Fever Status Detection using Artificial Neuron Network

Linos Nchena and

Dagmar Janacova

Department of Automation and Control Engineering, Tomas Bata University in Zlín, Czech Republic

Keywords: Assistive Technologies, Senior Citizen Assistance, Incident Detection System, Health Status Monitor,

Artificial Intelligence, Artificial Neural Networks, Machine Learning.

Abstract: This research paper proposes a monitoring system and a prototype that has been developed for detecting if a

when fever is present in senior citizens or any other specific groups of people requiring continuous care. With

various issues affecting the health of senior citizens, it is imperative to continuously monitor their health

status. The monitoring system is beneficial as it will make it feasible to enable the real time detection of fever

and thus allowing for the early treatment. Delaying treatment can lead to the underlining health issue going

beyond the remediable condition. Thus, quick detection is vital. There are various issues that might causes

illness in people. Some of the issues include virus outbreak, seasonal infections, disease, and old age. In this

paper our focus is mainly on old age. This group of people is much more at risk of getting ill or frequently

need more attention. In this project, the presence of fever or illness has been detected by using artificial

intelligence (AI). The AI technique that is utilized in this project is artificial neural networks. The computation

is done by first training the system and then secondly validating the trained system. After the training, the

system is supplied with a new set of data, with a known state, to validate that the training was successful. To

validate the system, it is provided with sample data to test its efficiency. If the system is well trained the

validation data would label that data correctly. That label is known before the validation test, as the sample

data had known labels. These known labels were not given to training but not validation system. The system

is function properly if its label matched the sample data label. The conducted experiment demonstrated a

successful detection with an efficiency rate of 82 percent.

1 INTRODUCTION

A status detection system is a computation system

used to monitor data activity and then assign a one

label on the occurring activity from given possible

data labels. An example of a data labelling is biodata

testing procedure which is intended to label the

sample as either infection found or not infection

detected. An automatic status detection system can be

used to monitor senior citizens activity status to avail

them quick support when needed. In most developed

countries, senior citizens make up a large portion of

the total population. Due to the advanced age, senior

citizens require more care and support more than

younger citizens. When taking care of senior citizen,

we must monitor their physical and health status to be

able to respond to their issues in the shortest possible

time frame. It is feasible to identify an issue, before it

becomes severe, and be able to respond to that issue.

An example of one such issue is a senior citizen fells

down and fails to get up on their own. A fell status

system can notify a caregiver who is away from the

house and then the caregiver can get back home to

assist the fallen senior in good time. A delay in this

situation could be fatal. Therefore, this work intent to

conduct early detection of fever before the situation

is out of controllable stage. This detection would

make it to response to issue at a stage when the issues

are still in remediable stages (Garçon et al., 2016).

There are four advantages why this system is very

helpful; Firstly, this system can help decongest care

giving facilities. Caregiver’s facility would get

congested, in a case where every person needing care,

goes to live at the facility for physical monitoring and

attention. Therefore, a system that can remotely

monitor issues can benefit caregivers’ facilities by

attending to more clients since remote monitoring

accommodates more clients than on-premises

monitoring. The second advantage is that for some

groups of people, for example senior citizen, such

fever is so much common that they require frequent

medical attention. Therefore, this monitoring system

would allow for identifying the right moment when to

776

Nchena, L. and Janacova, D.

Fever Status Detection using Artiﬁcial Neuron Network.

DOI: 10.5220/0010485407760781

In Proceedings of the 23rd International Conference on Enterprise Information Systems (ICEIS 2021) - Volume 1, pages 776-781

ISBN: 978-989-758-509-8

receive direct human attention. Without such a

monitoring system, such a needy person might delay

or ignore to identify the moment when to seek direct

human attention. Thus, a remote monitoring system

that can identify the fever in its earliest possible time

would allow treatment before the issue goes out of

control. The third advantage is that a detection system

can reduce the stress level on care givers. Caregivers

usually make frequent checks on the person under

their monitoring. In a case that a caregiver attends to

other works, it because difficult to monitor their

subject. Using an intelligent monitoring system, the

caretaker could receive real-time data about the

subject’s health status. Therefore, caregiver would

have less chances of missing status record data since

the system would be sending status updates in real-

time. Lastly, the importance of a monitoring system

is the freedom and independence that the subject

person gained from using the system. Since the

caregiver will not have to constantly monitor them,

this allows the subject to feel independent. They are

relieved of the feeling of having someone constantly

physically monitoring then. Some subjects would feel

guilty when caregivers give them more attention as

they feel they are inconveniencing the caregiver.

With this system applied, the subject might feel they

are not a burden to caregivers as the system assumes

most of the caregivers’ responsibilities (Paudel et

al.,2018; Hussein et al., 2014; Das et al., 2015).

This study is composed of the following parts:

The second chapter present fundamentals of artificial

neural networks. The third chapter presents related

works done by previous research works. Then the

fourth chapter presents a proposed methodology of

solving the problem. The fifth chapter gives the

experimental results. Sixth chapter presents

discussion of the results and prospects of the research.

Seventh chapter is the last and gives the conclusion.

2 ESSENTIALS OF ARTIFICIAL

NEURAL NETWORKS

An artificial neural network (ANN) is a data

processing system that is inspired by how the human

or animals’ brain processes data (Nasser et al., 2019).

The brain has several processing units, that are

interconnected and using these interconnections, they

can map input data to specified output data. Each one

of the units in these interconnections is known as a

neuron. The human brain has over 100 billion

neurons, which are interconnected in several ways. It

operates in such a way that it gets data from sensors,

and then passes that data to its processing mechanism.

The processing mechanism then manipulate this

collected data to generate output information. In a

similar approach ANN imitates the brains’ data

processing structure. Like the brain, the ANN has

three components: input layer, hidden layer, and

output layer (Nasser et al., 2019; Chatzimichail et

al.,2013; Ajerla et al., 2016).

An ANN has multiple layers, and each layer

processes data as a component of a processing layer

group. The data gets processed at every single layer.

The data gets processed as many times as there are

layers. If the network has only one layer, then the

processing would only occur in one step. In case of a

multilayer ANN, the processing is performed from

one layer to the next layer and then next layer until all

layers have had a turn in processing the data. This

means the data is processed repeatedly based on

number of hidden layers in that network. In a network

the number of neurons or number of layers are

designed based on the problem that is been solved.

However, the number of neurons should be decided

appropriately. Simply increasing the number of

neurons without some corresponding importance in

features used in the problems would not automatically

improve the performance of the network. Below is

Figure 1, which shows the three layers of an ANN.

Figure 1: The three main layers in ANN topology.

The hidden layer is usually the most complex of the

three layers of a network. The mapping of input to

output is achieved after the ANN has undergone some

training sessions. After the training, the network can

then be used to solve the problem it was trained to

solve. It solves the problem after it has mastered the

link between every input and to corresponding

outputs. ANN are used in solving a variety of

problems. Common problems solved by using ANN

include handwriting recognition, disease diagnosis,

process control, financial viability predictions, stock

market predictions, complex systems modeling, error

compensation in industrial processes, fire control,

Fever Status Detection using Artiﬁcial Neuron Network

777

security surveillance, etc. (Nasser et al.,2019; Ajerla

et al., 2019; Janku et al., 2018; Mehr et al., 2016).

3 RELATED WORK

Within the existing literature, the detection problem

has been solved in different ways by various

researchers. A list of research of particular interest

have been incorporated in presenting this research.

Chatzimichail et al. (2013), have discussed the

detection ways to determine the presence of Asma in

children under the age of five. The is done based on

recognized symptoms as features of presence of

Asma disease. The experiment was conducted by

collecting a sample from 112 records which have 48

features. To solve the issue the researchers decided to

reduce the number of features to nine from 48. This

was done because the removed features had little

impact on the results of the experiment. For analysis

purposes, the experiment was performed twice. First

with the full number of features at 48 and then

secondly only with the nine to illustrate the need to

have some features removed. During pre-processing,

data was divided in ten equal sets. Ten cycles of

training were performed using these ten datasets. For

every cycle, one data set is used as testing data where

the remaining nine sets are used as training data. The

total results obtained are then summated to obtain an

overage of the training accuracy. The experiment

results showed that removing the features that had a

smaller impact on results of experiment made the

ANN much more effective by raising accuracy from

83.87% to 96.77%.

Ajerla et al. (2016), considers an application for

providing various service to senior citizens using

artificial intelligence detection. The system offers

services that include fire detection, gas leak detection

and unaccompanied monitoring. The task was to

improve the performance of an algorithm if the sensor

was place on the waist rather than on the head or

wrist. This was because head or wrist is more accurate

but is less comfortable for the subject compared to the

wrist. The rest has more vector movements that the

head of waist of which these movements are the input

of the ANN. 525 data sets where collected. Because

they were of different sizes, some of the data was

disused and some of the data was normalized but

adding zeros where they had no entry to make all the

dataset have same size. The final data used was 120.

The 120 sets were divided into 90 as training data set

and 30 as testing set. An ANN of three hidden layers.

The ANN is trained to detect the occurrence or no-

occurrence of a fall. The experiment concludes that

the detection of fall from the waist and head in

previous experiment was at 95% while in this

experiment it was at 75%. The 75% detection

accuracy for the sensor on the wrist was considered

an improvement as the waist position is more

convenient than the head. A similar research is

conducted by Yoo et al. (2016). Both these systems

are used as a real-time motoring system for falling

and hence caregivers are updated immediately on

occurrence of falling.

Janku et al. (2018), presents a research about a

new method of fire detecting technique using neural

networks It focuses on the issue with current systems

that they have difficult in differentiating controlled

fires from dangerous fires. Controlled fires are fires

that are specifically started and are not a danger to life

or property. For instance, a fire from a welding

machine when using a welding thing in a warehouse.

For the experiment, the research required to use three

different types of sensors. A sensor for smoke, a

sensor for colour and a sensor for movement

direction. The three sensors would collect data from

the environment and then send it to the centre of the

ANN. The networks are of two kinds, the shallow nets

and deep leaning machine. The two of them differ in

the sense that the shallow nets are consist of only

three layers, while the deep learning machines has

more than three layers. The basic layers are input

layer, hidden layer, and output layer. In the deep

learning machines the hidden lawyer is not one but

several layers. The data from each of the three sensors

was used in this experiment. The researcher also

stated current systems use one sensor compared to the

three that this experiment is utilizing. Furthermore,

this work intended to remove a scenario of having a

high error values in the detection system. The cited

previous research works are said to have a lot of false

negatives and false positives. The study experiment

provides interesting results that proves a better

method to detect fires. The new method has provided

results fire detection with accuracy of 93%. This

system operated online hence a real-time motoring

system for fire and hence care takers are updated

immediately on the occurrence of fire.

In implementing our experiment, we shall take the

following direction. In training data, we shall set the

size of training data at 80% instead of 60% used by

Kajan et al., (2014) and 75% by Yoo et al., (2016).

This makes the system more specific and less generic

a good preference in this problem. We shall also limit

the parameters we select to those that have the highest

impact among the list of probable parameters.

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

778

4 RESEARCH METHODOLOGY

Senior citizens face several health issues. Neural

networks can be utilized to protect them from such

issues that they face in their daily lives (Shahid et

al.,2007; Amato et al., 2007; Mehr et al., 2016). The

application of ANN could improve and enhance

lifespan for the seniors. To develop a monitoring and

alert notification system, we shall design a prototype

for detecting fever in senior citizens and other citizen

in risk category. We shall use pseudonymized and

anonymized data to train and test this application.

4.1 Method and Procedure

To help identify whether a medication alarm should

be alerted, the following three tasks must be

performed. (1) The features of the sample data are

evaluated for relevancy. We use only those features

which have high relevancy to the identification of

medical condition. Using the square-mean error of

each features’ error, the features with small square-

mean error changes are removed. The regression

method is the method we shall apply to identify these

less significant features. (2) After this feature removal

operation, a feedforward-propagation ANN is trained

on this dataset to help identifies the status of the

medical condition. The dataset is divided into five

equal sub-datasets. In iterations, each of these five

sub-datasets is used as a test set while the rest is used

as a training set. This iteration is used to avoid cross-

validation and to make sure that every sub-dataset is

used at least once as a testing set. (3) After the training

sessions, the test dataset is feed into the system for

testing procedure. If the tests results are of some

specified accuracy level, then the design is successful.

The designed systems can then be used for a subject

dataset to predict status of medical condition. This is

the final step of the experiment. In summary, we used

three different datasets in this network. First is

training dataset, second is testing data set and last is

the subject dataset. The third dataset is the candidate

dataset that is under investigation for the status of

medical condition.

4.2 Experiment Data and Used Tools

The artificial neural network was developed in

Python programming language. Python has

supporting libraries for ANN implementation. In this

experiment the libraries that we used include keras,

matplotlib, pandas and TensorFlow. The computer

used had the following specifications: Processor:

Intel(R) Core (TM) i7-4510U, CPU: 2.00GHz, 2

cores, RAM Memory: 4 GB DDR3 1600 MHz, OS:

Windows 64 bits.

The data we used consists of one thousand

medical records. The records have 8 fields, each

representing a unique data feature of the record. We

divided the records into two groups. One group is to

be used as training dataset and the other as a

validation dataset. The records ratio is 8 to 2; where

training set is (80%) and validation set is (20%). The

experiment used data from different sources that

corelated to feature detection as advised by medical

specialists. The data’s features selected are not

selected at random. These are features that have been

identified to be direct or indirectly linked to the

presence or absence of fevers (Chatzimichail et

al.,2013). The ANN model has TensorFlow backend.

it has three layers, the first layer has 12 neurons,

second layer has 8 and last layer has two neurons. The

training dataset is online database available from

https://data.world/anaozp/diabetes. This dataset has

been used to train and test the ANN.

5 RESULTS

Based on the collected data we trained the neural

networks and attained an accuracy of 79.8%. Several

experiment trials were conducted. The accuracy of

79.8% is acceptable to successfully determine

presence or absence of fever.

Figure 2: Accuracy evolution of the training and testing

sessions.

Figure 2 shows the results for the training and testing

sessions that were conducted.

shows the level of

accuracy from this specific training and testing in the

machine learning experiment. As can be seen from

the figure, the training took 300 epochs with training

accuracy results ranging from 0 % to 82 %. However,

the feasible testing range was shorter, starting from

Fever Status Detection using Artiﬁcial Neuron Network

779

65 % ending at 82 %. Within the first 20 epochs, the

training accuracy rises sharply from 0 to 70 %. From

the 20th epoch, it steadily grows until it reaches 78 %

by the 250th epochs. It then marginally grows and

then stabilizes around 80% till the end of the

experimental period. As proven in the figure, there

are less data spikes in system training than in system

testing. The training-testing deviation starts at 67 %

on the 10th epoch. Spikes grow until around the 120th

epoch. They then gradually reduce to a 10%

difference by the 250th epoch. This spike reduction

indicates acceptable training performance. The

training and testing accuracy start to stabilize at about

the 150th epochs with 78.1% and 79.8% accuracy,

respectively. Accuracy starts from its lowest point

and goes upwards as the epochs increases. It stops

increasing at about 78.1% to 79.8% and stabilizes

there. Therefore, this experiment is a success as the

test performance indicates acceptable performance on

actual subjects’ data.

The training and testing loss start to stabilize at

around after the 120th epochs with 0.140 and 0.145

loss, respectively. The loss starts from the highest

point and starts declining. It continues a steady

decline until it reaches around 0.141 of the loss where

it then stabilizes.

Figure 3: Loss and accuracy evolution of the Training and

testing sessions.

In Figure 3, loss and accuracy metrics are shown.

The accuracy metric increases while the loss metric

decreases with the increase in the number of epochs.

6 DISCUSSION

The experiment’s result illustrates the usability of this

neural network in fever status detection as a vital

component of assistive technologies. Instances of

expected benefits of such a system includes. The

system allows early detection of an illness without

having a physically residing at a care facility. ANN

has been applied to identify illness status for subject

senior citizens. ANNs with a success rate at 81% did

manage to label the medical status based on specific

features. The subject citizen had sensors that are

taking record of changes in selected body parameters

(Ajerla et al., 2019; Yoo et al.,2018). The sensors can

be attached or unattached to the body depending on

preference. Rather than subject been located at a

monitoring care facility, sensors send data to the

facility making the monitoring process much more

convenient. This makes subject more self-reliant and

improves quality of life and reduces expensive for

cost of care and attention. This can also make the

monitoring process more robust and effective.

Furthermore, the ATs can make automatic diagnosis

and then alert relevant caregiver for further care and

prescription. This would as a result reduce or even

eliminates workload for personnel at care facilities.

With an increase in the population of senior

citizens, the need for monitoring systems would

increase. According to the World Health

Organization, (2002), in 2025, there would be a total

of about 1.2 billion people over the age of 60. By

2050 this number would increase to 2 billion with

80% of them living in developing countries. The

population growth is faster for older persons than for

the rest of the population (United Nations, 2007). In

1950 the total number of people over 60 years old was

8 percent. By 2007, this percentage had grown to 11

percent. By 2050, this number is projected to be at 22

percent. The increase in elderly population means

more people would require assistive living care than

before. Thus, detection systems are becoming vital

and would require more investment both academic

and financial. They could also become a usual home

appliance for the elderly family member of the home.

In future research we hope to focus on a system,

which includes a notification message and pill taking

schedule. Apart from fever detection system, this

application can be coupled with other application to

form a complex system offering several services. We

wish to integrate these into this fever detection as

separate system components.

7 CONCLUSION

In this paper, we have discussed fever status detection

using an artificial neural network. An experiment was

setup to perform fever status detection using neural

networks. The experiment has provided insights on

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

780

the viability of a fever detection system. The system

could be useful in reduction of cost in the elderly

citizens care and medical services. Additionally, we

outlined three key benefits for the use of assistive

technologies systems in rapid detection of health

status or condition of person under observation.

1) The remote and quick diagnosis would allow easier

and continuous surveillance and enable early

treatment of illness.

2) An automated system could minimise subject

persons’ dependence and reduce stress on both the

care giver and the person receiving care. Less stress

and less dependence for the concerned senior citizen

could be a good remedy for improvement of health.

3) An automated system would reduce the cost for

care and medical services. Employing machines

would costs less than employing human monitoring

assistants. Early diagnosis can also be achieved

without automated technologies; however, such a

method would require more costs than using the

automated detection system.

The common solution to senior citizens

monitoring is to have the citizen reside at a care

facility. Such solution costs more for the senior

citizen and is more workload for the care facility.

Hence a fever detection system is a better option as it

addresses both these issues. This detection services

could be installed at a home, housing a senior citizen.

This system could be able to support in the everyday

lives’ activities and care service delivery to senior

citizens and other vulnerable citizens. With such a

system in common usage, the lives of millions of

senior citizens across the globe could be improved.

ACKNOWLEDGEMENTS

This work was supported by IGA/CebiaTech/2021/

001, a research project of the Faculty of Applied

Informatics, Tomas Bata University in Zlín.

REFERENCES

Garçon, Loïc, Chapal Khasnabis, Lloyd Walker, Yukiko

Nakatani, Jostacio Lapitan, Johan Borg, Alex Ross, and

Adriana Velazquez Berumen. "Medical and assistive

health technology: meeting the needs of aging

populations." The Gerontologist 56, no. Suppl_2

(2016): S293-S302.

Paudel, Ramesh, Kimberlyn Dunn, William Eberle, and

Danielle Chaung. "Cognitive Health Prediction on the

Elderly Using Sensor Data in Smart Homes." In The

Thirty-First International Flairs Conference. 2018.

Hussein, Ali, Mehdi Adda, Mirna Atieh, and Walid Fahs.

"Smart home design for disabled people based on

neural networks." Procedia Computer Science 37

(2014): 117-126.

Das, Resul, Gurkan Tuna, and Ayse Tuna. "Design and

Implementation of a Smart Home for the Elderly and

Disabled." environment 1 (2015): 3.

Nasser, Ibrahim M., and Samy S. Abu-Naser. "Predicting

Tumor Category Using Artificial Neural Networks."

(2019).

Chatzimichail, Eleni, Emmanouil Paraskakis, and

Alexandros Rigas. "Predicting asthma outcome using

partial least square regression and artificial neural

networks." Advances in Artificial Intelligence 2013

(2013): 6.

Ajerla, Dharmitha, Sazia Mahfuz, and Farhana Zulkernine.

"A real-time patient monitoring framework for fall

detection." Wireless Communications and Mobile

Computing 2019 (2019).

Yoo, SunGil, and Dongik Oh. "An artificial neural

network–based fall detection." International Journal of

Engineering Business Management 10 (2018):

1847979018787905.

Janku, Peter, Zuzana Kominkova Oplatkova, Tomas Dulik,

Petr Snopek, and Jiri Liba. "Fire detection in video

stream by using simple artificial neural network." In

MENDEL, vol. 24, no. 2, pp. 55-60. 2018

Kajan, S., D. Pernecký, and J. Goga. "APPLICATION OF

NEURAL NETWORK IN MEDICAL

DIAGNOSTICS." SK: University of Technology in

Bratislava (2014).

Shahid, Nida, Tim Rappon, and Whitney Berta.

"Applications of artificial neural networks in health

care organizational decision-making: A scoping

review." PloS one 14, no. 2 (2019): e0212356.

Amato, Filippo, Alberto López, Eladia María Peña-

Méndez, Petr Vaňhara, Aleš Hampl, and Josef Havel.

"Artificial neural networks in medical diagnosis."

(2013): 47-58.

Mehr, Homay Danaei, Huseyin Polat, and Aydin Cetin.

"Resident activity recognition in smart homes by using

artificial neural networks." In 2016 4th international

istanbul smart grid congress and fair (ICSG), pp. 1-5.

IEEE, 2016.

World Health Organization. Active ageing: A policy

framework. No. WHO/NMH/NPH/02.8. Geneva:

World Health Organization, 2002.

United Nations, Population Division. "World Population

Ageing 2007." (2007).

Fever Status Detection using Artiﬁcial Neuron Network

781