This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

A Machine Learning Approach using Statistical

Models for Early Detection of Cardiac Arrest in
Newborn Babies in the Cardiac Intensive Care
Unit
Ketan Gupta1, Nasmin Jiwani2, Giovanni Pau3*, Mohammad Alibakhshikenari4
1,2
Department of Information Technology, University of The Cumberlands, Williamsburg, KY, USA
3
Faculty of Engineering and Architecture, Kore University of Enna, 94100 Enna, Italy
4
Department of Signal Theory and Communications, Universidad Carlos III de Madrid, 28911 Leganés, Madrid, Spain

Corresponding author: Giovanni Pau, Kore University of Enna, 94100 Enna, Italy (e-mail: [email protected]).

Funding Statement: This research was supported by the Tejankar Healthcare and Medical Research Institute Pvt. Ltd-
Department of Neonatology (Ref by THMRI/MD/22-23/170)

ABSTRACT Cardiac arrest in newborn babies is an alarming yet typical medical emergency. Early
detection is critical for providing these babies with the best care and treatment. Recent research has focused
on identifying the potential indicators and biomarkers of cardiac arrest in newborn babies and developing
accurate and efficient diagnostic tools for early detection. An array of imaging techniques, such as
echocardiography and computed tomography may help provide early detection of cardiac arrest. This
research aims to develop a Cardiac Machine Learning model (CMLM) using statistical models for the early
detection of cardiac arrest in newborn babies in the Cardiac Intensive Care Unit (CICU). The cardiac arrest
events were identified using a combination of the neonate’s physiological parameters. Statistical modeling
techniques, such as logistic regression and support vector machines, were used to construct predictive
models for cardiac arrest. The proposed model will be used in the CICU to enable early detection of cardiac
arrest in newborn babies. In a training (Tr) comparison region, the proposed CMLA reached 0.912 delta-p
value, 0.894 False discovery rate (FDR) value, 0.076 False omission rate (FOR) value, 0.859 prevalence
threshold value and 0.842 CSI value. In a testing (Ts) comparison region, the proposed CMLA reached
0.896 delta-p values, 0.878 FDR value, 0.061 FOR value, 0.844 prevalence threshold values and 0.827 CSI
value. It will help reduce the mortality and morbidity of newborn babies due to cardiac arrest in the CICU.

INDEX TERMS cardiac arrest, newborn, early detection, intensive care unit, statistical modeling

I. INTRODUCTION signs are present, it is essential to seek medical attention

Cardiac arrest in newborn babies is a devastating event immediately. Risk factors that may increase the likelihood
that can lead to severe complications and death. Early of cardiac arrest in newborn babies include low birth
detection of this condition is critical to provide the best care weight, a family history of cardiac arrest, preterm birth, a
for these infants and ensure their long-term health. In order difficult delivery, or a mother with a history of high blood
to ensure the early detection of cardiac arrest in newborn pressure during pregnancy [2]. A baby's medical history
babies, it is essential to understand the signs and symptoms should also be evaluated for any potential risks. In order to
associated with this condition and the risk factors that may ensure early detection of cardiac arrest in newborn babies,
put a baby at an increased risk of cardiac arrest [1]. The regular monitoring of the baby's heart rate and respiratory
most common signs and symptoms of cardiac arrest in rate is essential. It can be done through pulse oximetry, a
newborn babies are a rapid heart rate and difficulty non-invasive, painless procedure that measures the amount
breathing. Other signs that may indicate a baby is in cardiac of oxygen in the baby's blood [3]. Additionally,
arrest include a bluish tinge to the baby's skin, auscultation, or listening to the baby's heart rate and
unresponsiveness, or decreased movement. If any of these breathing with a stethoscope, can also help to detect any

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

irregularities in the baby's heart rate or breathing. Early calculate the odds ratio for each risk factor, which indicates
detection of cardiac arrest in newborn babies is vital to how much more likely an infant is to experience an arrest if
provide the best care for these infants and ensure their long- they have a particular risk factor. Another effective model
term health. By understanding the signs and symptoms of for the early detection of cardiac arrest in newborns is a
this condition and being aware of the risk factors that may support vector machine (SVM) [15]. This model type is
put a baby at an increased risk of cardiac arrest, parents and well-suited for binary classification tasks, such as
medical professionals can work together to ensure the best classifying an infant as either healthy or having experienced
possible outcomes for these babies [4]. The early detection a cardiac arrest. It can also be used to identify important
of cardiac arrest in newborn babies can be achieved using risk factors associated with cardiac arrest and predict the
Statistical Models. Statistical models are mathematical likelihood of an infant experiencing an arrest [16]. Finally,
techniques used to analyze and draw conclusions from data. artificial neural networks (ANNs) can also detect cardiac
These models are powerful tools in the medical field, as arrest in newborns. ANNs are powerful machine-learning
they can help predict, diagnose, and treat certain diseases models that can learn complex patterns from data. These
and conditions [5]. One example of a statistical model used models can be used to identify risk factors associated with
for the early detection of cardiac arrest in newborn babies is cardiac arrest and predict the likelihood of an infant
the Logistic Regression model. This model uses data experiencing an arrest. Logistic regression, support vector
collected from the baby's medical history, such as birth machines, and artificial neural networks are all effective
weight, gestational age, and gender, to create a predictive models for the early detection of cardiac arrest in newborns.
model to determine the likelihood of cardiac arrest [6]. This These models can be used to identify the most critical risk
model can help doctors identify those babies at risk and can factors associated with the condition and predict the
help them decide whether to treat the baby with medication likelihood of an infant experiencing an arrest. Therefore,
or perform surgery to correct the issue. Another model used these statistical models should be used to improve
for the early detection of cardiac arrest in newborn babies is newborns' early detection and intervention of cardiac arrest
the Naive Bayes model [7]. This model uses a probabilistic [17]. Machine learning is increasingly used to predict and
approach to analyze data and identify patterns to make detect cardiac arrest in newborn babies. Cardiac arrest is a
predictions. The model can identify high-risk babies and life-threatening condition in which the heart suddenly stops
help doctors determine the best course of action to take [8]. beating, and blood flow to the brain and other organs stops.
The Support Vector Machine model is another statistical It can lead to permanent brain damage or death. Due to the
model used for the early detection of cardiac arrest in complexity of the condition, early detection of cardiac
newborn babies. This model uses data collected from the arrest in newborns has been difficult. However, machine
baby's medical history and other sources to create a learning is changing that [18]. Machine learning algorithms
predictive model that can determine the likelihood of analyze large amounts of complex data, such as patient
cardiac arrest [9]. This model can identify those babies at medical histories, vital signs, and other physiological data.
risk and help doctors decide on the best course of treatment. The algorithms can detect patterns in the data indicative of
Statistical models are powerful tools that can be used for cardiac arrest and alert medical personnel. For example,
the early detection of cardiac arrest in newborn babies [10]. one study used machine learning to detect signs of cardiac
These models can help doctors identify those at risk so that arrest in newborns by analyzing their heart rates, breathing
they can provide the best possible treatment for the baby. patterns, and other vital signs. The algorithm detected signs
Furthermore, these models can help doctors determine the of cardiac arrest up to eight hours before conventional
best course of action to take in order to prevent or reduce methods. It could significantly improve the chances of
the likelihood of cardiac arrest [11]. survival for newborns and reduce the damage caused by the
condition. In addition, machine learning is used to predict
Cardiac arrest in newborns is a life-threatening medical newborns' risk of cardiac arrest. By analyzing large
condition that requires immediate medical attention. Early amounts of patient data, machine learning algorithms can
detection and intervention can improve the outcomes of identify risk factors associated with the condition. It can
these infants and reduce mortality rates. Statistical models help medical personnel identify newborns at an increased
are powerful tools that can be used to identify risk factors risk of cardiac arrest to receive the care they need. The
and predict the likelihood of cardiac arrest [12]. Logistic machine learning is revolutionizing the early detection of
regression is one of the best statistical models for the early cardiac arrest in newborns [19]. By analyzing large
detection of cardiac arrest in newborns. This model allows amounts of complex data, machine learning algorithms can
researchers to quantify the relationship between risk factors detect signs of cardiac arrest and identify newborns at an
and the probability of experiencing an arrest [13]. It can be increased risk of the condition. This technology could save
used to identify the most critical factors associated with lives and reduce the damage caused by cardiac arrest in
cardiac arrests, such as gender, gestational age, and birth newborns. The critical contribution of machine learning
weight [14]. Logistic regression can also be used to models used for the Early Detection of Cardiac Arrest in

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

Newborn Babies is that these models can detect subtle with acute decompensated heart failure include increased
changes in vital signs such as heart rate, respiratory rate, length of stay, healthcare utilization, mortality, and decreased
and oxygen saturation that are difficult to detect with the quality of life. The functional decline may also lead to higher
naked eye. This early detection can help to identify rates of re-hospitalization, as well as an increased risk of
newborns at risk of cardiac arrest and allow for timely institutionalization. Additionally, the functional decline may
intervention and treatment [20]. Additionally, machine lead to an increased risk of falls and delirium due to
learning models can be used to analyze patient data to decreased mobility and activity levels. Fonarow, G. C., et
provide personalized advice and care to patients, enabling al.[23] has discussed the Risk stratification for in-hospital
better long-term management of their condition. The mortality in acutely decompensated heart failure determines
following are the critical contribution of the proposed which patients are at higher risk of dying while in the
research works. hospital. It is done by using classification and regression tree
• Automated and accurately detected critical signs analysis. Classification and regression tree analysis is a type
associated with cardiac arrest in newborn babies. of predictive analytics that uses trees to classify and predict
• Ability to recognize subtle changes in the baby's outcomes. The trees are nodes representing various
vital signs that can indicate potential cardiac arrest. conditions, characteristics, or features associated with the
• Ability to identify high-risk babies likely to suffer outcome. Using a combination of these nodes, the model can
from cardiac arrest. determine the likelihood of a particular outcome occurring.
• Early detection of cardiac arrest, enabling timely The model can then be used to identify patients at a higher
interventions that can improve the outcome. risk of in-hospital mortality and to guide the treatment of the
• Reduction in the time and cost associated with patient.
traditional monitoring methods. Gaies, M. G., et al.[24] has discussed the Vasoactive–
• Improved patient outcomes due to early diagnosis inotropic score (VIS) is designed to predict morbidity and
and treatment of cardiac arrest. mortality in infants after cardiopulmonary bypass (CPB). The
The remaining chapters of the manuscript have VIS is calculated from the levels of vasoactive and inotropic
organized as the following. Chapter 2 provides information drugs administered to the infant during and after CPB. These
about the earlier works related to the research. Chapter 3 drugs are used to regulate the patient's blood pressure and
provides the construction of the proposed model. It includes heart rate. The VIS is believed to accurately predict post-
the proposed algorithm and flow chart. Chapter 4 details the CPB morbidity and mortality because it reflects the degree of
analytical discussion, and Chapter 5 shows the comparative hemodynamic instability in the infant. Higher VIS scores
analysis between the proposed and existing models. indicate greater hemodynamic instability and, therefore, a
Chapter 6 expresses the results and discussion, and finally, greater risk of morbidity and mortality. Studies have found
chapter 7 expresses the conclusion and future scope of the that higher VIS scores are associated with increased
proposed research. mortality, extended hospital stays, and an increased need for
vasopressor and inotropic support. The VIS is a significant
II. RELATED WORKS predictor of outcome after CPB and can help clinicians
Carlisle, M. A., et al.[21] has discussed the Heart failure is identify infants who may require closer monitoring and more
when the heart cannot pump enough blood to the rest of the aggressive management. Shah, S. J., et al.[25] has discussed
body. Various conditions, including high blood pressure, the Phenomapping is a novel classification system for heart
coronary artery disease, and diabetes, can cause it. Atrial failure with preserved ejection fraction (HFpEF). It is based
fibrillation is an arrhythmia (irregular heartbeat) in which the on the analysis of phenotypic characteristics, such as
heart’s upper chambers (atria) beat rapidly and irregularly. It demographics, clinical profile, laboratory values,
can cause a decrease in the amount of blood pumped to the electrocardiographic findings, echocardiography findings,
rest of the body, leading to symptoms such as shortness of and biomarkers. The goal of Phenomapping is to provide a
breath and fatigue. Atrial fibrillation is a common cause of more comprehensive and meaningful classification system
heart failure. Heart failure and atrial fibrillation treatment for HFpEF that is based on the distinct phenotypes of the
usually involve medications to control the heart rate and disease. This classification system will enable clinicians to
rhythm, lifestyle changes, and sometimes surgery to repair or more accurately diagnose and stratify patients with HFpEF,
replace the heart. Yaku, H., et al.[22] has discussed the Risk leading to better management and improved outcomes. The
factors for functional decline during hospitalization in very Phenomapping system also provides a platform for further
old patients with acute decompensated heart failure include research into the underlying path physiology of HFpEF,
age, gender, co-morbidities, and frailty. In addition, complex allowing for a better understanding of the disease and the
medical problems, the need for aggressive treatments, and potential for improved treatments. Lee, D. S., et al.[26] has
the presence of cognitive impairment may increase the risk of discussed the Heart failure with preserved or reduced ejection
functional decline. Clinical outcomes associated with a fraction (HFPEF or HFREF) is a form of heart failure in
functional decline during hospitalization in very old patients which the heart's ability to pump blood is impaired, but the

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

amount of blood pumped from the heart with each beat using metrics that favor specific outcomes or data sets that do
(ejection fraction) is either normal or reduced. The not reflect the full range of potential outcomes. Finally, there
underlying cause of this type of heart failure is poorly may be biases in how the algorithm is deployed, such as
understood, but various disease pathologies and risk factors using the algorithm to make decisions that favor specific
have been associated with it. Disease pathologies, such as outcomes.
coronary artery disease, hypertension, and diabetes, can lead Moor, M., et al.[30] has discussed the Sepsis is a life-
to HFPEF or HFREF. These conditions can impede blood threatening condition caused by an overwhelming immune
flow through the heart and its associated vessels, response to an infection, and it is one of the leading causes of
accumulating fluid in the lungs and other parts of the body. death in intensive care units (ICUs). Early detection and
Additionally, these diseases can cause damage to the heart intervention are essential for successful treatment, and
muscle and its inner lining, making it more difficult for the machine learning algorithms can help to identify sepsis
heart to pump blood efficiently. Risk factors for HFPEF or patients in the ICU at an earlier stage. To identify sepsis-
HFREF include advancing age, obesity, gender, smoking, associated patterns, machine learning algorithms can analyze
and alcohol consumption. People who are older, overweight, various data sources, such as patient vital signs, laboratory
and lead an unhealthy lifestyle are more likely to develop test results, and medical history. These algorithms can be
HFPEF or HFREF. Additionally, gender may play a role, as trained to identify sepsis patients earlier, allowing clinicians
women are more likely to develop HFPEF than men. to intervene before the patient's condition deteriorates. Using
Saeed, M., et al.[27] has discussed the Multiparameter machine learning algorithms in the ICU to predict and
Intelligent Monitoring in Intensive Care (MIMIC) is a system diagnose sepsis earlier can reduce mortality, morbidity, and
that uses artificial intelligence and machine learning length of stay. Furthermore, this technology can help to
algorithms to continually monitor a patient's health and vital reduce costs associated with sepsis-related complications and
signs in an intensive care unit (ICU) setting. MIMIC is improve the quality of care provided to patients. Deo, R. C.
designed to alert healthcare providers to changes in a et al.[31] has discussed the predictive model was then
patient's condition that may require medical intervention. It is validated by comparing its predictions to actual mortality
an effective tool for detecting subtle changes in a patient's rates among a separate group of patients. The model
condition that clinicians may find difficult to detect with only accurately predicted mortality among patients with heart
physical exams and laboratory tests. The system can monitor failure, indicating its effectiveness in identifying those at
vital signs, including heart rate, respiration rate, blood high risk of death. This study highlights the importance of
pressure, and temperature. Additionally, MIMIC can detect using predictive models to identify those at high risk of death
changes in a patient's oxygen saturation and provide alerts among patients hospitalized for heart failure. Such models
when abnormalities are detected. By combining data from can help healthcare professionals manage and treat patient’s
multiple sources and using advanced analytics, MIMIC can better, improving outcomes and patient safety.
provide early warnings and enable healthcare providers to Shankar K. et al. [33] have discussed deep synergic
make more informed decisions about a patient's care. Lee, D. learning as a powerful AI technique that allows machines to
S., et al.[28] has discussed the Predicting mortality among interpret data and create strategies for intelligent decision-
patients hospitalized for heart failure is an essential task for making. When applied to healthcare diagnosis for COVID-
healthcare professionals. It is essential to identify those at 19, this technology has the potential to identify the disease
higher risk of death so that they can receive more aggressive accurately and quickly. This approach can diagnose
treatment and better management of their condition. The individuals and populations in connected living and
study aimed to develop a clinical model to predict mortality imaginative city scenarios. The system applies layers to
among patients hospitalized for heart failure accurately. To detect patterns, identify trends, extract relevant information,
do that, researchers used data from an extensive database of and eventually find the best possible diagnosis and
management solutions. The machine can learn more
patient records to identify risk factors associated with
accurately and quickly by combining medical evidence with
mortality. They used various statistical methods to evaluate
other data sources, such as symptoms, environment,
the risk factors and develop a predictive model.
location, and lifestyle factors. Synergic deep learning can
Gianfrancesco, M. A., et al.[29] has discussed the Potential help to reduce diagnostic time, improve accuracy and
biases in machine learning algorithms using electronic health reliability, eliminate errors and delays, and identify insights
record (EHR) data can come from various sources. First, about the disease. It also helps healthcare providers make
there may be biases in the data due to sampling or coding better-informed choices in prescribing treatments and
errors. If the data used to train the algorithm does not further develop strategies and guidelines to prevent future
represent the population, the algorithm may be biased toward outbreaks. Hashash, O., et al.[34] has discussed MEC-
specific outcomes. Second, there may be biases in how the Based Energy-Aware Distributed Feature Extraction for
algorithm processes the data, such as favoring certain data mHealth Applications with Strict Latency Requirements, a
types or giving too much weight to certain variables. Third, technique in which feature extraction for mHealth
there may be biases in evaluating the algorithm, such as applications is distributed across multiple edge nodes using

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

Mobile Edge Computing (MEC) technology. MEC-based al.[38] has discussed that a machine learning early warning
feature extraction distributes a load of feature extraction score can significantly impact hospital mortality. This kind
among the edge nodes to reduce the energy consumed, of system would use data analysis to evaluate the health of
making the whole system energy-aware. Feature extraction a hospitalized patient and generate a score that reflects the
is made more efficient through data preprocessing, data patient's risk of death. Healthcare providers can then use
augmentation, ensemble training, and data cleansing this score to predict a patient's future health outcome,
techniques. Furthermore, the distributed feature extraction identify high-risk patients, and take appropriate action to
techniques ensure that the latency requirements are met, as reduce the mortality rate. By tracking vital signs and other
the feature extraction process is distributed among more physiological parameters and patient behavior, this type of
nodes, resulting in shorter latency for each node. system can provide valuable insight into patient health and
Chaudhary, P. et al.[35] has provided a motor imagery allow for early interventions to prevent more severe
classification using sparse nonnegative matrix factorization complications. By increasing the accuracy and scope of
(SNMF) and convolutional neural networks (CNNs) is a patient risk assessment, a machine learning early warning
technique used to classify motor imagery data acquired score can ultimately save more lives. Ahsan, M. M., et
from electroencephalography (EEG). The SNMF al.[39] has discussed a Machine Learning-based heart
decomposes the EEG signals into latent features that are disease diagnosis as a promising technique for identifying
then used as input to a CNN for classification. The patients with a high risk of developing cardiovascular
objective of this technique is to effectively capture the diseases or conditions. It uses the data collected from the
spatial and temporal patterns of motor imagery data while patient's medical history and other relevant sources to
maintaining interpretability. The technique is supervised in recommend treatments or preventive treatments. The
that the labels of the signal segments are known in advance. machine learning models are trained to recognize patterns
The SNMF is used to reduce the dimensionality of the data, and trends in the patient's data, such as the patient's body
making it more suitable for classification, while the CNNs weight and blood pressure readings. The models can
are used to learn an interpretable classification space. By identify patterns that indicate a high risk of developing
combining the two, a more efficient and effective motor heart disease and develop an alert system to alert the doctor.
imagery classification can be achieved. Bhatti, M. H., et By providing early detection and timely treatments,
al.[36] has discussed a Soft computing-based EEG machine learning-based heart disease diagnosis can help
classification by optimal feature selection and neural prevent the development of more severe complications.
networks, a technology that uses neural networks and Rooney, S. R., et al.[40] has discussed that Machine
evolutionary algorithms to classify brain activity in learning and high-frequency physiologic data can be used
electroencephalogram (EEG). The approach is based on to predict extubation failure in pediatric cardiac ICUs. This
optimizing feature selection and neural network structure to technique uses physiological variables captured at up to 10
find the most accurate model. It uses a combination of Hz to create a snapshot of a patient's health status. Then,
traditional machine learning algorithms and optimization machine learning algorithms are used to analyze the data,
techniques for feature selection, such as Genetic detect changes in the patient's condition, and determine the
Algorithms and Particle Swarm Optimization. Once the risk of extubation failure. It can help doctors identify those
optimal feature set and structure have been determined, a patients at risk for extubation failure and take measures to
feed-forward neural network is trained to recognize the prevent it. This technique can also be used to compare
patterns in the EEG signal and classify them. The network outcomes across different ICUs and can help inform
is trained using back-propagation, a standard gradient clinical decisions.
descent algorithm that measures the output error and adjusts
the weights accordingly. The result is an accurate, reliable From the above literature works, the following key issues
EEG classifier that can be used to detect and classify were identified. They are
various types of brain activities. • The complexity of data: Most of the existing
Luo, C., et al.[37] has discussed a machine learning- machine learning models must be able to accurately
based risk stratification tool for in-hospital mortality of interpret large and complex datasets to detect
intensive care unit patients with heart failure is a tool that cardiac arrest in newborn babies.
uses machine learning (ML) to identify patterns in patient • Data availability: Most machine learning models
data that can be used to determine their risk of mortality require large amounts of data to create accurate
while they are in the ICU. The tool aggregates patient data predictions. If data is unavailable in sufficient
such as age, medical history, ICU severity scores,
quantities, then model accuracy may suffer.
laboratory values, and other data to generate a risk
• Incorrect labeling: The existing machine learning
stratification score. The higher the score, the greater the
models are only as accurate as the data trained using
chance of mortality. The tool can help healthcare providers
predict and treat heart failure cases by identifying high-risk the input dataset. If the data has been incorrectly
patients. It, in turn, allows for earlier interventions and labeled, the model cannot make accurate
improved outcomes for these patients. Winslow, C. J., et predictions.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

• Over fitting: It is a common problem in machine cardiac arrest in newborns, allowing medical professionals to
learning models where the model needs to be more take steps to mitigate them. In addition, machine learning
specific and capture more noise from the data, models can monitor for signs of cardiac arrest in real time,
resulting in inaccurate predictions. allowing for prompt and appropriate interventions.
• Unbalanced data: Unbalanced data can also lead to
inaccurate predictions as the model may not be able III. PROPOSED MODEL
to detect unusual patterns. Constructing the proposed cardiac machine-learning model
• Model selection: Choosing a suitable model for the requires several steps. First, the data must be collected and
data is a challenging task. There are a variety of pre-processed. It includes gathering relevant cardiac data
models to choose from, and selecting the wrong such as electrocardiograms (ECG), other medical images,
model could lead to accurate predictions and better and any relevant patient information, such as age and gender.
results. The data must then be cleaned and transformed into a format
Early detection of cardiac arrest in newborn babies is suitable for machine learning algorithms, such as numerical
essential to ensure their survival and health. If cardiac arrest or categorical values. Once the data is ready, a machine-
is not identified and treated quickly, it can cause permanent learning model must be selected. It is typically a neural
damage to the baby's heart and other organs, leading to long- network model, as it can handle the complex relationships
term health problems. Early detection of cardiac arrest allows between the various data points. The model must then be
for immediate treatment, improving the chances of survival trained using the data and evaluated for accuracy. If
and long-term health for the newborn. Early detection also necessary, the model can be tweaked to improve its
allows the baby to receive necessary medical care to prevent performance. Finally, the model must be deployed. It
further complications or death. involves creating an application or web interface for the
• Establish a baseline for newborn babies' normal model to be used by medical professionals. The model
range of vital signs. This baseline can monitor any should also be continuously monitored for accuracy and
changes in the vital signs that could indicate a necessary adjustments. The block diagram of the proposed
potential cardiac arrest. machine learning model has shown in the following fig.1
• Developing a statistical model using machine
learning techniques to analyze the vital signs of
newborn babies. This model can identify patterns
and trends in the vital signs indicative of an
impending cardiac arrest.
• Utilizing predictive analytics to identify risk factors
associated with cardiac arrest in newborn babies. It
can help in predicting the likelihood of cardiac
arrest before it occurs.
• Develop an algorithm that can detect the subtle
changes in the vital signs indicative of an
impending cardiac arrest. This algorithm can alert
medical personnel when a potential cardiac arrest is
detected.
• Utilizing advanced signal processing techniques to
identify the subtle changes in the vital signs that
may indicate an impending cardiac arrest.
• Developing an early warning system that can alert
medical personnel during an impending cardiac
arrest. This system can help in providing timely
medical intervention and reduce mortality rates.
A machine-learning approach using statistical models is an
essential tool for the early detection of cardiac arrest in
newborn babies. By leveraging data collected from various FIGURE 1. Proposed block diagram
sources, such as medical records, vital signs, and test results,
machine learning models can accurately identify patterns and In the proposed method, the first detected patient symptoms
predict the occurrence of cardiac arrest. It can help healthcare are given as input. All these data are stored in the database,
providers take proactive steps to prevent the condition or and their volumes are categorized. These classifications
intervene to improve the baby's chances of survival. Machine provide information regarding the treatment provided in the
learning can also help identify risk factors that could lead to standard unit and the treatment provided in the emergency

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

unit, depending on the severity of the illness. The proposed VIS Vasoactive–inotropic score
algorithm tests these provided information blocks to predict CPB cardiopulmonary bypass
the severity of the patient's heart block problem. Accurate HFpEF heart failure with preserved ejection
results are obtained, and treatments are provided for him. It is fraction
documented and stored back in the database. MIMIC Multiparameter Intelligent
Monitoring in Intensive Care
Cardiac Machine Learning model (CMLM) enables the ICU intensive care unit
detection and analysis of heart disease via a combination of EHR electronic health record
several advanced machine-learning algorithms. At its core, MEC Mobile Edge Computing
CML involves creating a model which takes in large amounts SNMF sparse nonnegative matrix
of data, such as patient records, electrocardiogram (ECG) factorization
readings, clinical images of the heart, and other criteria and CNN convolutional neural network
processes it to detect and predict cardiac abnormalities. The EEG electroencephalography
core components of a typical CML model consist of feature ML machine learning
extraction, data pre-processing and model building. The FHR contains fetal heart rate
Feature Extraction step converts raw data into meaningful OS oxygen saturation
information or features, which a machine learning model can
FRR fetal respiratory rate
use for further analysis. Feature extraction uses pre-trained
PCA Principal Component Analysis
algorithms to categorize data points and extract traits that
may be essential features in a learning model. The critical RFE Recursive Feature Elimination
features depend on the problem domain; for example, the
data may contain patient records, ECGs, and clinical images A. Symptoms dataset
of the heart for heart disease detection. The model can be
better fitted to perform accurate data analysis by extracting Data construction and modeling are as follows:
essential features from these data points. Data pre-processing • A total of 84972 samples were selected from the
is a critical part of any machine learning model. It is data pcap files for the Newborn Babies in the
responsible for cleaning and preparing data for further Cardiac Intensive Care Unit [32]. These samples
processing by removing or replacing any noise or outliers. It contains fetal heart rate (FHR), oxygen saturation
helps the model focus on the essential features and improves (OS), fetal respiratory rate (FRR), body temperature
accuracy and efficiency. In the case of CML models, it is readings, and other medical constraints
essential to prepare the data for input by removing or • Totally 65% of data used for training (TR) the
replacing any missing values and normalizing the data in models and 35% of data used for testing (TS) the
order to detect any abnormalities in the data accurately. Once model
the data has been prepared, a model can be built using • It also used cross-validation on the training dataset
cardiac machine learning algorithm. With this algorithm, the with 20-folds.
model can classify the data and detect abnormalities,
providing an accurate analysis of the data points, which can B. Preprocessing
then be used for further medical diagnosis or treatment. The
core components of a CML model allow for the detection of Preprocessing is an essential step in the machine learning
cardiac abnormalities efficiently and accurately. Combining approach for Early Detection of Cardiac Arrest in Newborn
feature extraction, data pre-processing, and model building,
Babies in the Cardiac Intensive Care Unit. During this stage,
CML models can detect and predict heart conditions quickly
the data are prepared for machine learning analysis. This
and accurately, making them a valuable tool for medical
professionals. Table.1 shows the symbolic representation includes cleaning and normalizing the data, selecting the
details. appropriate features, and applying various transformations to
the data. Additionally, any missing values or outliers must be
Table.1: Symbolic representation detail identified and addressed. This process ensures that the data is
Abbreviation Meaning suitable for machine learning algorithms. The data must be
CMLM Cardiac machine learning model split into training, validation, and test sets so that the models
CICU Cardiac intensive care unit can be evaluated and compared. The preprocessing functions
Tr Training data would likely include:
Ts Testing data • Data cleaning: Remove any duplicate or irrelevant
FDR False discovery rate data, fill in missing values, and standardize data
FOR False omission rate formats.
CSI Critical success index • Missing value imputation: Replace any missing
SVM Support vector machine values with meaningful estimates such as the mean
ANN Artificial neural network of the feature or a prediction from another model.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

• Data scaling: Scale the data so that all features have as decision trees and logistic regression to identify the most
a similar range of values. This can be done using a important features from the data. This will help to identify
variety of techniques such as min-max scaling or the features that are predictive of the risk of cardiac arrest in
standardization. newborn babies. Furthermore, feature selection techniques
• Feature selection: Select only the most relevant such as Principal Component Analysis (PCA) and Recursive
features that will help the model make accurate Feature Elimination (RFE) can be used to further refine the
predictions. This can be done using techniques such feature set and reduce the number of features required for the
as Recursive Feature Elimination or Principal predictive model. Once the important features have been
Component Analysis. identified, they can be used to build a predictive model which
• Data transformation: Transform the data into a can be used to help predict the risk of cardiac arrest in
newborn babies. This model can then be used by healthcare
format that the machine learning model can
providers to take preventative measures to reduce the risk of
interpret. This can include one-hot encoding
cardiac arrest in newborn babies
categorical features or applying a logarithmic
transformation to skewed features.
D. Classification
Preprocessing functions are necessary in order to clean and
prepare the data for a machine learning model. Preprocessing Classification is a supervised learning technique used in
functions can include data cleaning, missing value machine learning for early detection of cardiac arrest in
imputation, data scaling, feature selection, data newborn babies. This technique is used to assign a label to a
transformation, and more. given input data based on its characteristics. In this case, the
input data consists of the baby’s vital signs such as heart rate,
respiratory rate, body temperature, oxygen saturation, and
C. Feature Extraction other medical parameters. The classification algorithm will
then analyze these parameters and output a label that
Feature extraction for Early Detection of Cardiac Arrest in indicates whether the baby is at risk of developing cardiac
Newborn Babies is the process of identifying the most arrest or not. The output label can be either positive or
important features from the data that can be used to predict negative, depending on the probability of the occurrence of
cardiac arrest in newborn babies. This is done by analyzing the condition. The classification operation of a machine
the patient’s medical history, vital signs, and other relevant learning approach for Early Detection of Cardiac Arrest in
data. These features can be extracted from the data using Newborn Babies involves several steps. First, a dataset is
various methods such as statistical analysis, machine learning collected containing information about newborns and their
algorithms, and feature selection techniques. Feature vital signs. This dataset is then split into two sets: a training
extraction for a machine learning approach for early set and a testing set. The training set is used to train a
detection of cardiac arrest in newborn babies can involve the machine learning model using supervised learning algorithms
following: such as support vector machine and random forest. The
• Fetal Heart Rate (FHR): The FHR can be used to model is then tested on the testing set to determine its
measure the health of the baby. Abnormal FHR accuracy. The model is evaluated on unseen data to assess its
patterns can be an indication of cardiac arrest. performance. This performance is then used as the basis for
• Fetal Movements: A decrease or lack of fetal making decisions about whether or not a newborn is at risk
movements can be a sign of cardiac arrest and can for cardiac arrest. It can be used to identify patterns in the
be monitored for early detection. data and accurately classify them into the correct category.
• Maternal Vital Signs: Maternal vital signs such as
blood pressure, heart rate, oxygen saturation, and
temperature can be used to indicate the health of the E. Proposed Algorithm
baby. Abnormalities in these vital signs can be an
indication of cardiac arrest. Cardiac machine learning algorithms are a type of artificial
• Umbilical Cord Blood Flow: The umbilical cord intelligence (AI) that can be used to detect early signs of
blood flow rate can be used to measure the baby's cardiac arrest in newborn babies. Fig.2 illustrates the flow
oxygenation levels. Any abnormalities in the blood information of proposed model. These algorithms use a
flow rate can be a sign of cardiac arrest. combination of data from electrocardiograms (ECGs), stress
• Ultrasound: Ultrasound images can be used to tests, and other sources to detect abnormalities and predict
detect any abnormalities in the baby's heart and can the likelihood of an adverse cardiac event. By analyzing
be used for early detection of cardiac arrest. historical data and identifying patterns, the algorithm can
These features can be used by the proposed machine learning provide early warning signs, before an arrest occurs. This
algorithms to detect signs of cardiac arrest in newborn early detection can help doctors to intervene in a timely
babies. For example, a machine learning approach for this manner and provide appropriate treatment for the baby. The
task may involve using supervised learning techniques such algorithm can also provide a better understanding of the

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

baby's condition and help doctors to assess their prognosis. ECG readings, and other clinical information. This data is
This can help to ensure that the baby receives the best then used to create a model which can accurately detect early
possible care and treatment. The algorithm.1 has shown the signs of cardiac arrest in newborn babies. The extracted data
structure of cardiac machine learning algorithm. is then used to create a predictive model which is capable of
detecting early signs of cardiac arrest in newborns. This
Algorithm 1: Cardiac machine learning algorithm model is based on various machine learning algorithms, such
1. Start; as logistic regression, neural networks, and support vector
2. Enter the dataset values; machines. The performance of the model is then evaluated on
3. Loading the Raw data inputs; a test dataset. This dataset contains the medical data of both
4. Initiate the preprocessing; healthy and sick newborn babies. The accuracy of the model
5. Dimensionality reduction using PCA; is then determined by comparing the predictions made by the
6. Obtain the feature vector ; model against the actual outcomes of the test dataset. Once
7. Resampling the inputs; the model has been evaluated, it can then be optimized to
8. Initiate the Training process; further improve its accuracy. This can be done using various
9. Start the Classification, regression and clustering techniques such as feature selection, parameter tuning, and
of the Input samples; ensembling. The performance of cardiac machine learning
10. Obtain the trained model; algorithms can be improved for early detection of cardiac
11. Initiate the Testing process arrest in newborn babies in a number of ways. First, the data
12. Start the Classification, regression and clustering used for training the algorithms should include both clinical
of the Input samples; data and imaging data. This will help the algorithms to better
13. Obtain the tested model; identify patterns in the data and make more accurate
14. Insert the evaluation parameters as per the training predictions. Second, the algorithms should be trained using a
and testing inputs; larger dataset. This will help them learn better and detect
15. If (Accuracy = Max.level) more subtle patterns in the data. Finally, the algorithms
16. Then predict the results for neonates; should be optimized to better handle imbalanced data, as well
17. Else go to step 7; as a variety of noise and artifacts present in cardiac data. By
18. End; taking these steps, the performance of cardiac machine
learning algorithms can be improved for early detection of
cardiac arrest in newborn babies. A precompiled method runs
based on the given inputs. Its model is explained in
Algorithm-1 below. First, the dataset inputs are given. Then
their data blocks are loaded, and raw data blocks are
available as input. These raw data blocks are now
preprocessed, and the resulting results are dimensionally
reduced. Thus the raw data block is now inputted with
feature vector processing formalized sample blocks. The
heart beat segment of neonates has computed with the help of
the following fig.3.

FIGURE 3. Heart beat segment of neonates

The heart beat segments of neonates are currently, divided

into two modules, namely training and testing. The training
model is configured with the training dataset module's
classification, regression and clustering operations.
Currently, test data model results are based on the results
FIGURE 2. Proposed Flow diagram
obtained through classification and regression in the testing
dataset module. Now the results obtained through training
and testing are compared. If the two results are compared,
The Cardiac machine learning algorithm first extracts
and more accurate results are obtained, the prognostic
features from the patient’s medical data such as vital signs,

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

methods of the disease are predicted. The data is then re-

sampled when perhaps in results. Currently, the proposed
method is working in this class

IV. ANALYTICAL DISCUSSIONS

Generally, convolutional input data creates temporal patterns
in the existing signals. Here the differential signal curves T
and the kernels value of the signal volume can be denoted as
K. The various representations of this are given below based
on eq.1.
æ T -1 ö
aix [ v ] = a i ç åbix d1i [ v - i ] + d ix ÷ (1)
è i =0 ø
Where, α is expressed an activation function, β indicated a
weight of the kernel and δ are indicated as the bias of the
FIGURE 4. Analyzed various intensive care ECG signals
input signal. d[v] expressed the input signal value. Here the
output signal values are fed into the convolution layer. Then
the kernel value is equal to double of the signal value. Now Finally the classification has to expressed as the following
eq.6
the non-overlapped segments of the input heat signals are
expressed in eq.2 1
C= (6)
ri [ v ] = max max {a [u ]}
- ( b *bG 1 +d )
x x v*U
(2) 1+ e
i u =( v -1)U +1 Here the cost function has computed to minimize the
Now to compute the output heart signal values as per the entropy values of binary heart rate signals without any
newborns present heart rate. If the newborn has the heart harmonic disturbance.
murmur issues, then the computation results have some
harmonic disturb from the murmur sound. So, we need to
cost å i
i i i ( i )
F = a é h ( original ) ln ln ( C ) + 1 - h ( original ) ln ln (1 - C ) ù
ë i
û
(7)

confirm the length of the output signal is equal to the non-

overlapping of other signals. This has shown in the following Where the hi(original) are the natural strength of the predicted
eq.3 signals values at the sample weights. The SVM prediction
six v = ri x v[] [] (3) has computed with the help of below eq.8
æ N
ö
hi ( Prediction )
= sign ç d + åbi * K ( u , ui ) ÷ (8)
Where, six [v] denotes the length of the output signals and the è i =1 ø
rix[v] expressed the non-overlapped signals. The upper
segment signals need to be reduced as per the computed The Various heart rates monitoring in continuous evaluation
redundant segment results. Now the second segment kernel has demonstrated in the fig.5.
values has expressed in the eq.4
æ K T -1 ö
a xj [ v ] = a j ç ååbix* j d1j [ v - i ] + d ix* j ÷ (4)
è j =1 i =0 ø

Where, α is expressed an activation function, β indicated a

weight of the kernel and δ are indicated as the bias of the
input signal. d[v] expressed the input signal value. Now we
need to obtain the feature vector values as per the received
convolutional signals. The Analyzed various intensive care
ECG signals has demonstrated in the fig.4
bG1 = max max {a xj [u ]} (5)

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

FIGURE 5. Various heart rates monitoring in continuous evaluation • The patient's diet should be balanced, rich in
calcium, magnesium, potassium and manganese
Medical Treatments: The deficiency cannot be cured by (most of them in oats, barley, buckwheat, apples
drugs. They are mainly used to relieve the symptoms of and prunes). At the same time, it is undesirable to
congestive heart failure. The following medications may also pay attention to salt and pickled foods, preservation.
be prescribed: It is better to eat in small portions, but often.
• Non-steroidal anti-inflammatory - with a • A child should go to bed on time, as proper rest
pronounced reaction to infection by the "forces" of significantly reduces the burden on the heart.
innate immunity; • The patient should be protected from situations that
• Angioprotectors - if vascular damage is observed; lead to overwhelm or frustration. It is also not
• Penicillin-based antibiotics - when the deficiency is recommended to mount it physically.
triggered by pathogenic bacteria; • If the weather is comfortable, regular walking is
• Cardiovascular therapy - for the treatment of acute essential.
failure.
The motivation for developing a Cardiac machine learning
Functional treatments: Surgical intervention is the only way algorithm is to provide a more accurate and reliable method
to completely eliminate defects in the cardiovascular system. for diagnosing and prognosis cardiac diseases. The
Sometimes the only way to save a child is to do it this way. challenges of developing such an algorithm include limited
The Surgery is recommended if you have the following availability of data, the need for robust feature selection, and
symptoms: the need to improve the efficiency and accuracy of the
• After the slightest physical exertion, the patient algorithms. Additionally, the lack of expert resources for
immediately develops shortness of breath, and other validating and verifying the machine learning models poses a
symptoms of insufficiency are also observed; significant challenge. Understanding and manipulating the
• Diagnostics shows the pathological expansion of complex relationships between the clinical features that need
any cardiac chambers and its task for "wear and to be captured to create a successful model is complex.
tear"; Finally, the models must be constantly updated to remain
• Pressure increases in one ventricle. relevant and practical. Motivation is essential in any machine
learning algorithm, especially for diagnosing, predicting, and
tracking cardiac health. It is important to remember that
sustaining motivation can be complex for any learner and
that the challenge posed by algorithms exists within that
larger context. To help overcome the motivation and
challenges the cardiac machine learning algorithm poses,
learners should focus on a few key steps.
• First and foremost, it is essential to understand the
data and algorithms that go into developing the
machine learning model. It can better understand
the problem domain and give the learner
meaningful feedback on their progress and
performance.
• Second, they should break the problem down into
smaller, achievable goals. When doing so, it is
essential to focus on the basics, such as
understanding the machine learning techniques,
implementing different models, and tuning the
parameters.
• Third, learners should consider using visualization
methods to understand the data better and enhance
their problem-solving skills. For example, they
create flow charts, timelines, or network diagrams
FIGURE 6. Preventive measures of the analytical results that can help explain complex relationships within
the data.
Preventive measures: The Preventive measures of the • Finally, learners should use reinforcement learning
analytical results have shown in the fig.6. It is impossible to to improve the machine learning model's
effectively treat the defect without following the correct performance over time. Reinforcement learning
regimen of the child: takes a trial-and-error approach to machine learning,
continually adjusting and refining its results.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

These steps will help learners reach their motivation and Signal detection connectivity 0.2
challenge goals when working on a cardiac machine learning Frame duration 150ms
algorithm. Busy state parameter of training session 6
The computational overhead of a cardiac machine-learning Idle state parameter of training session 5
model depends on the size of the dataset, the complexity of Busy state parameter of testing session 6
the model being used, and the number of layers the model Idle state parameter of testing session 5
needs for accurate predictions. In addition, any changes Maximum Interference Ratio 0.7
needed to be made to the computational architecture of the Permission availability 0.5
model would also need to be considered. As more complex
machine learning models are created, the computational A. Computation of delta-P (Δp)
overhead costs and complexity typically increase to account DeltaP (Δp) values are a measure of the difference in
for additional layers and complexity. Computational pressure between the left and right ventricles of the heart in
overhead in the cost and complexity of the cardiac machine newborn babies in the Cardiac Intensive Care Unit (CICU).
learning model refers to the resources required to implement, This measure is used to detect early signs of cardiac arrest in
maintain and operate the machine learning models. These newborns. The values are calculated by taking the difference
resources include hardware (e.g., computing clusters, GPUs, between the peak systolic pressure of the left ventricle and
or ASICs), software, and storage. Virtual machines, cloud the peak diastolic pressure of the right ventricle. A high
computing, and clustering components can simplify the cost deltaP (Δp) value can indicate a possible cardiac arrest, and
and complexity of the machine learning model. The can alert medical staff to take appropriate action to prevent
computational overhead is an essential aspect of cost and the condition from becoming more serious.
complexity and must be considered when designing a cardiac
machine-learning model. It affects the model's scalability, Dp = PPV + NPV - 1 (9)
accuracy, and response time and can be optimized via
parallelization, which entails running multiple models in
Where, PPV represents the positive predicted value and NPV
parallel. Additionally, computational overhead may include
represents the negative predicted value.
data ingestion, pre-and post-processing of the data, feature
engineering, and model training. Pt Nt
PPV = ; NPV = (10)
Pt + Pf Nt + N f
V. COMPARATIVE ANALYSIS
The proposed Machine Learning Approach (MLA) has Where, Pt represents the true positive prediction, Pf
represents the false positive prediction, Nt represents the true
compared with the existing Detection and Diagnosis of
negative prediction and Pf represents the false negative
Cardiac Amyloidosis (DDCA), Heart Failure and Atrial
prediction values. Now substitute the eq.10 in eq.9.
Fibrillation (HFAF) and Risk Factors for Heart Failure
(RFHF). Here the MATLAB r2022a is the tool used for ïì Pt ïü ïì N t ïü
simulation process. The table 2 express the simulation Dp = í ý+í ý -1 (11)
parameters of the proposed environment. ïî Pt + Pf ïþ ïî N t + N f ïþ

Table 2: Simulation parameters Table.3 shows the various delta-P evaluations between the
Parameter Value proposed and existing models. In that Tr represents the
Simulation duration 1500 fps training data and Ts represents the testing data.
Preamble duration 15ms
TABLE III
EVALUATION OF DELTA-P
No.of DDCA DDCA HFAF HFAF RFHF RFHF CMLA CML
Inputs (Tr) (Ts) (Tr) (Ts) (Tr) (Ts) (Tr) A (Ts)
100 0.456 0.625 0.547 0.568 0.678 0.743 0.852 0.936
200 0.472 0.602 0.552 0.586 0.664 0.751 0.867 0.926
300 0.488 0.579 0.557 0.604 0.65 0.759 0.882 0.916
400 0.504 0.556 0.562 0.622 0.636 0.767 0.897 0.906
500 0.52 0.533 0.567 0.64 0.622 0.775 0.912 0.896
600 0.536 0.51 0.572 0.658 0.608 0.783 0.927 0.886
700 0.552 0.487 0.577 0.676 0.594 0.791 0.942 0.876
800 0.568 0.464 0.582 0.694 0.58 0.799 0.957 0.866
900 0.584 0.441 0.587 0.712 0.566 0.807 0.972 0.856
1000 0.6 0.418 0.592 0.73 0.552 0.815 0.987 0.846

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

B. Computation of false discovery rate (FDR)

The False Discovery Rate (FDR) is a statistical measure used
to determine the accuracy of results in a clinical trial or
study. It refers to the proportion of false positives among all
positives identified in the trial or study. In the context of
Early Detection of Cardiac Arrest in Newborn Babies in the
Cardiac Intensive Care Unit, the FDR is a measure of the
accuracy of the study’s results in detecting cardiac arrest. The
lower the FDR, the more accurate the study’s results will be.
This is important because it helps clinicians to identify
patients at risk of cardiac arrest and intervene to save lives.

FDR = 1 - PPV (12)

Where, PPV represents the positive predicted value, Pt

FIGURE 7. Evaluation of Delta-P represents the true positive prediction, Pf represents the false
positive prediction. The PPV has obtained in the eq.10.
Fig.7 demonstrates the Evaluation of Delta-P in different
ïì Pt ïü
level of inputs. In a training (Tr) comparison region, the FDR = 1 - í ý (13)
proposed CMLA reached 0.912 delta-p value. Meanwhile the ïî Pt + Pf ïþ
DDCA reached 0.52, HFAF reached 0.567 and RFHF Table.4 shows the various false discovery rate (FDR)
reached 0.622 delta-p values respectively. In a testing (Ts) evaluations between the proposed and existing models. In
comparison region, the proposed CMLA reached 0.896 delta- that Tr represents the training data and Ts represents the
p value. Meanwhile the DDCA reached 0.533, HFAF testing data.
reached 0.64 and RFHF reached 0.775 delta-p values.
TABLE IV
EVALUATION OF FALSE DISCOVERY RATE
No.of DDCA DDCA HFAF HFAF RFHF RFHF CMLA CML
Inputs (Tr) (Ts) (Tr) (Ts) (Tr) (Ts) (Tr) A (Ts)
100 0.590 0.809 0.708 0.735 0.877 0.839 0.835 0.918
200 0.611 0.779 0.714 0.758 0.859 0.848 0.850 0.908
300 0.632 0.749 0.721 0.782 0.841 0.857 0.865 0.898
400 0.652 0.720 0.727 0.805 0.823 0.866 0.879 0.888
500 0.673 0.690 0.734 0.828 0.805 0.875 0.894 0.878
600 0.694 0.660 0.740 0.852 0.787 0.884 0.909 0.869
700 0.714 0.630 0.747 0.875 0.769 0.893 0.924 0.859
800 0.735 0.600 0.753 0.898 0.751 0.902 0.938 0.849
900 0.756 0.571 0.760 0.921 0.732 0.911 0.953 0.839
1000 0.776 0.541 0.766 0.945 0.714 0.920 0.968 0.829

FIGURE 8. Evaluation of false discovery rate

Fig.8 demonstrates the evaluation of false discovery rate in
different level of inputs. In a training (Tr) comparison region,
the proposed CMLA reached 0.894 FDR value. Meanwhile
the DDCA reached 0.673, HFAF reached 0.734 and RFHF
reached 0.805 FDR values respectively. In a testing (Ts)
comparison region, the proposed CMLA reached 0.878 FDR
value. Meanwhile the DDCA reached 0.690, HFAF reached
0.828 and RFHF reached 0.875 FDR values.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

C. Computation of false omission rate (FOR)

False omission rate values refer to the percentage of cardiac Where, NPV represents the negative predicted value, Nt
arrests in newborn babies in the Cardiac Intensive Care Unit represents the true negative prediction, Nf represents the false
that are not detected by the Early Detection System. This rate negative prediction. The NPV has obtained in the eq.10.
is an important measure of the effectiveness of the system in Nt
identifying and responding to cardiac arrests in newborns. A FOR = 1 - (15)
false omission rate of zero may indicate that the system is Nt + N f
very effective in detecting cardiac arrests, while high false Table.5 shows the various false omission rate (FOR)
omission rate values may indicate a need for improvement. evaluations between the proposed and existing models. In
that Tr represents the training data and Ts represents the
FOR = 1 - NPV (14) testing data.

TABLE V
EVALUATION OF FALSE OMISSION RATE
No.of DDCA DDCA HFAF HFAF RFHF RFHF CMLA CML
Inputs (Tr) (Ts) (Tr) (Ts) (Tr) (Ts) (Tr) A (Ts)
100 0.666 0.914 0.800 0.694 0.864 0.722 0.019 0.110
200 0.690 0.880 0.806 0.716 0.847 0.729 0.033 0.090
300 0.714 0.846 0.814 0.739 0.829 0.737 0.048 0.080
400 0.736 0.813 0.821 0.761 0.811 0.745 0.062 0.071
500 0.760 0.779 0.829 0.782 0.793 0.753 0.076 0.061
600 0.784 0.745 0.836 0.805 0.776 0.760 0.091 0.052
700 0.806 0.712 0.844 0.827 0.758 0.768 0.006 0.042
800 0.830 0.678 0.850 0.849 0.740 0.776 0.020 0.032
900 0.854 0.645 0.858 0.870 0.722 0.784 0.034 0.023
1000 0.876 0.611 0.865 0.893 0.704 0.791 0.049 0.013

The prevalence threshold is the minimum number of cardiac

arrests that must occur in a given population before Early
Detection of Cardiac Arrest in Newborn Babies in the
Cardiac Intensive Care Unit (CICU) is considered necessary.
The threshold is based on the incidence of cardiac arrest in
the population, and serves as a guide to determine when early
detection of cardiac arrest should be considered. In general, a
prevalence threshold of 2 or more cardiac arrests in a given
population is considered to be sufficient to suggest early
detection of cardiac arrest as a viable strategy.

ìï PFR üï
Pth = í ý (16)
ïî PTR + PFR ïþ

Where, PFR represents the positive false rate, PTR represents

FIGURE 9. Evaluation of false omission rate
the positive true rate, Pt represents the true positive
prediction, Pf represents the false positive prediction, Nt
Fig.9 demonstrates the evaluation of false omission rate represents the true negative prediction and Pf represents the
(FOR) in different level of inputs. In a training (Tr) false negative prediction values.
comparison region, the proposed CMLA reached 0.076 FOR
value. Meanwhile the DDCA reached 0.760, HFAF reached
æ Pt ö æ Pf ö
PTR = ç ÷÷ ; PFR = çç ÷÷ (17)
çP+N
è Pf + N t
0.829 and RFHF reached 0.793 FOR values respectively. In a
testing (Ts) comparison region, the proposed CMLA reached
è t f ø ø
Substitute eq.17 in eq.16.
0.061 FOR value. Meanwhile the DDCA reached 0.779,
HFAF reached 0.782 and RFHF reached 0.753 FOR values.

D. Computation of prevalence threshold (Pth)

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

ì æ Pf ö ü
ï ï Table.6 shows the various prevalence threshold evaluations
çç ÷÷ between the proposed and existing models. In that Tr
ïï è Pf + N t ø ïï
represents the training data and Ts represents the testing data.
Pth = í ý (18)
ï æ Pt ö æ Pf ö ï
ï çç ÷÷ + çç ÷÷ ï
ïî è Pt + N f ø è Pf + N t ø ïþ

TABLE VI
EVALUATION OF PREVALENCE THRESHOLD
No.of DDCA DDCA HFAF HFAF RFHF RFHF CMLA CML
Inputs (Tr) (Ts) (Tr) (Ts) (Tr) (Ts) (Tr) A (Ts)
100 0.572 0.785 0.688 0.597 0.852 0.621 0.803 0.882
200 0.592 0.756 0.693 0.616 0.835 0.627 0.817 0.873
300 0.613 0.726 0.700 0.636 0.817 0.634 0.831 0.863
400 0.632 0.698 0.706 0.655 0.799 0.641 0.845 0.854
500 0.653 0.669 0.713 0.673 0.782 0.648 0.859 0.844
600 0.673 0.640 0.719 0.692 0.765 0.654 0.874 0.835
700 0.692 0.611 0.726 0.711 0.747 0.661 0.888 0.825
800 0.713 0.582 0.731 0.730 0.729 0.668 0.902 0.816
900 0.733 0.554 0.738 0.748 0.712 0.674 0.916 0.807
1000 0.752 0.525 0.744 0.768 0.694 0.680 0.930 0.797

newborn babies in the Cardiac Intensive Care Unit (CICU).

The CSI is calculated by dividing the number of cardiac
arrests that have been detected early and treated successfully
by the total number of cardiac arrests that occurred during a
certain period of time. This index helps health care providers
to understand how successful their early detection efforts are
and where to focus their resources for improvement.

ìï Pt üï
CSI = í ý (19)
îï Pt + N f + Pf þï
Where, PFR represents the positive false rate, PTR represents
the positive true rate, Pt represents the true positive
prediction, Pf represents the false positive prediction, Nt
represents the true negative prediction and Pf represents the
FIGURE 10. Evaluation of prevalence threshold false negative prediction values.

Fig.10 demonstrates the evaluation of prevalence threshold in

different level of inputs. In a training (Tr) comparison region,
the proposed CMLA reached 0.859 prevalence threshold
value. Meanwhile the DDCA reached 0.653, HFAF reached
0.713 and RFHF reached 0.782 prevalence threshold values
respectively. In a testing (Ts) comparison region, the
proposed CMLA reached 0.844 prevalence threshold value.
Meanwhile the DDCA reached 0.669, HFAF reached 0.673
and RFHF reached 0.648 prevalence threshold values.

E. Computation of critical success index (CSI)

The critical success index (CSI) is a metric used to measure
the effectiveness of early detection of cardiac arrest in

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

that Tr represents the training data and Ts represents the

FIGURE 11. Evaluation of critical success index testing data.

Table.7 shows the various critical success index (CSI)

evaluations between the proposed and existing models. In
TABLE VII
EVALUATION OF CRITICAL SUCCESS INDEX
No.of DDCA DDCA HFAF HFAF RFHF RFHF CMLA CML
Inputs (Tr) (Ts) (Tr) (Ts) (Tr) (Ts) (Tr) A (Ts)
100 0.491 0.674 0.592 0.514 0.840 0.534 0.787 0.865
200 0.508 0.649 0.596 0.530 0.823 0.539 0.801 0.856
300 0.526 0.623 0.602 0.547 0.805 0.545 0.815 0.846
400 0.543 0.599 0.607 0.563 0.788 0.551 0.828 0.837
500 0.561 0.574 0.613 0.579 0.771 0.557 0.842 0.827
600 0.578 0.550 0.618 0.595 0.754 0.563 0.857 0.819
700 0.594 0.525 0.625 0.612 0.736 0.569 0.871 0.809
800 0.612 0.500 0.629 0.628 0.719 0.575 0.884 0.800
900 0.629 0.476 0.635 0.643 0.702 0.580 0.898 0.791
1000 0.646 0.451 0.640 0.661 0.684 0.585 0.912 0.781
specificity. The sensitivity measures the proportion of true
Fig.11 demonstrates the Evaluation of critical success index positives correctly identified by the model. The specificity
in different level of inputs. In a training (Tr) comparison measures the proportion of true negatives correctly identified
region, the proposed CMLA reached 0.842 delta-p value. by the model. The model's accuracy is then calculated as the
Meanwhile the DDCA reached 0.561, HFAF reached 0.613 arithmetic mean of these two values.
and RFHF reached 0.771 CSI values respectively. In a testing
(Ts) comparison region, the proposed CMLA reached 0.827
CSI value. Meanwhile the DDCA reached 0.574, HFAF
reached 0.579 and RFHF reached 0.557 CSI values.

F. Computation of Accuracy (A)

The accuracy of a Cardiac Machine Learning Model
(CMLM) is calculated by comparing the predicted results
with the observed results. The CMLM accuracy depends on
the quality of the model, the quality of the input data, the
performance of the validation mechanism used in the model,
and the performance of the prediction mechanism used in the
model. To calculate the accuracy of the CMLM, first, the
model's performance is assessed by splitting the data into
training and test sets. The data to train the model must be
sufficiently large and unbiased; otherwise, it may lead to over
fitting or invalid predictions. The test set is then used to FIGURE 12. Evaluation of Accuracy
validate the model's performance. It is done by comparing
the model's predicted outcomes with the observed outcomes Table.8 shows the various accuracy evaluations between the
in the test set. Two values are calculated from the test set to proposed and existing models. In that Tr represents the
calculate the accuracy of the CMLM for early detection of training data and Ts represents the testing data.
cardiac arrest in newborn babies in the Cardiac Intensive
Care Unit (CICU). These values include sensitivity and

TABLE VIII
EVALUATION OF ACCURACY
No.of DDCA DDCA HFAF HFAF RFHF RFHF CMLA CML
Inputs (Tr) (Ts) (Tr) (Ts) (Tr) (Ts) (Tr) A (Ts)
100 0.377 0.538 0.472 0.490 0.668 0.641 0.856 0.918
200 0.390 0.518 0.476 0.506 0.655 0.648 0.871 0.908
300 0.403 0.499 0.481 0.521 0.641 0.655 0.886 0.898
400 0.416 0.479 0.485 0.537 0.627 0.662 0.901 0.889

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

500 0.430 0.459 0.489 0.552 0.613 0.669 0.916 0.879

600 0.443 0.439 0.493 0.568 0.599 0.675 0.931 0.869
700 0.456 0.419 0.498 0.583 0.586 0.682 0.947 0.859
800 0.469 0.400 0.502 0.599 0.572 0.689 0.962 0.849
900 0.482 0.380 0.506 0.614 0.558 0.696 0.977 0.840
1000 0.496 0.360 0.511 0.630 0.544 0.703 0.992 0.830

Fig.12 demonstrates the Evaluation of accuracy in different

level of inputs. In a training (Tr) comparison region, the
proposed CMLA reached 0.916 accuracy value. Meanwhile
the DDCA reached 0.430, HFAF reached 0.489 and RFHF
reached 0.613 accuracy values respectively. In a testing (Ts)
comparison region, the proposed CMLA reached 0.879
accuracy value. Meanwhile the DDCA reached 0.459, HFAF
reached 0.552 and RFHF reached 0.669 accuracy values.

G. Computation of Precision (P)

The precision of a CMLM for early detection of cardiac
arrest in newborns in the CICU can be calculated by
measuring the true positives (TP) out of the total number of
positives (TP + False Positives (FP)). The true positives
measure the percent of cases the model correctly identifies as
cardiac arrest, while the false positives measure the percent
of cases the model incorrectly identifies as cardiac arrest. The FIGURE 13. Evaluation of critical success index
precision score is then calculated by dividing the true
positives by the total number of positives. Table.9 shows the precision evaluations between the
proposed and existing models. In that Tr represents the
training data and Ts represents the testing data.

TABLE IX
EVALUATION OF PRECISION
No.of DDCA DDCA HFAF HFAF RFHF RFHF CMLA CML
Inputs (Tr) (Ts) (Tr) (Ts) (Tr) (Ts) (Tr) A (Ts)
100 0.487 0.697 0.611 0.634 0.865 0.724 0.839 0.900
200 0.505 0.671 0.616 0.654 0.847 0.732 0.854 0.891
300 0.522 0.645 0.622 0.675 0.829 0.739 0.869 0.881
400 0.539 0.620 0.627 0.694 0.811 0.747 0.883 0.871
500 0.556 0.594 0.633 0.714 0.794 0.755 0.898 0.861
600 0.573 0.568 0.638 0.735 0.776 0.763 0.913 0.852
700 0.590 0.543 0.644 0.755 0.758 0.770 0.928 0.843
800 0.607 0.517 0.650 0.775 0.740 0.778 0.942 0.833
900 0.624 0.492 0.656 0.795 0.722 0.786 0.958 0.823
1000 0.641 0.466 0.661 0.815 0.704 0.794 0.973 0.813
The recall is a metric commonly used to evaluate
Fig.13 demonstrates the Evaluation of precision in different classification models such as the CMLM. It measures the
level of inputs. In a training (Tr) comparison region, the fraction of all positives (e.g., cardiac arrest babies) correctly
proposed CMLA reached 0.898 precision value. Meanwhile classified as positive by the model. To calculate recall, we
the DDCA reached 0.556, HFAF reached 0.633 and RFHF must first identify the true positives (TP) and false negatives
reached 0.794 precision values respectively. In a testing (Ts) (FN):
comparison region, the proposed CMLA reached 0.861
precision value. Meanwhile the DDCA reached 0.594, HFAF • TP = the number of babies the CMLM correctly
reached 0.714 and RFHF reached 0.755 precision values. identified as having a cardiac arrest,
• FN = the number of babies the CMLM incorrectly
H. Computation of Recall (R) labeled as not having a cardiac arrest.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

The recall score measures the model’s ability to identify true the proposed and existing models. In that Tr represents the
positives. It thus indicates the CMLM’s reliability for training data and Ts represents the testing data.
accurately detecting cardiac arrest in newborn babies in the
CICU. Table.10 shows the various recall evaluations between

TABLE X
EVALUATION OF RECALL
No.of DDCA DDCA HFAF HFAF RFHF RFHF CMLA CML
Inputs (Tr) (Ts) (Tr) (Ts) (Tr) (Ts) (Tr) A (Ts)
100 0.472 0.676 0.594 0.515 0.840 0.536 0.807 0.865
200 0.489 0.651 0.598 0.531 0.823 0.541 0.821 0.856
300 0.506 0.625 0.604 0.549 0.805 0.547 0.835 0.846
400 0.522 0.601 0.609 0.565 0.788 0.553 0.849 0.838
500 0.539 0.576 0.615 0.581 0.771 0.559 0.863 0.828
600 0.556 0.551 0.620 0.597 0.754 0.564 0.878 0.819
700 0.572 0.526 0.626 0.613 0.736 0.570 0.892 0.809
800 0.589 0.501 0.631 0.630 0.719 0.576 0.906 0.800
900 0.605 0.477 0.637 0.645 0.702 0.581 0.920 0.792
1000 0.621 0.452 0.642 0.663 0.684 0.587 0.934 0.782
compute the F1 score for the CMLM, we need to first look at
the precision and recall of our model. Precision is the fraction
of true positives (TP) out of all positives (TP+FP). The recall
is the fraction of true positives out of all actual positives
(TP+FN). In our case, TP is the number of cardiac arrests our
model correctly identifies, FP is the number of false arrests,
and FN is the number of cardiac arrests our model fails to
identify. The F1 score can range from 0 to 1, with 0
indicating a terrible model and 1 indicating a perfect model.
The higher the F1 score, the better the model identifies
cardiac arrests. The final F1 score can be used to compare
different models and determine the best identification of
cardiac arrests in the CICU.

FIGURE 14. Evaluation of recall

Fig.14 demonstrates the Evaluation of recall in different level

of inputs. In a training (Tr) comparison region, the proposed
CMLA reached 0.863 recall value. Meanwhile the DDCA
reached 0.539, HFAF reached 0.615 and RFHF reached
0.771 recall values respectively. In a testing (Ts) comparison
region, the proposed CMLA reached 0.828 recall value.
Meanwhile the DDCA reached 0.576, HFAF reached 0.581
and RFHF reached 0.559 recall values.
FIGURE 15. Evaluation of F1-score
I. Computation of F1-Score (F1)
The F1 score measures the accuracy of a model's prediction Table.11 shows the various F1-score evaluations between the
by combining precision and recall. Precision is the number of proposed and existing models. In that Tr represents the
true positives divided by the total number of positives, while training data and Ts represents the testing data.
recall is the number of true positives divided by the total
number of actual positives. The higher the F1 score, the
better the model can correctly identify true positives. To

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

TABLE XI
EVALUATION OF F1-SCORE
No.of DDCA DDCA HFAF HFAF RFHF RFHF CMLA CML
Inputs (Tr) (Ts) (Tr) (Ts) (Tr) (Ts) (Tr) A (Ts)
100 0.406 0.580 0.511 0.443 0.828 0.461 0.791 0.848
200 0.420 0.559 0.514 0.457 0.811 0.465 0.805 0.840
300 0.434 0.537 0.519 0.472 0.794 0.470 0.819 0.830
400 0.449 0.516 0.524 0.486 0.777 0.475 0.832 0.821
500 0.463 0.494 0.529 0.499 0.760 0.481 0.846 0.811
600 0.477 0.474 0.533 0.513 0.743 0.486 0.861 0.803
700 0.491 0.452 0.539 0.528 0.726 0.491 0.875 0.794
800 0.506 0.431 0.543 0.542 0.709 0.496 0.888 0.785
900 0.520 0.410 0.548 0.555 0.692 0.500 0.902 0.776
1000 0.534 0.388 0.552 0.570 0.674 0.505 0.916 0.766
ICU data to identify trends or patterns that may indicate a
Fig.15 demonstrates the Evaluation of F1-score in different higher risk of cardiac arrest. Using statistical models, medical
level of inputs. In a training (Tr) comparison region, the teams can identify potential cases of cardiac arrest early and
proposed CMLA reached 0.846 F1-score value. Meanwhile intervene quickly to provide the necessary care and
the DDCA reached 0.463, HFAF reached 0.529 and RFHF treatment.
reached 0.760 F1-score values respectively. In a testing (Ts)
comparison region, the proposed CMLA reached 0.811 F1- ì n ü
score value. Meanwhile the DDCA reached 0.494, HFAF ïï Ai ( x) ïï å
reached 0.499 and RFHF reached 0.481 F1-score values. C = í i =n1 ý (20)
ï å Z i ( x) ï
VI. RESULTS AND DISCUSSION ïî i =1 ïþ
Statistical models can identify potential cardiac arrest cases
in cardiac intensive care unit newborn babies. These models
Where, C represents the convergence of performance, ‘i’is
use data from monitoring devices, such as heart rate, blood
the initial input level, ‘n’ is the final input level. ‘A’
pressure, and oxygen saturation, to detect abnormalities that
represents the average value of the performance parameter
may indicate a cardiac arrest. If a potential case is identified,
and ‘Z’ represents the total number of training or testing
the medical team can intervene quickly to provide the
cycles. Table.8 shows the various convergence of
necessary care and treatment. Statistical models can detect
performance between the proposed and existing models. In
changes in vital signs over time and identify any risk factors
that Tr represents the training data and Ts represents the
that may increase the likelihood of cardiac arrest in a
testing data.
newborn. In addition, these models can be used to analyze

TABLE VIII
EVALUATION OF CONVERGENCE OF PERFORMANCE
Paramet DDCA DDCA HFAF HFAF RFHF RFHF CMLA CMLA
ers (Tr) (Ts) (Tr) (Ts) (Tr) (Ts) (Tr) (Ts)
Δp 0.0528 0.05215 0.05695 0.0649 0.0615 0.0779 0.09195 0.0891
FDR 0.06833 0.06749 0.0737 0.08399 0.07958 0.08795 0.09015 0.08735
FOR 0.07716 0.07623 0.08323 0.07936 0.07844 0.07565 0.00438 0.00574
Pth 0.06625 0.06546 0.07158 0.06826 0.07732 0.06508 0.08665 0.08396
CSI 0.05688 0.05621 0.06157 0.05872 0.07622 0.05598 0.08495 0.08231
A 0.04362 0.04491 0.04913 0.056 0.06063 0.0672 0.09239 0.08739
P 0.05644 0.05813 0.06358 0.07246 0.07846 0.07588 0.09057 0.08568
R 0.05471 0.05636 0.06176 0.05889 0.07622 0.05614 0.08705 0.08235
F1 0.047 0.04841 0.05312 0.05065 0.07514 0.0483 0.08535 0.08074

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

success index values, 0.056 accuracy, 0.07246 precision,

0.05889 recall and 0.05065 F1-score and RFHF reached
0.0779 delta-P value, 0.08795 false discovery rate, 0.07565
false omission rate, 0.06508 prevalence threshold, 0.05598
critical success index values, 0.0672 accuracy, 0.07588
precision, 0.05614 recall and 0.0483 F1-score respectively.
Hence mean value of the performance parameters has shown
in the following eq.21.
M = Ctr - CTs (21)

Where, M represents the mean value, Ctr represents the

convergence of performance training data and Cts represents
the convergence performance of the testing data. Table.9
expresses the mean value of the performance parameters

TABLE VIII
FIGURE 12. Convergence of performance EVALUATION OF MEAN OF PERFORMANCE
Parameters DDCA HFAF RFHF CMLA
Fig.12 demonstrates the convergence of performance Δp 0.00065 -0.00795 -0.0164 0.00285
between existing and proposed models. In a training (Tr) FDR 0.00084 -0.01029 -0.00837 0.0028
comparison region, the proposed CMLA reached 0.09195 FOR 0.00093 0.00387 0.00279 -0.00136
delta-P value, 0.09015 false discovery rate, 0.00438 false Pth 0.00079 0.00332 0.01224 0.00269
omission rate, 0.08665 prevalence threshold, 0.08495 critical CSI 0.00067 0.00285 0.02024 0.00264
success index values, 0.09239 accuracy, 0.09057 precision, A -0.00129 -0.00687 -0.00657 0.005
0.08705 recall and 0.08535 f1-score. Meanwhile the DDCA P -0.00169 -0.00888 0.00258 0.00489
reached 0.0528 delta-P value, 0.06833 false discovery rate, R -0.00165 0.00287 0.02008 0.0047
0.07716 false omission rate, 0.06625 prevalence threshold, F1 -0.00141 0.00247 0.02684 0.00461
0.05688 critical success index values, 0.04362 accuracy,
0.05644 precision, 0.05471 recall, 0.047 F1-Score, HFAF
reached 0.05695 delta-P value, 0.0737 false discovery rate,
0.08323 false omission rate, 0.07158 prevalence threshold,
0.06157 critical success index values, 0.04913 accuracy,
0.06358 precision, 0.06176 recall and 0.05312 F1-Score and
RFHF reached 0.0615 delta-P value, 0.07958 false discovery
rate, 0.07844 false omission rate, 0.07732 prevalence
threshold, 0.07622 critical success index values, 0.06063
accuracy, 0.07846 precision, 0.07622 recall and 0.07514 F1-
Score respectively. In a testing (Ts) comparison region, the
proposed CMLA reached 0.0891delta-P value, 0.08735 false
discovery rate, 0.00574 false omission rate, 0.08396
prevalence threshold, 0.08231 critical success index values,
0.08739 accuracy, 0.08568 precision, 0.08235 recall and
0.08074 F1-score. Meanwhile the DDCA reached 0.05215
delta-P value, 0.06749 false discovery rate, 0.07623 false
omission rate, 0.06546 prevalence threshold, 0.05621 critical FIGURE 13. Mean values of the performance parameters
success index values 0.04491 accuracy, 0.05813 precision,
0.05636 recall and 0.04841 F1-score, HFAF reached 0.0649 The proposed CMLA reached 0.00285 mean delta-P value
delta-P value, 0.08399 false discovery rate, 0.07936 false (m-Δp), 0.0028 mean false discovery rate (m-FDR), -0.00136
omission rate, 0.06826 prevalence threshold, 0.05872 critical mean false omission rate (m-FOR), 0.00269 mean prevalence

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

threshold (m-Pth), 0.00264 mean critical success index (m- provide early intervention that may help to avert a tragic
CSI) values, 0.005 accuracy, 0.00489 precision, 0.0047 recall outcome. Early detection of cardiac arrest can also reduce the
and 0.00461 F1-score. Meanwhile the DDCA reached amount of time an infant spends in the CICU, helping to
0.00065 mean delta-P value (m-Δp), 0.00084 mean false reduce costs and improve outcomes. Future enhancements of
discovery rate (m-FDR), 0.00093 mean false omission rate the proposed model will focus on using real-time data to
(m-FOR), 0.00079 mean prevalence threshold (m-Pth), identify critical indicators of cardiac arrest. It can involve
0.00067 mean critical success index (m-CSI) values, - collecting various data types such as heart rate, breathing
0.00657 accuracy, 0.00258 precision, 0.02008 recall and rate, temperature, and other physiological measures. The
0.02684 F1-score, HFAF reached -0.00795 mean delta-P cardiac machine learning algorithms can then be used to
value (m-Δp), -0.01029 mean false discovery rate (m-FDR), analyze this data to develop models that can accurately
0.00387 mean false omission rate (m-FOR), 0.00332 mean
predict the likelihood of cardiac arrest. The proposed model
prevalence threshold (m-Pth) and 0.00285 mean critical
can then be used to alert medical staff in order to allow for
success index (m-CSI) values, -0.00687 accuracy, -0.00888
earlier and more effective interventions. Future
precision, 0.00287 recall and 0.00247 F1-score and RFHF
reached -0.0164 mean delta-P value (m-Δp), -0.00837 mean enhancements may also include using artificial intelligence to
false discovery rate (m-FDR), 0.00279 mean false omission detect patterns in the data and make more accurate
rate (m-FOR), 0.01224 mean prevalence threshold (m-Pth) predictions. It could incorporate data from other sources,
0.02024 mean critical success index (m-CSI) values, - such as previous records and medical histories. Finally, these
0.00129 accuracy, -0.00169 precision, -0.00165 recall and - models could be used to develop personalized interventions
0.00141 F1-score respectively. While compared with the for individual patients, allowing for more effective
existing models, the proposed CMLA model achieved better treatments. Enhancing the proposed machine learning
convergence and mean values. The Cardiac Machine algorithm could also pave the way for predicting potential
Learning Model for Early Prediction of Sepsis in the ICU is complications in fetuses or newborns. A healthcare team can
an innovative predictive model that can accurately identify determine risk levels for specific cardiac abnormalities
sepsis in Intensive Care Units (ICUs) before it becomes life- before a baby is even born, which helps provide better
threatening. This model uses a combination of supervised interventions during the prenatal period. In addition, the
and unsupervised machine learning algorithms to identify proposed machine learning algorithm could be used to
sepsis in ICU patients at an early stage. The model uses improve diagnostics and treatments. By studying historical
clinical data, such as vital signs and laboratory parameters, patient data, diagnostics can be improved, and doctors can be
along with patient notes to detect sepsis accurately. This data presented with more accurate and up-to-date information
is then used to create a predictive model that can accurately when diagnosing a patient. It can lead to earlier interventions,
identify sepsis at an early stage. The model can also identify better patient outcomes, and more cost-effective treatments.
subtle changes in patients’ conditions that could indicate
sepsis, allowing for early intervention and treatment. The
Cardiac Machine Learning Model for Early Prediction of
Sepsis in the ICU using Machine Learning is an innovative REFERENCES
and accurate predictive model that can save lives by [1] Choi, E., Schuetz, A., Stewart, W. F., & Sun, J. (2017). Using
detecting sepsis before it becomes life-threatening. It is a recurrent neural network models for early detection of heart failure
highly effective tool for doctors and nurses in ICUs looking onset. Journal of the American Medical Informatics Association,
24(2), 361-370.
to quickly and accurately detect sepsis in their patients [2] Johnson, K. W., Torres Soto, J., Glicksberg, B. S., Shameer, K.,
Miotto, R., Ali, M., ... & Dudley, J. T. (2018). Artificial intelligence
VII. CONCLUSION in cardiology. Journal of the American College of Cardiology,
The proposed machine learning-based statistical model is 71(23), 2668-2679.
[3] Yu, C. M., Wang, L. I., Chau, E., Chan, R. H. W., Kong, S. L., Tang,
essential for the early detection of cardiac arrest in newborn M. O., ... & Lau, C. P. (2005). Intrathoracic impedance monitoring in
babies in the Cardiac Intensive Care Unit (CICU) because patients with heart failure: correlation with fluid status and feasibility
they enable the efficient and accurate identification of infants of early warning preceding hospitalization. Circulation, 112(6), 841-
848.
at high risk of cardiac arrest. Machine learning models can [4] Bonafide, C. P., Localio, A. R., Roberts, K. E., Nadkarni, V. M.,
accurately identify subtle changes in vital signs, such as heart Weirich, C. M., & Keren, R. (2014). Impact of rapid response system
and respiration rates, that may indicate an impending cardiac implementation on critical deterioration events in children. JAMA
pediatrics, 168(1), 25-33.
arrest. In a training (Tr) comparison region, the proposed [5] Bernard, S. A., Gray, T. W., Buist, M. D., Jones, B. M., Silvester,
CMLA reached 0.912 delta-p value, 0.894 FDR value, 0.076 W., Gutteridge, G., & Smith, K. (2002). Treatment of comatose
FOR value, 0.859 prevalence threshold value and 0.842 CSI survivors of out-of-hospital cardiac arrest with induced hypothermia.
value. In a testing (Ts) comparison region, the proposed New England journal of medicine, 346(8), 557-563.
[6] Attia, Z. I., Noseworthy, P. A., Lopez-Jimenez, F., Asirvatham, S. J.,
CMLA reached 0.896 delta-p values, 0.878 FDR value, Deshmukh, A. J., Gersh, B. J., ... & Friedman, P. A. (2019). An
0.061 FOR value, 0.844 prevalence threshold values and artificial intelligence-enabled ECG algorithm for the identification of
0.827 CSI value. The proposed cardiac machine learning patients with atrial fibrillation during sinus rhythm: a retrospective
analysis of outcome prediction. The Lancet, 394(10201), 861-867.
model to identify at-risk infants, healthcare providers can

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

[7] Rajkomar, A., Oren, E., Chen, K., Dai, A. M., Hajaj, N., Hardt, M., [24] Gaies, M. G., Gurney, J. G., Yen, A. H., Napoli, M. L., Gajarski, R.
... & Dean, J. (2018). Scalable and accurate deep learning with J., Ohye, R. G., ... & Hirsch, J. C. (2010). Vasoactive–inotropic score
electronic health records. NPJ digital medicine, 1(1), 18. as a predictor of morbidity and mortality in infants after
[8] Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, cardiopulmonary bypass. Pediatric critical care medicine, 11(2), 234-
P. A., ... & Jodoin, P. M. (2018). Deep learning techniques for 238.
automatic MRI cardiac multi-structures segmentation and diagnosis: [25] Shah, S. J., Katz, D. H., Selvaraj, S., Burke, M. A., Yancy, C. W.,
is the problem solved?. IEEE transactions on medical imaging, Gheorghiade, M., ... & Deo, R. C. (2015). Phenomapping for novel
37(11), 2514-2525. classification of heart failure with preserved ejection fraction.
[9] Pollard, T. J., Johnson, A. E., Raffa, J. D., Celi, L. A., Mark, R. G., Circulation, 131(3), 269-279.
& Badawi, O. (2018). The eICU Collaborative Research Database, a [26] Lee, D. S., Gona, P., Vasan, R. S., Larson, M. G., Benjamin, E. J.,
freely available multi-center database for critical care research. Wang, T. J., ... & Levy, D. (2009). Relation of disease pathogenesis
Scientific data, 5(1), 1-13. and risk factors to heart failure with preserved or reduced ejection
[10] Christodoulou, E., Ma, J., Collins, G. S., Steyerberg, E. W., fraction: insights from the framingham heart study of the national
Verbakel, J. Y., & Van Calster, B. (2019). A systematic review heart, lung, and blood institute. Circulation, 119(24), 3070-3077.
shows no performance benefit of machine learning over logistic [27] Saeed, M., Villarroel, M., Reisner, A. T., Clifford, G., Lehman, L.
regression for clinical prediction models. Journal of clinical W., Moody, G., ... & Mark, R. G. (2011). Multiparameter Intelligent
epidemiology, 110, 12-22. Monitoring in Intensive Care II (MIMIC-II): a public-access
[11] Huang, J., Osorio, C., & Sy, L. W. (2019). An empirical evaluation intensive care unit database. Critical care medicine, 39(5), 952.
of deep learning for ICD-9 code assignment using MIMIC-III [28] Lee, D. S., Austin, P. C., Rouleau, J. L., Liu, P. P., Naimark, D., &
clinical notes. Computer methods and programs in biomedicine, 177, Tu, J. V. (2003). Predicting mortality among patients hospitalized for
141-153. heart failure: derivation and validation of a clinical model. Jama,
[12] Kalscheur, M. M., Kipp, R. T., Tattersall, M. C., Mei, C., Buhr, K. 290(19), 2581-2587.
A., DeMets, D. L., ... & Page, C. D. (2018). Machine learning [29] Gianfrancesco, M. A., Tamang, S., Yazdany, J., & Schmajuk, G.
algorithm predicts cardiac resynchronization therapy outcomes: (2018). Potential biases in machine learning algorithms using
lessons from the COMPANION trial. Circulation: Arrhythmia and electronic health record data. JAMA internal medicine, 178(11),
Electrophysiology, 11(1), e005499. 1544-1547.
[13] Krittanawong, C., Zhang, H., Wang, Z., Aydar, M., & Kitai, T. [30] Moor, M., Rieck, B., Horn, M., Jutzeler, C. R., & Borgwardt, K.
(2017). Artificial intelligence in precision cardiovascular medicine. (2021). Early prediction of sepsis in the ICU using machine learning:
Journal of the American College of Cardiology, 69(21), 2657-2664. a systematic review. Frontiers in medicine, 8, 607952.
[14] Miotto, R., Wang, F., Wang, S., Jiang, X., & Dudley, J. T. (2018). [31] Deo, R. C. (2015). Machine learning in medicine. Circulation,
Deep learning for healthcare: review, opportunities and challenges. 132(20), 1920-1930.
Briefings in bioinformatics, 19(6), 1236-1246. [32] https://www.kaggle.com/datasets/andrewmvd/heart-failure-clinical-
[15] Atkins, D. L., Everson-Stewart, S., Sears, G. K., Daya, M., Osmond, data.
M. H., Warden, C. R., ... & Resuscitation Outcomes Consortium [33] Shankar, K., Perumal, E., Elhoseny, M., Taher, F., Gupta, B. B., &
Investigators. (2009). Epidemiology and outcomes from out-of- El-Latif, A. A. A. (2021). Synergic deep learning for smart health
hospital cardiac arrest in children: the Resuscitation Outcomes diagnosis of COVID-19 for connected living and smart cities. ACM
Consortium Epistry–Cardiac Arrest. Circulation, 119(11), 1484- Transactions on Internet Technology (TOIT), 22(3), 1-14.
1491. [34] Hashash, O., Sharafeddine, S., & Dawy, Z. (2021, June). MEC-based
[16] Nadkarni, V. M., Larkin, G. L., Peberdy, M. A., Carey, S. M., Kaye, energy-aware distributed feature extraction for mHealth applications
W., Mancini, M. E., ... & National Registry of Cardiopulmonary with strict latency requirements. In 2021 IEEE International
Resuscitation Investigators. (2006). First documented rhythm and Conference on Communications Workshops (ICC Workshops) (pp.
clinical outcome from in-hospital cardiac arrest among children and 1-6). IEEE.
adults. Jama, 295(1), 50-57. [35] Chaudhary, P., Varshney, Y. V., Srivastava, G., & Bhatia, S. (2022).
[17] Chan, P. S., Jain, R., Nallmothu, B. K., Berg, R. A., & Sasson, C. Motor imagery classification using sparse nonnegative matrix
(2010). Rapid response teams: a systematic review and meta- factorization and convolutional neural networks. Neural Computing
analysis. Archives of internal medicine, 170(1), 18-26. and Applications, 1-11.
[18] Shickel, B., Tighe, P. J., Bihorac, A., & Rashidi, P. (2017). Deep [36] Bhatti, M. H., Khan, J., Khan, M. U. G., Iqbal, R., Aloqaily, M.,
EHR: a survey of recent advances in deep learning techniques for Jararweh, Y., & Gupta, B. (2019). Soft computing-based EEG
electronic health record (EHR) analysis. IEEE journal of biomedical classification by optimal feature selection and neural networks. IEEE
and health informatics, 22(5), 1589-1604. Transactions on Industrial Informatics, 15(10), 5747-5754.
[19] Kilgannon, J. H., Jones, A. E., Shapiro, N. I., Angelos, M. G., [37] Luo, C., Zhu, Y., Zhu, Z., Li, R., Chen, G., & Wang, Z. (2022). A
Milcarek, B., Hunter, K., ... & Emergency Medicine Shock Research machine learning-based risk stratification tool for in-hospital
Network (EMShockNet) Investigators. (2010). Association between mortality of intensive care unit patients with heart failure. Journal of
arterial hyperoxia following resuscitation from cardiac arrest and in- Translational Medicine, 20(1), 136.
hospital mortality. Jama, 303(21), 2165-2171. [38] Winslow, C. J., Edelson, D. P., Churpek, M. M., Taneja, M., Shah,
[20] Abdelghany, M., Abdelhamid, M., Allam, A., El Etriby, A., Hafez, N. S., Datta, A., ... & Halasyamani, L. K. (2022). The impact of a
S., Ragy, H., & Sobhy, M. (2023). Detection and Diagnosis of machine learning early warning score on hospital mortality: a
Cardiac Amyloidosis in Egypt. Cardiology and Therapy, 1-17. multicenter clinical intervention trial. Critical Care Medicine, 50(9),
[21] Carlisle, M. A., Fudim, M., DeVore, A. D., & Piccini, J. P. (2019). 1339-1347.
Heart failure and atrial fibrillation, like fire and fury. JACC: Heart [39] Ahsan, M. M., & Siddique, Z. (2022). Machine learning-based heart
Failure, 7(6), 447-456. disease diagnosis: A systematic literature review. Artificial
[22] Yaku, H., Kato, T., Morimoto, T., Inuzuka, Y., Tamaki, Y., Ozasa, Intelligence in Medicine, 102289.
N., ... & Kimura, T. (2020). Risk factors and clinical outcomes of [40] Rooney, S. R., Reynolds, E. L., Banerjee, M., Pasquali, S. K.,
functional decline during hospitalisation in very old patients with Charpie, J. R., Gaies, M. G., & Owens, G. E. (2022). Prediction of
acute decompensated heart failure: an observational study. BMJ extubation failure in the paediatric cardiac ICU using machine
open, 10(2), e032674. learning and high-frequency physiologic data. Cardiology in the
[23] Fonarow, G. C., Adams, K. F., Abraham, W. T., Yancy, C. W., Young, 32(10), 1649-1656.
Boscardin, W. J., & ADHERE Scientific Advisory Committee.
(2005). Risk stratification for in-hospital mortality in acutely
decompensated heart failure: classification and regression tree Acknowledgement: The authors would like to thank Tejankar
analysis. Jama, 293(5), 572-580. Healthcare and Medical Research Institute Pvt. Ltd for funding this
research project.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3286346

ACCESS, Wireless Networks (Springer), EURASIP Journal on Wireless

Communications and Networking (Springer), Wireless Communications
and Mobile Computing (Hindawi), Sensors (MDPI), and Future Internet
(MDPI), to name a few.

Ketan Gupta is PMP certified with 10+ years Mohammad Alibakhshikenari (Member,
of experience as a professional and research IEEE) was born in Mazandaran, Iran, in February
scholar in Information Technology, Supply 1988. He received the Ph.D. degree (Hons.) with
Chain, Medical, and Healthcare domains. He European Label in electronics engineering from
completed Bachelor's in Pharmaceutical Sciences the University of Rome ‘‘Tor Vergata’’, Italy, in
in India, followed by three master’s degrees – February 2020. He was a Ph.D. Visiting
MBA/ PGDM International Business, India, MS Researcher at the Chalmers University of
Supply Chain Management, United States, and Technology, Sweden, in 2018. His training
MS Information Technology with Honors, during the Ph.D. included a research stage in the
United States. Ketan is pursuing a Ph.D. in Swedish company Gap Waves AB. He is
Information Technology at the University of the currently with the Department of Signal Theory
Cumberlands, United States, focusing on and Communications, Universidad Carlos III de
Artificial Intelligence and Machine Learning tools in the medical and Madrid (uc3m), Spain, as the Principal Investigator of the CONEX
healthcare domains. He has 20+ research papers published in renowned (CONnecting EXcellence)-Plus Talent Training Program and Marie
journals as a research scholar. Ketan is also an active peer reviewer for Skłodowska-Curie Actions. He was also a Lecturer of the electromagnetic
scholarly journals with excellent ratings on WOS. Currently, Ketan is fields and electromagnetic laboratory with the Department of Signal
working as a Program Manager in the company Meta (formerly Theory and Communications for academic year 2021–2022 and he
Facebook), where he is accountable for enriching CBOM management, received the “Teaching Excellent Acknowledgement” Certificate for the
ERP system architecture implementation, and research model design, course of electromagnetic fields from Vice-Rector of studies of uc3m.
including AI, ML, ANN, and Deep Learning. His primary accountability is Now he is spending an industrial research period in SARAS Technology
equally towards product pricing, risk opportunities identification, cost- Limited Company, located in Leeds, United Kingdom, which is defined as
benefit analysis, supplier plan of record, and data analysis using data his secondment plan by CONEX-Plus Program and Marie Skłodowska-
visualization for Augmented Reality (AR), Virtual Reality (VR), and Curie Actions. His research interests include electromagnetic systems,
Portal Products. Ketan’s strong inclination and perspicacity toward antennas and wave-propagations, metamaterials and metasurfaces,
research in the medical and healthcare sector are recognized in addressing synthetic aperture radars (SAR), multiple input multiple output (MIMO)
the current IT challenges and community welfare. systems, RFID tag antennas, substrate integrated waveguides (SIWs),
impedance matching circuits, microwave components, millimeter-waves
Nasmin Jiwani is an IT professional and a and terahertz integrated circuits, gap waveguide technology, beamforming
Second Year Ph.D. student at the University of matrix, and reconfigurable intelligent surfaces (RIS). He was a recipient of
The Cumberlands, USA. Her research explores the three years research grant funded by Universidad Carlos III de Madrid
building machine learning algorithms to and the European Union’s Horizon 2020 Research and Innovation
determine existing and future healthcare issues. Program under the Marie Skłodowska-Curie Grant started in July 2021,
She has 10+ years of professional experience as the two years research grant funded by the University of Rome ‘‘Tor
an IT Project Manager managing various Vergata’’ started in November 2019, the three years Ph.D. Scholarship
complex and large projects in the Pharmaceutical funded by the University of Rome ‘‘Tor Vergata’’ started in November
Life science, Healthcare, Insurance, and Supply 2016, and the two Young Engineer Awards of the 47th and 48th European
Chain sector. She also holds two master's degrees Microwave Conference were held in Nuremberg, Germany, in 2017, and
in ISEM and Project Management with a PMP, in Madrid, Spain, in 2018, respectively. His research article entitled
CSPO, and CSM certification with 25+ research ‘‘High-Gain Metasurface in Polyimide On-Chip Antenna Based on CRLH-
papers published in renowned journals and books. Her current role at a TL for Sub Terahertz Integrated Circuits’’ published in Scientific Reports
Pharma and Life science company called Viatris Pharmaceutical, formally was awarded as the Best Month Paper at the University of Bradford, U.K.,
known as Pfizer, allows her to implement her research in healthcare and in April 2020. He is serving as an Associate Editor for (i) Radio Science,
Machine learning knowledge to determine various aspects of applying IT and (ii) IET Journal of Engineering. He also acts as a referee in several
to improve drug trials. Her most recent research explores how technology highly reputed journals and international conferences.
plays symbolic and practical roles in nature's exploitation, imitation,
reproduction, and "healing." She plans to continue her research in the
human development and safety field from external environmental attacks
from viruses and diseases.

Giovanni Pau (Member, IEEE) received the

bachelor’s degree in telematic engineering from
the University of Catania, Italy, and the master’s
(cum Laude) and Ph.D. degrees in telematic engi-
neering from Kore University of Enna, Italy. He
is currently an Associate Professor at the Faculty
of Engineering and Architecture, Kore University
of Enna. He is the author/coauthor of more than
80 refereed papers published in journals and con-
ference proceedings. His research interests
include wireless sensor networks, fuzzy logic
controllers, intelligent transportation systems, the
Internet of Things, smart homes, and network security. He has been
involved in several international conferences as a session co-chair and a
technical program committee member. He serves/served as a leading guest
editor in special issues of several international journals. He is an Editorial
Board Member and an Associate Editor of several journals, such as IEEE

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/