CAMISA : An AI Solution for COVID-19

Bhanuteja G
Electronics and Communication
MVJ College of Engineering Kaustubh Anand Kandi Srinidhi K
Bangalore, India Micro Electronics Design Facility Electronics and Communication
bhanuteja.g@mvjce.edu.in URSC, ISRO MVJ College of Engineering
Bangalore, India Bangalore, India
Paulson Premsingh S kaustubh@ursc.gov.in srinidhikrao05@gmail.com
Electronics and Communication
MVJ College of Engineering Dikshitha R Anil Kumar R
Bangalore, India Electronics and Communication Electronics and Communication
paulsonpremsingh7@gmail.com MVJ College of Engineering MVJ College of Engineering
Bangalore, India Bangalore, India
dikshitharavisha@gmail.com anilkumarr0475@gmail.com
Abstract—The COVID-19 pandemic has created an unpar- In many countries the vaccination drives are happening
alleled need for remote patient monitoring and has primarily at a slow pace, this exposes the majority of the population to
impacted the world as the mortality rate has increased rapidly. COVID-19. The availability of beds has also narrowed
As long as coronavirus exists, mutations of the virus continue down in hospitals, and the health status of COVID-19 in-
to happen, which also insists on the need for remote monitor- fected patients are not remotely monitored as expected. The
ing. Healthcare sectors require the help of many new technolo- need for patient monitoring which is both easy to use and
gies such as IoT, Artificial Intelligence, Neural Networks, and also produces accurate results, is a solution that can dissolve
sensor technology which can play an important role. The pro-
the problem. We try to provide a solution that solves this is-
posed system predicts the COVID-19 symptoms in a patient
with the integration of sensor technology and AI. This system is
sue which is very inexpensive for measuring the health sta-
effective in solving the crisis. It includes a shirt and a mask tus of a patient. We aim to produce an AI-based platform for
measuring the heart rate, blood oxygen level, and respiration monitoring the patient who has been infected by the coron-
avirus. The integration of the AI model on wearable devices
solves the problem to some extent.
This project was supported by Karnataka State Council for Science
As coronavirus was recognised only in the year 2019, the
and Technology, under "Student Project Programme - 44th Series”, with
Reference Number 44S_BE_4342 information available on the symptoms was very less. Con-
sidering the above situation, the previous work on this prob-
lem is highlighted where algorithms used for detecting
rate. In addition to this is the predictable AI model, where the
symptoms predict whether the patient is COVID-19 positive or
COVID-19 such as LSTM in [1] (Long Short Term Mem-
not. ory) and Fuzzy rule-based predictions, are having smaller
datasets. In order to overcome this, we provide a solution
Keywords—COVID-19, IoT, Contactless Health Monitoring, that evaluates the symptoms experienced by an individual,
Neural Network, Sensor Technology, Lilypad Arduino, Pulse by using a relatively larger dataset. We perform algorithms
Oximeter like backpropagation which is a simple technique and is fast
in execution. This is an important aspect in resolving the
I. INTRODUCTION COVID-19 pandemic by integrating the neural network
Coronavirus has caused much havoc around the world. model with the hardware that consists of a shirt along with a
As on May 20, 2021, there have been 165,565,058 con- nebulizer.
firmed cases, resulting in 3,434,004 deaths and disturbing When the patient uses our proposed system during the
life all around the world. The losses are exacerbating day by incubation period, the data obtained from the sensors is sent
day. With very little research done globally on the pandemic to the user-friendly application for real-time monitoring.
since SARS COV-2, there is no cure and less vaccination This also gives alert notifications whenever the patient’s
until recent days. Controlling the growth by early testing in- health is at risk to hospitals and guardians.
dividuals and quarantining the coronavirus-affected individ-
uals is the only implicit solution against the infectious Section II gives us an overview of the previous work,
COVID-19. However the ease to proceed with this strategy Section III contains the detailed discussion of our proposed
in order to control the deadly virus is not viable. There is a system, and Section IV gives the result and compares the so-
high risk of spread of the virus in crowded places amongst lution with the previous work carried out. Finally, Section V
people with low immunity is still a major concern. The need provides the conclusion of this project.
for testing this virus in an early stage is still a key differen-
tiator in most of the countries in order to reduce the pan- II. LITERATURE SURVEY
demic curve. As per the previous research for health monitoring sys-
tems, the technology used changes day by day to be more
accurate and use advanced methodologies. The IoT has been
a helping hand in achieving this goal today with many pa-
tients being monitored remotely in real-time. Continuous
care is given to the patients as their well-wishers are also
aware of their health status.
 Maneesh Gupta and Hana Qudsi [2] share an important
aspect from their perspective where the thermistor is used to
measure the breathing rate as COVID-19 affected patients
are targeted with the damage of lungs. This device measures
an individual’s respiration rate by detecting changes in tem-
perature which is mounted on the base of the face mask.
However, a delay was observed as the main drawback and is
being solved by advancing the hardware used.
AI 4 COVID-19 [3] by Ali Imran, throws light on the
fact that AI can solve the problem of COVID-19 and focus
on AI over the healthcare system to predict the contagious
disease using various algorithms. The ResNet-100 Convolu-
tional Neural Network, a deep learning technique together
with a Logistic Regression classifier, is employed to spot the
coronavirus pandemic rapidly. With the help of AI, we can
categorize a person’s health situation as having few symp- Fig. 1. Block diagram of Shirt and Mask
toms of COVID-19 will be under continuous monitoring.
· Here we use the MAX30100 Pulse Oximeter sensing
Dr.R.Poovendran, in his paper [4], has mentioned cost- device. It’s a measuring device that obtains its read-
effective hardware components that are use to detect heart ings from received intensity of light which is later
rate, breathing rate and SpO2, and informs the guardian converted to electrical signals. The LEDs are used to
about the patient's situation through alert or alarm call-out get the data by placing the sensor on the fingertip.
system. The paper mentions that the patient uses a headset The sensor is embedded on the shirt circuit, con-
with the mask for determining the breathing rate and heart nected by an elongated cable that reaches the tip of
rate using a sensor. It has a user-friendly application to col- the finger measuring accurate values.
lect the data, design all these wearables with the integration · The STM32 Blue Pill is an ARM Cortex M3 Micro-
of an application where the alarm or alert system can save controller. The STM board operates on a 3.3V power
patients immediately, even in a remote location. supply and uses a 32-bit processor. With the thermis-
Vishal Varun [6] in his paper mentions that measuring tor attached, a code was built for measuring the respi-
respiratory rate as one of the most tedious tasks. It requires ration rate.
utmost accuracy in order to prove that the patient is suffer- · The Ublox NEO6M GPS module used in this project
ing from breathlessness or not. The calculation of baromet- determines its location and we obtain the output, that
ric pressure and the usage of thermistor provides an accu- is the latitude and longitude of its position. It exhibits
racy of 93% in this research. However the author faces is- high sensitivity when used indoors. The battery in the
sues with the data communication from the mask to the de- GPS module is used for power backup, and the EEP-
vice like mobile/tablet etc. The usage of signal processing ROM present in the GPS module stores the config-
algorithms is seen here monitoring lung functions. ured settings providing the needful results.
· There is a patch antenna present in the module which
Muhammad E. H. Chowdhury [7] has mentioned solving has a sensitivity of -161dBm. A U.FL cable is used to
the problem of COVID-19 through the datasets of SARS us- connect the antenna and the module. This allows
ing CNN model and chest X-rays through image. Since the great flexibility when the GPS module is mounted
dataset collected was only about 319 CheXNet dataset, it onto our shirt.
wasn’t accurate enough to detect pneumonia in humans. The · The temperature sensor that is embedded in the shirt
pre-dataset being trained is not sufficient to provide high ac- is the DS18B20 and has a 1-Wire bus used for com-
curacy rate as the image dataset available from the period of munication with the microcontroller. It operates over
pneumonia is ineffective in detecting coronavirus. the temperature range of -55°C to +125°C and accu-
racy is ±0.5°C for the range from -10°C to +85°C.
III. PROPOSED ARCHITECTURE · For heart rate and SpO2 (Blood Oxygen level) which
A. Hardware is the major parameter for COVID-19, if the SpO2
level drops below 90%, it will create an alert system
For the hardware which is the combination of a shirt and through WiFi module or by sending an SMS to the
a mask the following components are considered and are patient’s guardian and hospital.
connected to give us the results of pulse, blood oxygen
level, breathing rate, real time location and temperature. B. Software
The proposed block diagram is illustrated in Fig.1. The software used for programming the components of
· The microcontroller used is Lilypad Arduino, which the shirt and mask is Arduino IDE. The ease with which an
is based on the ATmega328V. It’s designed for e- Arduino can obtain sensor values is one of the features that
textiles, thereby it can be sewn on fabric with other makes it very useful. ThingSpeak is used to aggregate, visu-
components using a conductive thread. The LilyPad alize, and analyze live data streams, which are sent to the
Arduino can be programmed with the Arduino IDE. app over the Internet. PlatformIO is another software used,
· The NodeMCU is used to collect the sensory outputs which is an open-source ecosystem for IoT development.
from the microcontrollers and send it to the App over C. Camisa - Shirt
the Internet. It runs on ESP8266 firmware having
128KB RAM and 4MB of flash memory to store the With the ongoing swath of the COVID-19 pandemic sit-
data. uation, health care monitoring has become an essential part
of human lives. A few health monitoring systems including
wearable devices like watches, waist belt, shirt, mask, etc.
are effective when used for remote applications. One such
confirmation for real-time health monitoring [5] is the use of
a shirt that is embedded with sensors. The circuit of shirt is
shown in Fig. 2. This gives us accurate results of patients
when they are in their quarantine period and get alerts when
the sensor values reach a threshold level indicating at-risk
condition.

Fig. 3. Circuit for Mask

The STM32 is an ARM Cortex M3 Microcontroller that

is programmed with a thermistor to collect the breathing pat-
tern of an individual. We try to analyze the number of ‘no
Fig. 2. Shirt Circuit which can be sewn on textile. breaths' conditions, using the parameter shown in Table II.
After consecutive ‘no breaths’, an alert is notified to the per-
The need to measure blood oxygen saturation decreasing son monitoring.
during the incubation period gives way to the pulse oxime-
ter. Based on the light signals which are reflected from the TABLE II. THRESHOLD VALUES FOR MASK
blood cells, it converts them to electrical signals. The output
is mathematically processed to measure oxygen saturation Ideal versus Risk Values for Alert Systems
level and heart rate. The patient's temperature is recorded by S. No.
Specified
the temperature sensor. Parameters At Risk
Value

If the threshold is exceeded, the real-time location of the Breathing Rate (respirations
1. 12 - 20 > 20 a
patient is shared with the hospital to rush the patient for im- per min)
mediate intensive care. The threshold values of sensors in
the shirt is given in Table I. II.

TABLE I. THRESHOLD VALUES FOR SHIRT a. Varies between individuals

Ideal versus Risk Values for Alert Systems

S. No.
Parameters Ideal At-Risk

1. SpO2 Level (%) 94 -100 < 91

2. Heart Rate (bpm) 70-100 >110 a

3. Temperature ( in °F) 97 - 99 >100

I.
a. Irregularity is experienced, varies among individuals

D. Camisa - Mask
The use of face masks and shields has reduced the risk of
virus transmission. Adding onto this primary solution, our
focus is to develop a smart nebulizer, as illustrated in Fig. 3,
which can be worn by the patient during their quarantine pe-
riod. This is helpful in order to continuously monitor their
breathing pattern and temperature using the NTC thermistor Fig. 4. Transfer of data to the application
that is attached inside the nebulizer [6].
After the data is collected from the shirt and the mask,
the data is visualized on the application for remote monitor-
ing. The transfer of data to the application is illustrated in
Fig. 4.
E. Neural Network Implementation
An ANN model is used to design and compute mathe-
matical algorithms resembling how the brain reacts to sen-
sory inputs. The brain consists of neurons that are intercon- and so on. Thereby giving us maximum of 20 different
nected to form a huge network. Neural Network, in general, parameters to predict the condition of a single user. The
give us various advantages in numerous fields, one such ad- parameters are set giving the input layer size over
vantage is the healthcare system, which uses the neural net- 200,000 inputs. We labelled the data as per the require-
work in prediction and spread of disease. A neural network ment and get the final raw data for further tuning.
is made up of millions of artificial neurons and is intercon-
nected by nodes, therefore called a processing unit.
Initially, a mathematical model of all the neurons and the
interconnections between them are created. Then the inputs
are provided to the network, and the neurons convolute with
the inputs systematically, to generate an output, which is fed
into the next level of neurons, as shown in Fig. 5. The
process repeats itself until the output is reached, in the pre-
defined hidden layers.

Fig. 7. Formatted Input Dataset File

c) Database formatting to make it compatible with AI:

The raw input data is made compatible for AI by prepro-
cessing. The raw data is processed and fed to the 3 layer
Fig. 5. Neural Network Architecture feed forward network, where forward propagation is per-
formed below as in Fig. 8.
The ANN model can also be used to estimate the confir-
mation of the coronavirus through the available datasets [7].
In order to limit the propagation speed and the propagation
power of the coronavirus, it is important to predict and iden-
tify the infected patients and isolate them. The flowchart of
neural network is explained in Fig. 6. Below are the steps
carried out for the Predictive AI model using C language
and Python.
a) Collection of data from various sources: Since the
virus exhibits different symptoms in different countries
due to the mutations, the dataset is obtained from a stan-
dardized database. The network is tuned by training it on
this dataset built on health conditions of 300,000 people
from different geographic locations.
Fig. 8. A 3 Layer Feed Forward Network

d) Theoretical modelling of ANN: For performing the

theoretical analysis, the arrays are set up by assigning
random weights. This is a 2D array matrix where one col-
umn represents a single node, thus obtaining a 10x20 ma-
trix. This final matrix obtained is used to train the model.
Equations (1) is the input vector fed into the network.
Equations (2) and (4) calculate the weighted sum of in-
puts of the respective layers which is subsequently used
to measure the activation function given by (3) and (5).

(1)
x i=a i ,i ∈1 , 2 ,3 (1)
Fig. 6. Flowchart for Neural Network
(2) (1) (1)
z =W x +b (2)
b) Database formatting and creation of labelled data:
The data obtained is formatted for inputs with 20 different
a(2)=f (z (2) ) (3)
parameters as shown in Fig. 7, where each bit in the every
input corresponds to a symptom experienced by an indi- (3) (2 ) (2) (2)
vidual like cold, fever, cough, breathlessness, diarrhoea, z =W a +b (4)
 (3) (3 )
a =f (z ) (5)

Initially, a loop is started to run through the data used for

training in each layer. In each iteration the order in which
the training data runs through is randomized, to ensure that
local minima do not have a union.
The data is fed through the network in order to calculate
the activation function of the hidden layers and output
layer’s nodes (in our case, sigmoid function as in (6))[9].
1
σ (x)=
1+e
−x (6)

Later on updating the associated weights and comparing

the output after multiple iterations, on how well the neural
network performed as to the given training sample. This is Fig. 10. Error Plot of the trained model
referred to as the cost function, given by (7) and illustrated
in Fig. 9. f) Training the model and validation: The model and
algorithms are trained on 60% of the dataset where the
1 n dataset is split. The remaining 40% of the dataset is used
Cost Function (J)¿ n Σ i=0 ¿ for validation, to prove the predictability of the algorithm.
(7) g) Multiple iterations and parameter tuning to obtain
the least error: This is the final step of training the
model. The output thus obtained is an error function that
is obtained after every sample is trained. The Fig. 11. il-
lustrates the output file generated, showing 1-bit binary
output for every combination of input parameters. Zeroes
and ones represent absence and presence of COVID.

Fig. 9. Cost Function

The backpropagation algorithm checks for the least

value of the error function in the weights vector. On back-
propagating the error function to the hidden layer, the algo-
rithm tunes the network in accordance with the error rate
[10]. This calculates the gradient of the obtained loss func-
tion with respect to weights and thus minimizes the error.
∂J
Gradient Descent ¿ ∂ w (8)

The gradient descent for a given algorithm is an opti-

mised algorithm used to find the parameters that reduce the
cost function as shown above in (8). Fig. 11. Output file generated showing the presence of COVID-19
e) Parameter tuning: Using the obtained values, we
Neural network gives us a total of (20x10) + (5x1) = 205
perform parameter tuning in order to find the error func- trainable parameters in the model. This is almost similar to a
tions. It is observed that the error functions are in a de- 205th-degree equation since it is highly impossible to fit in
creasing trend as seen in Fig.10. The obtained outputs of such a high degree equation using mathematical methods or
the error function are compared as the values of various any conventional methods. Neural networks giving a very
parameters such as learning rates, i.e, alpha are varied efficient solution to this problem is a good tool to model
(9). such diverse datasets.

∂J To conclude this model we integrate the output obtained

dx = alpha ¿∨ ∂ w ∨¿ (9) from the neural network, which gives us information about
whether a person has COVID-19 or not based on the symp-
toms.
Our project shows that it is possible to acquire a qual-
ity model for the prediction of disease using AI, with inputs
as symptoms experienced by an individual. The predictions
prove that AI models can also solve problems in prediction
of any disease. The application of AI methods should also
be modelled keeping in mind the present and future spread
of diseases and in an attempt to predict and prevent the im-
pact of such infections. Our future plan widely leans towards
making the model available worldwide solving the problem
of COVID-19.
F. User Friendly Application
The health monitoring app combines the output from
sensors and modules, then visualizes the data received. After
the predictive model is trained, it is deployed on the applica-
tion. The app consists of two main section:
a) Self Assessment: It has a questionnaire filled using
toggle switches that record users' responses, which corre-
spond to the COVID-19 symptoms experienced by the
user. Based on which the app predicts an output of
whether the patient is COVID-19 positive or not. Fig. 12
illustrates the questionnaire.

Fig. 13. Health Monitoring on the App

IV. RESULT COMPARISON

The comparison of key characteristics of prior work and
proposed work is shown in Table III. As the parameters con-
sidered vary from each other, our work tries to overcome the
cons and provide accurate results.

TABLE III. RESULT COMPARISON

Reference Reference Reference Proposed

Parameters
[5] [2] [3] Work

Heart Rate,
Temperature,
Heart Rate, Breathing ANN Net-
Application Blood Oxygen
Temperature Rate work
level, Breath-
ing Rate

Heart Rate Pulse Oxime-

Main Sensor, Neural ter, Thermis-
Thermistor
Integrant Temperature Network tor, Tempera-
Sensor ture Sensor
Fig.12. Self Assessment Questionnaire
Sensor Tech-
Comparison nology & Pre-
b) Real-Time Health Monitoring: The health parame- Working Sensor
of Relative
Prediction
dictive Analy-
ters obtained from each of the sensors are visualized on Principle Technology by LSTM
Voltages sis using Neu-
the app using the Thingspeak IoT Platform. The app ral Network
alerts the guardians and shares the real-time location of
Microcon- Lilypad
the patient when the patient is at risk. troller
Atmega328 TI-MSP430 -
Arduino
Fig. 13 shows heart rate and blood oxygen saturation in
real time. Similarly the other parameters, breath rate, tem-
perature and location of the patient is also visualized on the V. CONCLUSION
application.
In today’s world due to rising threats of COVID-19, the
need to solve the problem with the help of different tech-
nologies is essential. The solution provided above satisfies
the need and also can be further considered for advance-
ments.
Artificial intelligence is a very important and promising
tool for detecting early COVID-19 infections and monitor-
ing the condition of infected ones remotely. The advent of
useful algorithms greatly improves the sequence of process-
ing and decision-making. The participation of the Internet of
Things that integrates all the above technologies provides a
good result. In this paper, we’ve introduced a low-power
IoT wearable devices for the COVID-19 application called
Camisa (Contactless Patient Monitoring System). We've
identified the foremost components of the proposed system
and explained its implementation details. Performance rat-
ings indicate that wearable shirts and masks are an inexpen-
sive device. Future scope to this area will include ensuring
data access and security to shield the privacy of hospitals
and patients, further as enabling voice recognition for the
proposed wearable device which can be more advanced.
REFERENCES
1. Shahid Farah Anleela Zameer and Muhammad Muneeb “Predictions
for COVID-19 with Deep Learning Models of LSTM GRU and B1-
LSTM” Chaos Solitions &Fractals 140(2020)
2. Gupta, Maneesh, and Hana Qudsi. "Low-cost, thermistor based respi-
ration monitor." 2013 39th Annual Northeast Bioengineering Confer-
ence. IEEE, 2013.
3. Imran, Ali, et al. "AI 4COVID-19: AI enabled preliminary diagnosis
for COVID-19 from cough samples via an app.”
4. Nachiar, Cecil C., et al. "Design of Cost-effective Wearable Sensors
with Integrated Health Monitoring System." 2020 Fourth International
Conference on I-SMAC (IoT in Respiratory Monitoring System using
thermistor 1 Gagandeep kour , 2Mohammad
Rouman ,Geetha.M31,2UG Students, Assistant Professor Department
of Biomedical Engineering BIHAR, BIST, Bharath UniversityChen-
nai- 6000073.
5. Aziz, Kahtan, et al. "Smart real-time healthcare monitoring and track-
ing system using GSM/GPS technologies." 2016 3rd MEC Interna-
tional Conference on Big Data and Smart City (ICBC). IEEE
6. Kalavakonda, Rohan Reddy, et al. "A Smart Mask for Active Defense
Against Coronaviruses and Other Airborne Pathogens." IEEE Con-
sumer Electronics Magazine (2020).
7. Huang, Chiou-Jye, et al. "Multiple-input deep convolutional neural
network model for covid-19 forecasting in china." MedRxiv (2020).
8. Che, Zhen-Guo, Tzu-An Chiang, and Zhen-Hua Che. "Feed-forward
neural networks training: a comparison between genetic algorithm and
back-propagation learning algorithm." International Journal of Inno-
vative Computing Information and Control 7.10 (2011): 5839-5850.
9. Tsai, Chang-Hung, et al. "A hardware-efficient sigmoid function with
adjustable precision for a neural network system." IEEE Transactions
on Circuits and Systems II: Express Briefs 62.11 (2015): 1073-1077
10. Erb, Randall J. "Introduction to backpropagation neural network com-
putation." Pharmaceutical research 10.2 (1993): 165-170.