Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP.

48-61, 2023

Physical Activity Monitoring for Older Adults through IoT and

Wearable Devices: Leveraging Data Fusion Techniques
Hayder Mahmood Salman1, Hasan Faleh Hamdan2*, Raed Khalid3, Sanaa Al-Kikani4
1
Al-Turath University College, Baghdad, 10021, Iraq
2
Department of computer engineering techniques, Mazaya University College, Thi Qar, Iraq
3
Department of medical instrument engineering techniques, Alfarahidi University, Baghdad, Iraq
4
Department of Physical Education and Sport Science, Al Mustaqbal University College, 51001 Hilla, Babylon,
Iraq
Emails: haider.mahmood@turath.edu.iq; eti.hassan940@gmail.com; raed.khalid@alfarahidiuc.edu.iq;
Dr.Sanaa@uomus.edu.iq,

Abstract

The emergence of low-cost individual sensing devices has facilitated the application of data fusion methods to
yield insights useful for score-level, rank-level, or hybrid-level fusion. Intelligent tools for fusion processing,
such as fuzzy methods and optimization algorithms, may be used to the deluge of raw data generated by these
devices. The use of numerous sensors allows for multi-level/hybrid-level fusion, and the combination of several
models for intelligent systems allows for fusion system design optimized for score improvement. Multimedia
data fusion applications and machine learning methods can be used to accomplish data fusion in cloud settings.
For older people in independent living conditions, a physical activity assessment framework (PAAF) that uses
deep learning models for fusion to identify activity and evaluate progress based on the spectral domain of each
window is needed. This study highlights the significance of data fusion in outlining the needs for IoT devices in
networked computers for distant patient monitoring. In order to provide for the health of the elderly without
compromising their comfort or freedom of choice, we need a seniors network based on the Internet of Things
and wearable health technology. The sensors' functionality was investigated by analyzing data gathered from the
environment and the organisms within it. The proposed PAAF-IoT architecture has many layers, each one
connected to a different device, with the most important part being the integration of data from all of them to
classify types of physical activity. Cloud services geographically close to the customer are used to process the
resulting mountain of data, reducing end-to-end delay and facilitating prompt responses from healthcare
professionals. Data fusion in healthcare and remote patient monitoring are demonstrated through the deployment
of an app that allows doctors to remotely administer prescriptions and maintain track of patients' medical
histories.

Keywords: Elder Person; Free-Living; IoT; Physical Activity; Fusion data; Fusion-Based Techniques

1. Introduction
The Internet of Things (IoT) has several important uses, one of the most important being healthcare monitoring,
which enables constant monitoring of patient states in an effort to keep them out of the hospital [1]. This is now
feasible because to the incorporation of sensors and off-site servers into e-health systems. However, the direct
interaction between the entry point client and the sensor network is not yet possible in traditional medical
systems [2]. IoT has enabled the routing of many sensing devices across the internet, thanks to developments
like IPv6. Recognizing human actions remains challenging despite technological advancements because humans
conduct actions in different locations and for varying amounts of time. Extraction of spatio-temporal knowledge
from massive datasets can help. In the same way that mobile phones have revolutionized people's lives,
wearable devices have done the same for healthcare, activity monitoring, and entertainment, and they are
expected to become much more common in the future years [3]. Technology has made it more safer for seniors
to live alone, yet falls are still a huge issue for their health. To solve this problem, we can employ multi-sensor
fusion system topologies and smart fusion processing approaches [4]. Fusion scores and intelligent system
decision-making can both benefit from the use of several models combined with deep learning. Multimedia data

Doi: https://doi.org/10.54216/FPA.110204 48
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

fusion applications, such as those pertaining to E-Systems, might benefit from data fusion techniques in cloud
settings and fuzzy approaches utilizing optimization algorithms. As a result, healthcare monitoring utilizing IoT
and wearable devices might benefit greatly from employing data fusion techniques, particularly for the elderly
[5].

When individual slips or stumbles while moving or standing, it is often referred to as a fall. Elderly individuals,
according to studies, begin to experience one fall per year. The danger of falling grows as people become older,
and most falls occur at home. If a person falls and stays on the ground for an extended period, health problems
such as exhaustion, internal injuries, and freezing may occur, and half of them will die within six months [6]. A
fall can result in various injuries, reduce a human's functional ability, and enhance their chance of dying young.
According to statistics, one out of every five falls can result in serious injuries, concussions, or fractured ribs.
Muscle pain is a significant risk factor for elderly persons, and individuals with low muscles have a higher
chance of falling [7]. Existing care processes rely on human effort to handle all difficulties. The older society
reaches a new high point in the coming years, and human work progressively diminishes in society [8]. The
existing model is no longer viable and works on a transition that merges operating systems immediately to
increase efficiency. Conventional fall detection systems necessitate specific wearable equipment [9]. Most fall
monitoring systems are mounted to the hips or breast at the center of the specimen to obtain great precision. The
primary goal of this study is to use high-tech items to improve people's standard of living. A fuzzy control
threshold-based system for fall methods relies on the triaxial accelerometer in a wearable device [10].
Individuals' heights, mass, and running gestures are used to calculate the limit in this method. When a fall
happens, a beacon uses a Received Signal Strength Indicator (RSSI) to conduct indoor localization, which can
immediately pinpoint the position of a care recipient.
The following are the achievements of this study:
• The study shows the promise of data fusion techniques, including the application of intelligent
techniques like fuzzy approaches and optimization algorithms, for processing the volume of raw data
acquired from low-cost individual sensing gadgets.
• Using many sensors and a fusion system architecture targeted for score improvement, as well as cloud
environments and machine learning approaches for data fusion, are all aspects of multi-level/hybrid-
level fusion that are highlighted in this study.
• The study suggests a PAAF-IoT model that can classify physical exercise based on the results of all
devices, and it discusses the potential advantages of data fusion in healthcare and remote patient
monitoring, such as the ease with which doctors can prescribe medications and keep tabs on their
patients' health over time.
The remainder of the article is as follows: section 2 illustrates the background to the physical activity analysis
models. The proposed physical activity analysis framework is designed and illustrated in section 3. Section 4
discusses the software outcome evaluation and analysis of the suggeted system. The conclusion and future study
of the model are discussed in section 5.

2. Related Works
To simplify the presentation of the following system, this section introduced the technological solutions used in
this research, including an accelerometer, monitoring devices, and indoor location [11]. The study's key test
aims were evaluated and described for several approaches. IoT technologies that didn't use event detection
technologies can merely identify the patient's condition, and they were ineffective and therefore unable to
forecast the patient's aberrant condition. Support vector machines (SVM) have been employed in several recent
types of research to categorize abnormal behavior using collected data dependent on tactile sensors within a
residence, although these were individually labeled. Padikkapparambil et al. suggested a method for recognizing
aberrant conditions in homes aimed mainly at the aged and those who live alone [12]. This research was the
most relevant to the job. The fundamental premise of this work was to show that by just examining the various
positions of the occupants within a home, it was feasible to educate an automated process on the behaviors of
these customers and so evaluate whether their actions are unusual [13]. On the other hand, the method was more
precise since it considers data over time and location—the scholars' above-utilized sensors to gather patient
activity data. However, the system employed two signals that anticipated the patient's next position and more
reliably identified normal/abnormal status when combined with an ECG signal [14]. Vargemidis et al. suggested
an abstraction architecture for an IoT medical system that transforms a patient's observational values into a
welfare structure that groups people with chronic behaviors for symptoms assessment [15]. A forecasting
mechanism for anomalies was included in their design, although a theoretical schema had not been deployed.
The related studies either provided an abstract architecture or were insufficiently precise; they also required a
large amount of previous data to determine a patient's aberrant condition [16]. Several studies have looked into
the possibility of detecting aberrant behavior by looking for behavior patterns different from learned typical

Doi: https://doi.org/10.54216/FPA.110204 49
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

patterns [17]. Many research studies have shown that teaching a classifier to recognize a particular event, mainly
falls, is possible. Clustering methods were also used to spot unusual patterns of behavior.
Hung et al. presented a dynamic online data habit modeling and detection (ODHMAD) model for lonely older
individuals in their living environments to conduct daily activity identification, habit designing, and outlier
detection [18]. The online activities recognition (OAR) framework and the dynamic day habit modeling
(DDHM) element make up ODHMAD. Itanalyzed sensor information in real-time to detect old people activities
and major incidents for the aged. This technology cannot forecast abnormal characteristics and necessitates
additional sensors to observe the patient's behavior [19]. These forecasting approaches were critical since they
can foresee a patient's aberrant condition and assist them to live.
Kang et al. described well-being to track the activities of elderly persons [20]. They used appliance-based
behaviors to make real-time activity identification in older adults and assessed the health functionality for these
individuals. In the user's environment, six variety of sensors are used. They identified inappropriate behavior
based on informationgathered from the surroundings and the usage of the health feature, albeit their accuracy
was limited [21]. It employed just two sensors in the research, ECG and stationary detectors, amid a preset cell,
to forecast the ECG signal's future position. MobiCart, a medical system design presented by Al-Khafajiy et al.,
included a complete variety of health-related activities to provide treatment of portable patients [22]. (1) health-
related solutions in healthcare gadgets for remote setup, self-activations, re-configuration, or perhaps even self-
repairment with current health technologies and services; (2) secure and private, flexible simulation tool
enhance or inform assistance, implemented to the codebase of a diagnostic components; and (3) distant
enrollments and (re)configurations of sensing devices and distant electronic health record services, like client
health document downloads and intermediate stage. This system was only focused on online services and lacked
accurate predictions.
By implementing a Marko-vian logic circuit hierarchial, Mozaffari et al. discovered an aberrant condition for the
client [23]. They believed that different detectors were located in various things throughout the house. The
purpose of their technique was to determine the patient's state by taking into account: 1) the patient's belongings;
2) the arrival times in a location; 3) the time of service in a place; 4) the participant's activities, and 5) the
potential of undertaking tasks simultaneously. They used two learning approaches: detecting various activities
and deriving rules that reflect the link between components and abnormalities of the status. As a result, this
approach had a significant overhead. They also didn't use ECG signals, commonly used in IoT medical systems.
Malwade et al. gathered data on bodily body movements, categorized them, and combined them with ECG
signals to diagnose a patient's aberrant condition [24]. They believed that a participant's activity indicated their
condition; nevertheless, their technique necessitated the patient wearing four additional sensors for action
detection, which was inconvenient [25]. Based on the categorization methodologies utilized, specific research
investigations have been recommended. An Artificial Neural Network was employed to design a fall detector,
and the method was used to enhance reliability and decrease the number of false alarms. An Eigenspace mimic
human drop detection method employed a multi-class SVM classifier to correctly categorize movements and
predict fall events. The Gaussian mixture model (GMM) carried out specific fall or non-fall depending on form
dynamic loading during the collapse. Motion assessment and interfering concerns were solved using a multiview
fall algorithm to identify the layered hidden Markov model. A better activity analysis module was needed to
monitor older adults.

3. Proposed physical activity analysis framework based on Fusion data

The physical activity analysis framework comprises the Local Environment Building Block (LEBB) and the
Cloud Building Block (CBB).
The LEBB is in charge of acquiring raw input on actual sensors, regardless of their techniques or modulation
schemes, and analyzing it to calculate consolidated data for the top layers. In further detail, it is in charge of
extracting the diversity of physical systems to give the basic logic with a standard view of data. It is supplied
with relevant software systems that can interface with sensing devices as per the necessary standards and
frameworks, allowing for this strategy. This capability enables the development of new modules to extend the
application's operation to emerging innovations. The information gathered by each module is subsequently
communicated to the CBB via an application programming interface (API). The LEBB is implemented in
various ways, including using the user's cellphone or a more capable device, including a windows machine,
mounted in the user's house to gather original information from fundamental physical sensing or a fitness
tracker. The platform's operational core can process incoming data and notify registered members in the event of
potentially catastrophic conditions. The detailed information is also saved in a bespoke server and made
available to doctors and approved users' relatives via an application programming interface. Even though the
developed personal information collecting system can collect data for all of the groups indicated, only the
companies listed below have been evaluated in this study.
Indoor positioning service: The LEBB's indoor navigation service operates on the smartwatch (or immediately
on the smartphone application) and is based on a Bluetooth low energy (BLE) device network. It identifies the

Doi: https://doi.org/10.54216/FPA.110204 50
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

user's location and sends it to the server platform. The data about translation is saved and given access to other
applications here.
Motility services: It can identify the user's bodily information and status. Notably, it can recognize the
participant's movement with high precision and reliability by utilizing the accelerometer sensor incorporated in
the smartwatch. The outcome of the processing cycle is then transferred to the CBB for additional examination
using the LEBB's mobile communications channel. These program data aids in identifying senior people's
behavior, their alterations, which serves as an early warning sign of more severe problems. Only low-cost and
accessible hardware were used to construct the indoor tracking and motility subsystems, discussed in the
following Sections. The preference was for the new SensorTag, which is multi-standard. The motion tracking
gadget includes a 3-axis gyroscopes, 3-axis accelerometers, 3-axis magnetometers, and a Digital Signal
Processors. Because of its small size, the Sensor Tag is ideal for wearing a bracelet since it does not obstruct the
user's motions. Furthermore, the included BLE transponder makes connecting mobile telephone touchscreens
much easier. Lastly, the redesigned Sensor Tag is driven by a Cortex M3 Microcontroller, which ultralow
energy usage enables the Sensor Tag to run simultaneously rechargeable battery for an extended period.

Figure 1: The training and testing of Fused model of the proposed PAAF-IoT model
The training and testing model of the proposed PAAF-IoT model is depicted in Figure 1. The proposed PAAF-
IoT model is first trained, and the trained model is evaluated with the test model. This framework comprises
both equipment and software. A diagram of the system illustrates the principle of the program.

3.1 System design

The hardware consists of three components: (1) a smartwatch made to wear by the consumer that quantifies
average physical data and endorses wearable sensors and placement; (2) a Wi-Fi-based beacon machine for
indoor localization and alerting transferring; and (3) an exterior air quality detector that quantifies the existing
local air quality in this region and notifies if the pollution levels are ideal for the outdoor task. A caretaker and
receiver cell phone app, the caretaker, can get a wealth of details about the care recipient from the platform. The
app determines the position whenever the care recipient has an injury, and the treatment recipient can get
medicine from the smartphone. A cloud RESTful web service that is centered on a website page is user-friendly.
It has a Database management system to save valuable data. According to the data gathered, the wearable gadget
is a popular high-tech product, similar to a smartphone or tablet. The free software Android operating system is
increasingly prevalent; Smartwatch is an app created to meet the need for a smartwatch. The primary purpose of
this device is to synchronize data with a cellular telephone. After connecting a cell device and a fitness tracker,
the wearable technology receives all the data and responds. The Wearable application can also be immediately
run on a smartwatch.
A fuzzy framework comprises a fuzzy component, a fuzzy reasoning engine, a fuzzy rule collection, and a
defuzzy component, also known as a fuzzy system or fuzz logical controller. Crisp raw data is converted into
correct fuzzy semantics via a fuzzy component. A fuzzification engine uses approximate thinking or fuzzy
inferences to replicate expert thinking and decision methods. A fuzzy rule collection contains information and
rules used to solve problems. A defuzzy module transforms fuzzy data into a crisp value from a fuzzy inference
system.

Doi: https://doi.org/10.54216/FPA.110204 51
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

Figure 2: The activities of the elderly people

The activities of the older adults identified by the proposed PAAF-IoT model are shown in Figure 2. The
identified activities are difficulty in bending, difficulty in moving head, limitation of stamina, lack of
coordination, depression, loss of sight, and loss of hearing. The fuzzy inference base comprises a set of if-then
fuzzifications that explain the platform's input-output connection. A multiple-input/multiple-output device can
be broken down into several single-input/single-output systems. There are two types of fuzzy systems used in a
multi-input/single-output scheme: (1) Sugeno Fuzzy Rule (2) Mamdani Fuzzy Guideline
The most successful and extensively used method is fuzzy management. In various fields, fuzzy control tackles
hard to handle with classical control theories. A control technique is transformed to an appropriate control
method, simply a set of IF-THEN rules relying on expert information and actions. THEN is the result of a fuzzy
system; IF is a prerequisite of a fuzzy system, i.e., a collection of circumstances. This original study collapse
detection method uses simple statistical modeling to address issues, and the design is essentially based on a
fuzzy inference system (FIS). Data is categorized using fuzzy sets, fuzzy reasoning, and fuzzy IF-THEN
principles to produce an IF-THEN fuzzification collection. As a result, the accident detection algorithm's limit
selection becomes much more versatile and tailored to every individual.
The number and structure of fuzzy sets and the antecedents and subsequent sentence parameters in the fuzzy
inference collection are defined first. An antecedents sentence in this approach comprises two factors: a and n; a
describes the current speed, and n represents the upcoming velocity input. There is only one variable in the
following statement: s is the current user status. Consider the following five fuzzy sets for every assertion
parameter (a, n, s): negative significant (NB), negative small (NS), zero (Z), positive small (PS), and positive
big (PB).
3.2 Fall detection algorithm
The following description of a fall is used considering a review of prior research: an unintended loss of balance
that results in contact with the floor or a lower level. This research does not identify a fall using deep learning
since it produces quick battery drain and danger of watch burning. A watch does not do intensive computation.
The following describes the entire wearable device with the fall detection method. When the program detects a
fall, the seriousness is assessed. Different procedures are used depending on the intensity: they extend from
storing databases to informing and providing geolocation data to the correct caregiver.
A triaxial accelerometer calculates an SVM used to detect falls. At a particular instant, the arm swing magnitude
is determined using this Equation to assess the likelihood of intense shaking. 𝐷𝑖 , 𝐷𝑗 and 𝐷𝑘 reflect the
accelerometer's velocity at instant x on the i-axis, j-axis, and k-axis, respectively. The accelerometer estimate of
the Influence of gravity is not derived when a watch senses gravity velocity. The computation and identification
are dependent on the watch's initial measurement. Since smoothing calculation simulation does not warrant
substantial processing resources, the screen is deleted after studying the outcomes after filtering. As a result, the
objective of this project is on the method configuration, which minimizes the watch computation difficulty and
energy consumption rates. The SVM function is denoted in Equation (1)
𝑆𝑉𝑀(𝑥) = √𝐷𝑖2 (𝑥) + 𝐷𝑗2 (𝑥) + 𝐷𝑘2 (𝑥) (1)

Doi: https://doi.org/10.54216/FPA.110204 52
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

The three-dimensional displacement of the older person is denoted 𝐷𝑖 (𝑥), 𝐷𝑗 (𝑥)𝑎𝑛𝑑 𝐷𝑘 (𝑥). When a typical
individual walks, the wristband accelerometer SVM result swings between 6–14 m/s2 with a 25 percent
variance, as seen in this graph. During jogging, the SVM value swings dramatically between 3–27 m/s2 with a
25% variation. After a drop, it takes around one second for a body to lose equilibrium, lean forwards, start
falling, and contact the floor, as depicted in this diagram. An individual can walk or slowly stands depending on
the hit severity. As a result, the wristwatch speed sampling frequency has been set to 250 milliseconds.
The speed of a fall varies from the maximum of regular action, given the strong impact. As a result, the
activities are categorized using a binary tree depending on the cutoff setting. The active or static condition is
identified first, and then an active ingredient condition or gesture is defined depending on the criterion. Falling,
jogging, and walking are active states, whereas standing, resting, and lying are examples of stationary gestures.
The body angle is computed by splitting the k-axis by the SVM. It is in a standing posture whenever the body
inclination is between 0° and 60°, and it is in a reclining position when the head angle is between 60° and 120°.
The angular displacement is denoted in Equation (2)
∅ = 𝑎𝑟𝑐 tan(𝐷𝑘 ) (2)
The height displacement is denoted 𝐷𝑘 . For the distinctive peaks, a method is created. Five independent
variables and a timeframe are established after the tri-axial information is retrieved. 𝐿𝑛−𝑎𝑐𝑡𝑖𝑣𝑒 , 𝐿𝑙𝑜𝑤 and 𝐿ℎ𝑖𝑔ℎ
are three accident detection criteria.
First, check to see if the SVM is in an exciting phase. The calculation comes to a halt if the SVM is inactive.
Check if the SVM is in an exciting phase and if it surpasses 𝐿𝑙𝑜𝑤 . Whenever the SVM is among 𝐿𝑛−𝑎𝑐𝑡𝑖𝑣𝑒 and
𝐿𝑙𝑜𝑤 , it may imply a shift in gesture or a little action. The clock is initiated if the SVM reaches 𝐿𝑙𝑜𝑤 , and count
activity is increased by one. Timer high grows by one whenever SVM activates a limit that surpasses 𝐿ℎ𝑖𝑔ℎ . The
next step is to see if the time is within a specific range. The triaxial information is continually collected for
identification if the period falls inside a timespan. If the time goes beyond the timeframe, it checks to see if the
clock is inside the fall counter's scope, more than MIN and less than MAX. It can tell if the value is approaching
a stable level if the clock is inside the range.
Multiple volunteers of different elevations and weights were selected to justify this approach for this report's
testing. Patterns are discovered by analyzing the data. A 10% data variance is conceivable based on the obtained
activity data. To improve reliability, this experimental procedure is then used to the 𝐿ℎ𝑖𝑔ℎ and 𝐿𝑙𝑜𝑤 criteria.
Then, detailed topic information and threshold choices are reviewed and assessed.
3.3 Automation system
Data gathering, preprocessing with training data, machine learning algorithms, and actual deployment are all
part of the computerized method.
3.3.1. Data acquisition
The information is gathered via an automated system that employs a Zigbee protocol on the garment design and
the base station. This module operates on the 2.4 GHz band, with a communication range of 10 to 25 meters
indoors and up to 250 meters outdoors. During information gathering, ten distinct participants participated in 6
various functions. Volunteers range from 45 to 70 years old, with five males and five females. All of the tests
were done with the people's permission. Data must be preprocessed once obtained to use deep learning
techniques.
3.3.2. Data preparation
Pre-processing, information balance, and feature extraction are all used to prepare data. Pre-processing entails
translating the information into an input-shaped format for every simulation with a 15% chance of getting a
sample throughout the action. It also divides the Database into training and testing datasets. Three males and
three ladies make up the train set, accounting for 75% of the total Database. The other 25% of the Database,
which consists of two men and two girls and is wholly separate from the training data, was used as a test set to
assess predictive accuracy.
The native source data exhibits an unequal training and testing data sources distribution. It had a good aim but
low specificity due to imbalanced data gathered for different activities. Data balancing methods were applied to
increase accuracy, and the concept of adequacy was selected according to an unexpected halt. Only one thing it
thought about before stopping the balance was that any activity ratio would be less than 2. The data augmented
method is denoted in Algorithm 1.

Doi: https://doi.org/10.54216/FPA.110204 53
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

Algorithm 1
Aug data (𝑫, 𝒙, 𝒚)
𝑫𝒙+𝟏 = 𝑫𝒙
𝑹 = 𝒓𝒂𝒏𝒅(𝟏, 𝒙)
𝑫𝒙+𝟏 = 𝒓𝒂𝒏𝒅 𝒔𝒉𝒊𝒇𝒕 (𝒌, 𝑫)
𝑹 = 𝒓𝒂𝒏𝒅 (𝟏, 𝒚)
𝑫𝒙+𝟏 = 𝒂𝒅𝒅 (𝑹 − 𝒏𝒐𝒊𝒔𝒆 (𝑹, 𝑫𝒙+𝟏 ))

The balancing database method is denoted in Algorithm 2.

Algorithm 2
Balance (𝑫, 𝒙, 𝒚)
𝑹 = 𝒓𝒂𝒏𝒅(𝟏, 𝒙)
For activity belonging to the Database, do
If activity_length x2 > maximum activity
then
Continue
If activity_length x2 < maximum activity
then
𝑺𝒂𝒎𝒑𝒍𝒆 = 𝒈𝒆𝒕 𝒅𝒂𝒕𝒂 𝒇𝒓𝒐𝒎 𝒂𝒄𝒕𝒊𝒗𝒊𝒕𝒚
For 𝒍 <= 𝑹 do
𝑺𝒂𝒎𝒑𝒍𝒆_𝒏𝒆𝒘 = 𝑨𝒖𝒈 𝒅𝒂𝒕𝒂 (𝑺𝒂𝒎𝒑𝒍𝒆, 𝒙, 𝒚)

The balanced algorithm's fundamental software is also employed in feature extraction due to the absence of
sample points for every operation. This program aims to increase comparable data to enrich the Database after
taking important information for any activity. To accomplish so, it introduces changes to the original
information and moves it left and right at randomized. The correlation analysis was considered for computing
the distortion and shifting quantity. A correlations matrix has been built for each action to retain the closeness
degree of every action kind in the gathered information.

Figure 3: The workflow of the proposed PAAF-IoT model

The workflow of the proposed PAAF-IoT model is designed and shown in Figure 3. The raw data is collected
and filtered, the preprocessed data is processed, and long short-term memory (LSTM) layers are used for the
classification, and the best result is stored. Multiple values in correlation coefficients were reported after
leveling and augmenting the Database, and that was the reference point for regulating and augmenting the
randomness boundary.
3.3.4 LSTM models

Doi: https://doi.org/10.54216/FPA.110204 54
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

Several studies use wearables to make use of the capacity to evaluate sequence information for activity
detection. The most often utilized architecture is proposed for constructing LSTM models. At moment p, an
LSTM memory block consists of three gates: an input gate (ig), a forget gate (fg), and an output gate (og) that
rewrite, retain, or extract the data from the memory module (mc).
Equations (3) and (4) are used to calculate the input gateway (𝑖𝑔𝑥 ) and forgetting gate (𝑓𝑔𝑥 ):
𝑖𝑔𝑥 = 𝑠𝑔𝑛 (𝐵𝑖𝑔 × 𝑖𝑥 + 𝐺𝑖𝑔 × 𝑔𝑥−1 + 𝐹𝑖𝑔 × 𝑓𝑥−1 + 𝑊𝑖𝑔 ) (3)
𝑓𝑔𝑥 = 𝑠𝑔𝑛 (𝐵𝑓𝑔 × 𝑖𝑥 + 𝐺𝑓𝑔 × 𝑔𝑥−1 + 𝐹𝑓𝑔 × 𝑓𝑥−1 + 𝑊𝑓𝑔 ) (4)
The bias weight of the input gate and forget gate are denoted 𝐵𝑖𝑔 𝑎𝑛𝑑 𝐵𝑓𝑔 . The quality of the input gate and the
forget gate are denoted 𝐺𝑖𝑔 𝑎𝑛𝑑 𝐺𝑓𝑔 . The fuzzy function for input gate and forget gate are denoted 𝐹𝑖𝑔 𝑎𝑛𝑑 𝐹𝑓𝑔 .
The input and quality of the data are denoted 𝑖𝑥 𝑎𝑛𝑑 𝑔𝑥 .
Following that, the present memory cell (𝑓𝑥 ) is modified for remembering and the fresh memory (𝑓̂) is also
changed by utilizing a quantity of the prior information (𝑓𝑥−1 ). Equations (5) and (6) denote the predicted
quality function.
𝑓̂𝑥 = tanh[𝐵𝑓̂ × 𝑖𝑥 + 𝐺𝑓̂ × 𝑔𝑥−1 + 𝑊𝑓̂ ] (5)
𝑔𝑥 = 𝑐𝑥 ⨀𝑓𝑥−1 + 𝑖𝑥 ⨀𝑓̂𝑥 (6)
The predicted bias function predicted quality function, and the weighing function is denoted 𝐵𝑓̂ , 𝐺𝑓̂ 𝑎𝑛𝑑 𝑊𝑓̂ .
The current input and previous data are denoted 𝑖𝑥 𝑎𝑛𝑑 𝑔𝑥−1 . The predicted memory cell is denoted 𝑓̂𝑥 , and the
previously stored data is denoted 𝑓𝑥−1 . The computed function is denoted 𝑐𝑥 .

Figure 4: The mathematical model of the function 𝑔𝑥

The mathematical model of the function 𝑔𝑥 is depicted in Figure 4. It uses present and previous memory
functions and the input received from older adults. The predicted data is accurate with the help of the LSTM
function. Lastly, the outputting gate (𝑜𝑔𝑥 ) that controls the array of data produced is used to determine the
ultimate activation at the present location (𝑔𝑥 ). The output gate function is denoted in Equation (7), and the
current location of the older adults is denoted in Equation (8).
𝑜𝑔𝑥 = 𝑠𝑔𝑛(𝐵𝑜𝑔 × 𝑖𝑥 + 𝐺𝑜𝑔 × 𝑔𝑥−1 + 𝐹𝑜𝑔 × 𝑓𝑥−1 + 𝑊𝑜𝑔 ) (7)
𝑔𝑥 = 𝑜𝑔𝑥 ⊙ tanh(𝑓𝑥 ) (8)
While S is the source duration. The intake activity at the present moment (x−1) is 𝑖𝑥 , while the output
installations at the previous time (x) and 𝑔𝑥−1 . B, G, and F values matrix are intake to gates, recurring links, and
cells to entryways. They think things data can be provided from the present state of a typical LSTM network.
Bidirectional LSTM networks need inputs to be of the exact dimensions. Furthermore, the present state can
determine their upcoming data input. Bidirectional LSTM's fundamental principle links two concealed layers
facing opposing orientations to the same exit. The output unit accesses results from past and upcoming states
and understands them better because of this architecture.
3.4 Computing speed of motion and detection of fall
The relative change in velocity among two successive frames is determined using collected locations of critical
points. Also calculated for every feature point are the angles concerning the x-axis (in a counter-clockwise
direction), and the axis of rotation can be calculated using these calculations.

Doi: https://doi.org/10.54216/FPA.110204 55
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

If the angle is between 𝜃 > 0° 𝑎𝑛𝑑 𝜃 < 180° , the motion is upward; else, it is downwards. The sampling unit is
an excellent example of this (a). The length between the original feature point and the exact location recorded in
subsequent frames is calculated to discover the lateral amplitude of critical points. The criterion is compared to
the most significant displacement. Using the assumption that a person's height limit is two-thirds the length of
the framework, the most significant distance traveled by the feature point in the event of a fall is calculated and
utilized as a comparative criterion.
An example video illustrates the most significant distance traveled by a critical point in the event of fall action.
The length across the region of interest in the starting frame (with plank position) and the falling frame is
displayed in this example. The upward and low movement points are also calculated throughout the subsequent
frames. The system uses these numbers to identify the prevalent motion orientation.
The most recent locations of critical points are also examined to determine if they are close to the floor. Because
the films are taken at residence, the camera is installed so that the ground of the space is covered for one-third of
the house's length from the base. It is sometimes referred to as the zone near the floor in video frames. Suppose
the comparative difference in movement performance in successive frames is substantial. The role of the
restricting key points is close to the floor, the displacements of the feature point are higher than the critical, and
the predominate orientation of movement is downward, the exercise is classified as a drop action; otherwise, it is
classified as a non-fall action.
The proposed PAAF-IoT model is designed with a fuzzy system, support vector machine for feature extraction,
and long short-term memory for the feature classification model. The internet of things model helps connect
different physical activity monitoring model components.
4. Software outcome evaluation model
The findings were achieved with the identical hardware configuration, including an Intel i5 2.4 GHz CPU, 16
GB RAM, and 8 GPU. The Python language and the Tensorflow library created all the applications. All neural
learning algorithms on databases were evaluated once the data was prepared, and several factors and comparison
factors were explored to test it. The experiment utilized red, green, and blue (RGB) films from the fall database.
The Database contains 70 films, 40 of which are non-fall and 30 collapse.

Figure 5: F score analysis of the proposed PAAF-IoT model

Doi: https://doi.org/10.54216/FPA.110204 56
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

Figure 6 : Precision analysis of the proposed PAAF-IoT model

Figures 5(a) and 5(b) show the F score and precision analysis of the proposed PAAF-IoT model, respectively.
The older adults monitored six physical activities: sleeping, showering, breakfast, spare time, lunch, and
toileting. The performance of the proposed PAAF-IoT model is evaluated, and the outcomes are compared with
the existing ODHMAD model. The proposed PAAF-IoT model is designed with IoT, fall detection, feature
classification, and physical activity analysis. The proposed PAAF-IoT model outperforms the existing system in
the entire simulation environment.
Table 1: Precision analysis of the proposed PAAF-IoT model
Activity ODHMAD (%) PAAF-IoT (%)
Sleeping 65 89
Showering 48 92
Breakfast 57 94
Spare time 61 86
Lunch 69 87
Toileting 53 91

Table 1 depicts the precision analysis of the proposed PAAF-IoT model. The older adults are continuously
monitored, the physical activities are computed, and the result is tabulated. The monitored activities are
compared with the existing model. The proposed PAAF-IoT model is initially trained with the training dataset,
and the trained model is evaluated using the testing dataset. The proposed PAAF-IoT model with a fuzzy
interference system simplifies the architecture and reduces the computation time. The fall detection model is
used to analyze continuous physical activity.

Doi: https://doi.org/10.54216/FPA.110204 57
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

Figure 7: The computation complexity of the proposed PAAF-IoT model

Figure 8: The computation time of the proposed PAAF-IoT model

Figures 6(a) and 6(b) show the computation complexity and computation time of the proposed PAAF-IoT
model, respectively. The proposed PAAF-IoT model is designed with the help of internet of things modules and
machine learning models such as support vector machine (SVM) and long short-term memory (LSTM). The
machine learning models enhance the computation speed of the proposed PAAF-IoT model, and the fuzzy
computation system reduces the complexity. The proposed PAAF-IoT model outperforms all the given
simulation environments and test conditions.
Table 9: Physical activity evaluation of the proposed PAAF-IoT model
Physical activity Self-efficacy ratio
Number of persons
ratio (%) (%)
5 52 46
10 54 48

Doi: https://doi.org/10.54216/FPA.110204 58
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

15 56 52
20 58 56
25 60 49
30 61 57
35 63 45
40 64 58
45 65 51
50 67 59

Table 2 describes the physical activity evaluation of the proposed PAAF-IoT model. The number of elderly
persons in the simulation is varied from a minimum of 5 persons to a maximum of 50 persons with a step size of
5 persons. As the number of persons in the simulation increases, the respective physical activity ratio and self-
efficacy ratio of the proposed PAAF-IoT model also increase. The higher physical activity ratio leads to better
physical and mental health for older adults. The higher self-efficacy ratio of the person leads to higher feature
extraction and classification model.

Figure 9: Physical activity ratio analysis of the proposed PAAF-IoT model

Figure 10: Self-efficacy ratio analysis of the proposed PAAF-IoT model

Doi: https://doi.org/10.54216/FPA.110204 59
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

Figures 7(a) and 7(b) show the physical activity ratio and self-efficacy ratio analysis of the proposed PAAF-IoT
model, respectively. The numbers of elderly persons for the simulation analysis are varied from a minimum of 5
persons to a maximum of 50 persons with a step size of 5 persons. The proposed PAAF-IoT model produces
better simulation outcomes with the highest accuracy. The IoT module helps to better connectivity of the
devices, and the fuzzy system reduces the system complexity of the proposed PAAF-IoT model. The machine
learning model such as support vector machine (SVM) and long short term memory (LSTM) reduces the error at
the output and produces better results at higher accuracy.
The proposed PAAF-IoT model is developed, analyzed, and evaluated in this section. The simulation findings
show that the proposed PAAF-IoT model outperforms the existing model. The proposed PAAF-IoT model with
IoT, fuzzy system, SVM based feature extraction model, and LSTM based feature classification model.

5. Conclusion
This research paper introduces a novel approach to personal health forecasting, which relies on the proposed
PAAF-IoT model with two signals, and uses a tree-structured system to estimate the health condition of nursing
older people. The method intuitively detects actions in every cell without utilizing specific sensors to capture the
patient's overall status, items, or ambient factors. The article also highlights that the ECG data collected is not
precise and a client's health condition cannot be predicted reliably based solely on the information from this
device. The research is another addition to the dispersed online forecasting of node mobility data in IoT-based
healthcare systems, which prior research has not studied. The proposed approach employs intelligent techniques
for fusion processing and fusion system design, which include multi-level/hybrid-level fusion, multi-
classifier/decision-level fusion, and fusion with deep learning models. The fusion optimization is done by
combining multiple models for intelligent systems, which can improve the fusion score and rank-level fusion.
The research also discusses various fusion system architectures, such as multi-sensor fusion system architectures
and multimedia data fusion applications.

Furthermore, the article highlights the use of fusion in robotics and decision-making, and how data fusion can be
implemented in cloud environments. The research also discusses the use of machine learning for data fusion and
fuzzy approaches for data fusion applications. Optimization algorithms for data fusion are also explored in this
research. Overall, this paper presents an innovative approach to personal health forecasting that utilizes various
intelligent techniques for fusion processing and design, highlighting the importance of data fusion in the field of
IoT-based healthcare systems. Existing systems that use action recognition algorithms to identify a patient's
medical condition are not cost-effective because they require a variety of sensors that are both expensive and
cumbersome for individuals to wear. The efficiency of using the date of delivery within a cell, ECG output,
length, and position to improve the accuracy of the suggested technique is demonstrated by simulated results. It
is intended to forecast patient mobility at varying day moments and use smartphone sensors attached to their
bodies to anticipate uncommon health conditions. It uses a complex tree-structured system to make it easier to
forecast node mobility information, the patient's condition state, and the addressing/routing/forwarding of
wireless sensor information from and/or to the gateway. In a senior treatment center, forecasting is supplied by a
customized proposed PAAF-IoT model of two outcomes applying these principles to a participant's telemetry
data and mobile ad hoc network information. The length of every state is taken into account by the proposed
PAAF-IoT model, which improves the accuracy of the forecast. The findings demonstrate that the suggested
technique is better in the process of the online health state of individuals without the need for any additional
sensors. The simulations show that the suggested method is employed for wearable sensors. The model's
performance is increased by specifying more parameters and employing various learning strategies in the future.

Funding: “This research received no external funding”

Conflicts of Interest: “The authors declare no conflict of interest.”

References
[1] Qi, J., Yang, P., Waraich, A., Deng, Z., Zhao, Y., & Yang, Y. (2018). Examining sensor-based physical
activity recognition and monitoring for healthcare using Internet of Things: A systematic
review. Journal of biomedical informatics, 87, 138-153.
[2] Shams N. Abdul-wahab, Mostafa Abdulghafoor Mohammed, & Omar A. Hammood. (2021).
Theoretical Background of steganography. Mesopotamian Journal of CyberSecurity, 2021, 22–32.
https://doi.org/10.58496/MJCS/2021/005
[3] Tun, S. Y. Y., Madanian, S., & Mirza, F. (2021). Internet of things (IoT) applications for elderly care: a
reflective review. Aging clinical and experimental research, 33(4), 855-867.
[4] Jaber, M.M., Yussof, S., Ali, M.H., Abd, S.K., Jassim, M.M., Alkhayyat, A. and Mubarak, H., 2022.
PIRAP: PPDA-FAF: Maintaining Data Security and Privacy in Green IoT-Based
Agriculture. International Journal of Cooperative Information Systems.

Doi: https://doi.org/10.54216/FPA.110204 60
Received: November 13, 2022 Accepted: March 23, 2023
Fusion: Practice and Applications (FPA) Vol. 11, No. 02. PP. 48-61, 2023

[5] Naeemullah Khan, Ismael Khaleel, & Elika Daghighi. (2021). Improved feature selection method for
features reduction in intrusion detection systems . Mesopotamian Journal of CyberSecurity, 2021, 9–
15. https://doi.org/10.58496/MJCS/2021/003
[6] Qian, K., Zhang, Z., Yamamoto, Y., & Schuller, B. W. (2021). Artificial intelligence internet of things
for the elderly: from assisted living to healthcare monitoring. IEEE Signal Processing Magazine, 38(4),
78-88.
[7] Zhong, C. L. (2020). Internet of things sensors assisted physical activity recognition and health
monitoring of college students—measurement, 159, 107774.
[8] Hosseinzadeh, M., Koohpayehzadeh, J., Ghafour, M. Y., Ahmed, A. M., Asghari, P., Souri, A., ... &
Rezapour, A. (2020). An elderly health monitoring system based on biological and behavioral
indicators in the internet of things. Journal of Ambient Intelligence and Humanized Computing, 1-11.
[9] Yacchirema, D. C., Sarabia-Jácome, D., Palau, C. E., & Esteve, M. (2018). A smart system for sleep
monitoring by integrating IoT with big data analytics. IEEE Access, 6, 35988-36001.
[10] Huang, H., Li, X., Liu, S., Hu, S., & Sun, Y. (2018). TriboMotion: A self-powered triboelectric motion
sensor in wearable Internet of Things for human activity recognition and energy harvesting. IEEE
Internet of Things Journal, 5(6), 4441-4453.
[11] Rahman, M. A., & Hossain, M. S. (2019). A cloud-based virtual caregiver for elderly people in a
cyber-physical IoT system. Cluster Computing, 22(1), 2317-2330.
[12] Alsudani, M.Q., Jaber, M.M., Ali, M.H., Abd, S.K., Alkhayyat, A., Kareem, Z.H. and Mohhan, A.R.,
2023. Smart logistics with IoT-based enterprise management system using global
manufacturing. Journal of Combinatorial Optimization, 45(2), p.57.
[13] Boukhennoufa, I., Amira, A., Bensaali, F., & Esfahani, S. S. (2020). A novel gateway-based solution
for remote elderly monitoring. Journal of Biomedical Informatics, 109, 103521.
[14] Ali, M.H., Al-Azzawi, W.K., Jaber, M., Abd, S.K., Alkhayyat, A. and Rasool, Z.I., 2022. Improving
coal mine safety with internet of things (IoT) based Dynamic Sensor Information Control
System. Physics and Chemistry of the Earth, Parts A/B/C, 128, p.103225.
[15] Vargemidis, D., Gerling, K., Spiel, K., Abeele, V. V., & Geurts, L. (2020). Wearable physical activity
tracking systems for older adults—a systematic review. ACM Transactions on Computing for
Healthcare, 1(4), 1-37.
[16] Kang, S. (2020). A study on smart homecare for daily living ability and safety management of the
elderly. In Information Science and Applications (pp. 707-710). Springer, Singapore.
[17] Wan, J., Al-awlaqi, M. A., Li, M., O'Grady, M., Gu, X., Wang, J., & Cao, N. (2018). Wearable IoT
enabled a real-time health monitoring system. EURASIP Journal on Wireless Communications and
Networking, 2018(1), 1-10.
[18] Hung, L. P., Chao, Y. H., & Chen, C. L. (2019). A Hybrid Key Item Locating Method to Assist Elderly
Daily Life Using Internet of Things. Mobile Networks and Applications, 24(3), 786-795.
[19] Forkan, A. R. M., Branch, P., Jayaraman, P. P., & Ferretto, A. (2019). An Internet-of-Things solution
to assist independent living and social connectedness in the elderly. ACM Transactions on Social
Computing, 2(4), 1-24.
[20] Kang, S. (2018). A Study on Programs Applying the Internet of Things (IoT) for Prevention of Falls in
the Elderly. In IT Convergence and Security 2017 (pp. 49-53). Springer, Singapore.
[21] Mardini, M. T., Iraqi, Y., & Agoulmine, N. (2019). A survey of healthcare monitoring systems for
chronically ill patients and elderly. Journal of medical systems, 43(3), 50.
[22] Al-Khafajiy, M., Baker, T., Chalmers, C., Asim, M., Kolivand, H., Fahim, M., & Waraich, A. (2019).
Remote health monitoring of the elderly through wearable sensors. Multimedia Tools and
Applications, 78(17), 24681-24706.
[23] Mozaffari, N., Rezazadeh, J., Farahbakhsh, R., Yazdani, S., & Sandrasegaran, K. (2019). Practical fall
detection based on IoT technologies: A survey. Internet of things, 8, 100124.
[24] Malwade, S., Abdul, S. S., Uddin, M., Nursetyo, A. A., Fernandez-Luque, L., Zhu, X. K., ... & Li, Y.
C. J. (2018). Mobile and wearable technologies in healthcare for the ageing population. Computer
methods and programs in biomedicine, 161, 233-237.
[25] Praveen, S.P., Ali, M.H., Jaber, M.M., Buddhi, D., Prakash, C., Rani, D.R. and Thirugnanam, T., 2022.
IOT-Enabled Healthcare Data Analysis in Virtual Hospital Systems Using Industry 4.0 Smart
Manufacturing. International Journal of Pattern Recognition and Artificial Intelligence.

Doi: https://doi.org/10.54216/FPA.110204 61
Received: November 13, 2022 Accepted: March 23, 2023