optimal learning path rather than having to process impulse and provides real time fall detection alerts to
all the 6 models parallelly. This makes the work set for the caretakers and health assistance. The finger splint
the real-world challenges. device is connected via bluetooth which transfers the
data through the EdgeML cloud, which is sent to the
model through API keys. The wearable is made by
5. Results and Discussions
considering ergonomics, adjustable fit, durability and
In this research, we achieve the best optimal learning battery life as to give the subject a safe experience. The
algorithm r-squared score using Decision Tree classifier results for the Decision Tree model is the most optimal
at a sliding window of size k = 4, which gives a solution with an accuracy of 95% which is the highest
high accuracy score of 95% using the r-squared metric. compared to other models mentioned in (4.3). The alert
Other, classifiers such as Naive Baye’s, Random Forest message is sent to the caretaker and registered health
and K-Neighbors classifiers also give an accuracy of assistance through as shown in Fig 6. using the chat bot

EAI Endorsed Transactions on

5 Scalable Information Systems
Online First
S. Mishra et al.

developed on Telegram to handle the emergency as soon and Fahim, Muhammad and Waraich, Atif (2019) Remote
as possible. health monitoring of elderly through wearable sensors
(Multimedia Tools and Applications)
[7] Nweke, Henry and Wah, Teh and Al-Garadi, Mohammed
and Alo, Uzoma (2018) Deep Learning Algorithms for
Human Activity Recognition using Mobile and Wearable
Sensor Networks: State of the Art and Research Chal-
lenges (Expert Systems with Applications), 18th vol
10.1016/j.eswa.2018.03.056
[8] Kallimani, Rakhee and Pai, Krishna and Raghuwanshi,
Prasoon and Iyer, Sridhar and Alcaraz López, Onel
(2023) TinyML: Tools, Applications, Challenges, and Future
Research Directions (Sensors)
[9] EdgeML - An open source and browser based toolchain for
machine learning on microcontrollers https://edge-ml.
org/
[10] Pinyarash Pinyoanuntapong and Prabhu
Janakaraj and Ravikumar Balakrishnan and
Minwoo Lee and Chen Chen and Pu Wang
Figure 6. Alert Message via Telegram Bot to registered (2022) EdgeML: Towards network-accelerated federated
caretakers learning over wireless edge (Computer Networks)
doi.org/10.1016/j.comnet.2022.109396
[11] Murshed, M. G. Sarwar and Murphy, Christopher and
In the future we also aim to consider corner cases
Hou, Daqing and Khan, Nazar and Ananthanarayanan,
such as false alarms, i.e., the subject dropping the finger
Ganesh and Hussain, Faraz (2019) Machine Learning at
splint device which could result in the case of a false the Network Edge: A Survey (ACM Computing Surveys),
alarm of a fall detection being trigerred. For this case 54th vol
in future researches we aim to add a ground truth [12] Christ, Maximilian and Braun, Nils and Neuf-
validation mechanism using a camera module attached fer, Julius and Kempa-Liehr, Andreas (2019) Time
that would capture snapshots in case of an event and Series FeatuRe Extraction on basis of Scalable Hypothe-
cross verify it using object detection models or region sis tests (tsfresh – A Python package) (Neurocomputing)
based convulational models (R-CNNs). 10.1016/j.neucom.2018.03.067
[13] Nakamura, Tetsuya and Makio, Kosuke and Uehara,
References Kuniaki (2006) Discovering and Translating Skills from
Motion Data
[1] Ha, Minjeong and Lim, Seongdong and Ko, Hyunhyub [14] Nath, Mahendra Prasad and Mohanty, Sachi Nandan
and Daly P.W. (2018) Wearable and flexible sensors for user- and Priyadarshini, Sushree Bibhuprada B. (2021)
interactive health-monitoring devices (Journal of Materials Application of Machine Learning in wireless Sensor Network
Chemistry B), 6th vol. (2021 8th International Conference on Computing for
[2] Li, Junde and Ma, Qi and Chan, Alan and Man, Siu Shing Sustainable Global Development (INDIACom))
(2019) Health monitoring through wearable technologies for [15] Mohanty, Sachi Nandan (2021) Machine Learning for
older adults: Smart wearables acceptance model (Applied Healthcare Applications
Ergonomics), 10.1016/j.apergo.2018.10.006 [16] Seo, Jeong-woo and Kim, Taeho and Lee, Jinsoo and
[3] Dias, Duarte and Cunha, João Paulo (2018) Wearable Kim, Junggil and Choi, Jin Seung and Tack, Gyerae
Health Devices—Vital Sign Monitoring, Systems and (2019) Fall prediction of the elderly with a logistic regression
Technologies (Sensors), 18th vol 10.3390/s18082414 model based on instrumented timed up & go(Journal of
[4] Casilari-Pérez, Eduardo and Santoyo Ramón, José Mechanical Science and Technology), 33rd vol
and Cano-Garcia, Jose Manuel (2017) UMAFall: A [17] Zemblys, Raimondas and Niehorster, Diederick and
Multisensor Dataset for the Research on Automatic Fall Holmqvist, Kenneth (2016) Detection of oculomotor events
Detection (Procedia Computer Science), 110th vol using random forest
10.1016/j.procs.2017.06.110 [18] Sanchez, Veralia and Skeie, Nils-Olav (2019) Decision
[5] Lin, Chung-Chih and Chiu, Ming-Jang and Hsiao, Trees for Human Activity Recognition in Smart House
C.-C and Lee, Ren-Guey and Tsai, Yuh-Show (2006) Environments 10.3384/ecp18153222
Wireless Health Care Service System for Elderly With [19] Shivraj, Prapulla and Sinha, Dipayan and C,
Dementia (Information Technology in Biomedicine, IEEE Guruprasad and Shobha, G. and T C, Thanuja (2019)
Transactions), 10th vol Multi Mobile Agent Based Remote Health Monitoring
[6] Al-khafajiy, Mohammed and Baker, Thar and Chalmers,
Carl and Asim, Muhammad and Kolivand, Hoshang

EAI Endorsed Transactions on

6 Scalable Information Systems
Online First