IMBALANCED DATASET EFFECT ON CNN-BASED CLASSIFIER PERFORMANCE FOR FACE RECOGNITION

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.15, No.1, January 2024
DOI:10.5121/ijaia.2024.15101 1
A COMPREHENSIVE SYSTEMATIC REVIEW FOR
CARDIOVASCULAR DISEASE USING MACHINE
LEARNING TECHNIQUES
Islam D. S. Aabdalla1
and Dr.D.Vasumathi2
1
PhD. scholar of CSE JNTUH University
2
Professor of CSE JNTUH University
ABSTRACT
The global upswing in cardiovascular disease (CVD) cases presents a critical challenge. While the
ultimate goal remains elusive, improving CVD prediction accuracy is vital. Machine learning and deep
learning are crucial for decoding complex health data, enhancing cardiac imaging, and predicting disease
outcomes in clinical practice. This systematic literature review meticulously analyses CVD using machine
learning techniques, with a particular emphasis on algorithms for classification and prediction. The meta-
analysis covers 343 references from 2020 to November 2023, preceding a thorough examination of 65
selected references. Acknowledging current hurdles in CVD classification methods that impede practical
use, this systematic literature review (SLR) is conducted.
The study provides valuable insights for researchers and healthcare professionals, facilitating the
integration of clinical applications in machine learning settings related to CVD. It also aids in promptly
identifying potential threats and implementing precautionary measures. The study also recognizes
prevalent classical machine learning methods, emphasizing their clinically relevant diagnostic outcomes.
Deliberating on current trends, algorithms, and potential areas for future research offers a comprehensive
insight into the present state of affairs.
KEYWORDS
Machine learning, Classification,CVD, Deep learning, ECG, Disease classification.
1. INTRODUCTION
Cardiovascular diseases (CVD), encompassing coronary heart disease (CHD), cerebrovascular
disorders, peripheral vascular ailments, rheumatism-induced heart disease, and congenital heart
disease, pose a significant global threat, contributing to widespread mortality [1]. The
anticipation of these illnesses presents a formidable challenge. While individuals may experience
variations in their vascular pressure and heart rate, the average heart rate generally varies between
60 and 100 heartbeats per minute, and the average blood pressure is approximately 120/80. Any
disruption to the average blood circulation in the heart can result in severe consequences, referred
to as cardiovascular diseases [2]. According to reports from the World Health Organization
(WHO), approximately 17.9 million deaths occur annually. They are attributed to heart-related
complications, with over four-fifths of cardiovascular disease (CVD) deaths linked to heart
attacks and strokes. Unhealthy lifestyle choices, such as poor dietary habits, insufficient physical
activity, excessive alcohol consumption, and tobacco use, contribute to an elevated risk of heart-
related complications. CVD is a severe global disorder, and specific medical conditions
significantly increase the likelihood of acquiring cardiovascular conditions, particularly heart
disease [3].Risk factors include hypertension, a family history of CVD, low levels of HDL (good)

2
cholesterol, a high-fat diet, and elevated levels of LDL cholesterol. As individuals age, the
probability of developing cardiac disease increases, making it a more common occurrence [4].
Early detection is crucial for saving lives, and identifying individuals at an elevated risk of CVD
in its initial phases can prevent unexpected and premature deaths. There is a growing interest in
employing ML methodologies for cardiovascular disease classification, offering potential
improvements in accuracy and efficiency [2]. Typical attributes for assessing heart disease
include age, gender, fasting blood pressure, chest pain classification, resting electrocardiogram,
number of vessels highlighted by resting blood pressure, serum cholesterol, thalach, ST
depression, pain loc, fasting glucose, exercise-induced angina (exang), smoking status, dietary
habits, body weight, height, and obesity [5].CVD presents a significant challenge, and while
achieving the ultimate goal remains tough, enhancing prediction accuracy is crucial. ML and DL
are essential in decoding complex health data, improving cardiac imaging, and predicting disease
outcomes. This review explores ML techniques for CVD, focusing on classification and
prediction algorithms essential for early detection and effective treatment. ML algorithms find
widespread cardiovascular disease classification applications, extracting valuable features from
datasets[6]. and enhancing the accuracy and efficiency of medical identification. Incorporating
ML into clinical practice poses difficulties, but it discerns complex relationships for precise
predictions, adeptly handles diverse data types for improved accuracy in identifying at-risk
individuals, and enables early intervention. ML-based classification aids clinicians with decision
support systems, contributing to CVD management and diagnosis. Although ECG is used for
CVD diagnosis, visually spotting irregularities requires significant time and effort[7].ML-based
classification revolutionizes diagnostics and forecasting. The review analyzed PubMed,
MEDLINE, and Web of Science datasets, addressing conflicts and evaluating criteria including
clinical significance, patient considerations, algorithms, validation, clinical usefulness, and
efficacy [8]. The analysis extensively examines ML-based CVD classification, emphasizing
accurate classification for early detection and effective treatment [9]. It highlights benefits and
addresses challenges in ML-based cardiovascular disease classification, holding
significant potential for improving diagnosis, treatment, and patient outcomes[10].However,
model performance uncertainty concerning bias, fairness, and intelligibility persists [11].
Figure 1. depicts the evaluation criteria utilized in current literature to assess CVD.
Figure 1 depicts prevalent CVD, backed by studies that emphasize heart arrhythmia and coronary
artery disease as central clusters. Krittanawong et al.[38],employed decision trees and random
forests for cardiovascular disease classification. Recognizing trends, methodologies, gaps, and
ML-based heart disease classification opportunities is vital. The meta-analysis covers 343

3
references from 2020 to November 2023 after thoroughly examining 65 selected references.
Identifying challenges in CVD classification methods hindering practical application, this
systematic literature review (SLR) provides valuable insights for researchers and healthcare
professionals. It aids ML integration in clinical CVD settings, facilitating threat identification and
precautionary measures. The study acknowledges prevalent classical ML methods, emphasizing
clinically relevant diagnostic outcomes. Deliberating on current trends, algorithms, and potential
research areas offers a comprehensive view of the present state. The metadata analysis aims to
address various aspects in the subsequent in-depth analysis papers:
1. Provide an integrated, synthesized overview of the existing knowledge on cardiovascular
disease (CVD) classification techniques.
2. Highlight current methodological strategies and unique perspectives.
3. Describe research insights, identify existing gaps, and propose future research directions,
illustrating the study's significance. Furthermore, it is to offer a comprehensive survey of
the current landscape of ML applications in cardiovascular disease classification. By
examining existing studies' methodologies, datasets, and performance metrics, this
review strives to distill critical insights into the strengths and limitations of ML-based
approaches.
Research Questions:
RQ1: What difficulties exist in cardiovascular disease (CVD) classification research?
RQ2: What are the commonly employed metrics for evaluating performance in CVD
classification?
RQ3: What algorithms are utilized to classify cardiovascular disease (CVD)?
The subsequent sections of the document are structured as follows: Section 2 briefly summarizes
the review methods, followed by the examination of clinical CVD in Section 3. The results and
responses to the research questions are then presented and discussed in Section 4, 5 limitations
concluding with the future work outlined in Section 6.
2. METHODOLOGY
A comprehensive analysis of the literature using systematic review entails creating inquiries and
employing systematic, clear procedures to locate, choose, and critically assess pertinent research.
The aim is to gather and assess data from the studies incorporated in the review [12]. This
approach is chosen for its accurate and reliable synthesis of academic literature, recognized
across various research fields. The systematic literature review (SLR) adheres to the Preferred
Reporting Information for Systematic Reviews and Meta-Analysis (PRISMA) recommendations,
ensuring clarity and transparency in the presentation of systematic literature reviews[7].Our
systematic search encompassed vital databases, including PubMed, IEEE Xplore, and Scopus,
employing a structured set of search terms related to machine learning and cardiovascular disease.
The inclusion criteria span studies published between 2020 and 2023, concentrating on ML
models designed for CVD diagnosis and risk assessment. Evaluating study quality and data
extraction adhered to established guidelines for systematic reviews, ensuring a rigorous and
systematic approach.
2.1. Data Collection
The studies included in this review were gathered from different Internet-based repositories,
including Scopus, Web of Science, PubMed, Google Scholar, and arXiv, from biomedical

4
journals, machine learning journals, and deep learning journals. Additionally, proceedings from
various conferences were considered. The data collection process involved the following steps:
Identification of data identification:
Standard and specific databases relevant to this research were identified, encompassing
biomedical journals, machine learning and deep learning journals, and conference proceedings.
The selection of keywords was based on the research questions, including cardiovascular disease,
ECG, machine learning, classification, and algorithms. The search query used was
"cardiovascular disease" OR "CVD" OR "ECG" OR "classification" OR "machine learning" OR
"algorithms." The search criteria were designed to ensure both accuracy and recentness.
Screening and Determining Eligibility of Initial Data
They commenced the search with specified keywords and uncovered 343 articles. I independently
evaluated the titles and abstracts using a standardized extraction form and resolved conflicts
through discussion. Articles unrelated to machine learning or heart disease were omitted,
including reviews and non-human studies. Following the title and abstract screening, 65 articles
were thoroughly examined, all meeting the inclusion criteria. Figure 3 outlines the exclusion and
inclusion processes employed in this study. Reasons for article exclusion during full-text
screening encompassed a need for more imbalanced data analysis methods, a singular focus on
models' performance without addressing potential limitations, the absence of peer-reviewed
status, and inaccessibility to the full text.
Observations and Findings
Outcomes from analyzing metadata and observations are detailed below, stemming from content
analysis of 65 publications and a metadata study of 343 papers screened for cardiovascular
disease (CVD) classification relevance. Varied CVD patterns pose challenges for universal
models due to inconsistent data quality, noise, and recording variations, impacting algorithm
accuracy. The inclusion and exclusion criteria in Figure 2 guide the process, while Figure 3
shows papers on CVD classification from 2020 to 2023. The investigation into CVD has sparked
interest, with authors proposing integrated methods (Table 4). Analyzing clinical data involves
interpreting information from clinical sources to gain insights into cardiovascular health,
including occurrence, risk factors, diagnosis, treatment, and outcomes. The overall strategy is
shown in Figure 2.

5
Figure 2. Depicts the schematic diagram employed in the review.
Figure 3. Quantity of published papers from 2020 to 2023 in CVD classification
Uses of Cardiovascular Imaging
AI, including machine and deep learning, revolutionizes CVD medicine, particularly cardiology
imaging. Electrocardiography, crucial for early heart disease detection, is swift, painless, and
widely used, revealing CVD lesion pressure gradients cost-effectively. Utilizing acoustic waves,
echo is a safe method, employing 2D, standard Doppler ultrasonography, and 3D ultrasound to
capture heart muscle images, examining the four chambers. A semantic classification model
trained on left ventricle tracings now segments video frames without prior human tracings [4].
Narula et al. [11] demonstrated that algorithms with supervised learning, particularly when
utilizing STE data, could more accurately distinguish between athlete heart disease and
hypertrophic cardiomyopathy than traditional measurement systems. ML models in

6
electrocardiography also hold promise for another application, specifically in heart valve disease
[16]. Moghaddasi et al.[17]Introduced novel features for recognizing micropatterns in ECG
images. The method suggested attains a specificity rate of 99.63% and a sensitivity rate of 99.38%
for detecting the severity of MR.
Additionally, Playford et al. [18].Evaluated AI's ability to estimate aortic valve area using
additional echocardiographic data, excluding the left ventricular outflow tract measurement
requirements, achieving a performance level of 0.95 in recent development, Ouyang et al.
[19].They developed EchoNet-Dynamic, an advanced video-based deep learning algorithm with
over 0.92 accuracy. This 3D convolutional neural network precisely evaluates cardiac function by
segmenting the left ventricle and estimating the ejection fraction from echo videos. It matches or
surpasses the accuracy of human experts. I also introduced CADNet, a sophisticated framework
for automated coronary artery calcification and shadow border detection in IVUS images.
CADNet incorporates Convolutional Block Attention Modules (CBAM) and Atrous Spatial
Pyramid Pooling (ASPP) in an encoder-decoder U-Net structure, effectively identifying vessel
components in the presence of artifacts and complex lesions in IVUS images with up to 360° of
calcification. Experimental results on 1097 IVUS images from 12 patients demonstrate CADNet
efficacy, with mean intersection over union (mean IoU) at 0.7894, dice coefficient at 0.8763,
precision at 0.8768, and recall at 0.8774. Compared to alternative techniques, CADNet excels in
IVUS image segmentation accuracy, showcasing robust calcification and shadow border
detection performance—a promising field advancement [20].AI reduces observer variability and
ensures accurate diagnoses in electrocardiography by automating measurement and image
segmentation of cardiac parameters, potentially revealing clinically meaningful insights [2].AI-
based image quality classification is crucial for accurate diagnosis, particularly within the
cardiovascular domain. An FDA-approved AI-powered echocardiographic device aids users in
capturing high-quality images. However, cardiologist review and approval of acquired images
remain essential for thorough patient assessment[21]. Electrocardiography evolution has
increased parameters and examination complexity, requiring nonspecialists to rely on findings.
Building effective AI models demands a significant amount of labeled echocardiographic data,
and ongoing efforts focus on collecting such data[6].Deep learning models exhibit promise in
accurately predicting ejection fraction, even when trained on mislabeled data, showcasing their
ability to mimic human recognition successfully[22].
Magnetic Resonance Imaging (MRI)
Within heart MRI, the segmentation of ventricles emerges as a promising domain for applying
ML methods. This approach streamlines the measurement of ventricular volumes, contributing to
heightened precision and consistency in clinical evaluations [23].Avendi et al. [24]utilized deep
learning algorithms trained on cardiac MRI data for the automatic classification and segmentation
of the right ventricular chamber to improve algorithmic accuracy. Similarly, various automated
NN has been effectively created for left ventricle classification, especially in cardiac cine MRI.
Another noteworthy use of machine learning in cardiac MRI includes identifying chronic
myocardial or sub acute scars [25]. Dawes et al. employed supervised ML to evaluate systolic
cardiac motions in three dimensions, enabling the prediction of adverse outcomes in individuals
with pulmonary diseases, such as early mortality or right heart failure. This prediction was
independent of conventional risk factors. Further details on cardiac computed tomography were
not provided in the text.
Machine learning (ML) image analysis is increasing in cardiac CT for diagnosing coronary artery
disease and atherosclerosis. This involves techniques like coronary artery calcium scoring and
fractional flow estimation in noninvasive coronary computed tomographic angiography. However,
CCTA often overestimates stenosis severity compared to invasive angiography. Moreover,
angiographic stenosis does not consistently correlate with the hemodynamic significance

7
indicated by fractional flow reserve (FFR). Various ML models[26]enhance CCTA by accurately
reclassifying aerodynamically
-
non significant stenosis.
Regarding coronary plaque characterization, ML models applied to automatic coronary artery
calcium scoring help reduce false positives and minimize interobserver variability. Utilizing
supervised machine learning, a CNN was used to directly detect and quantify coronary artery
calcification (CAC), calculating the Agatston score directly from CT scans without prior
segmentation. In cardiac CT, machine learning has applications in prognosis and myocardial
infarction detection through texture analysis methods. The SMARTool Project introduces an
innovative approach to managing CAD patients, encompassing diagnosis, prognosis, and
treatment based on machine learning risk stratification and computational biomechanics. Machine
learning analysis, utilizing historical and future-oriented data, including clinical details, humoral
information, and CCTA imaging, aims to differentiate individuals at low risk from those at
medium-to-high risk for CAD. The CAD diagnosis involves three-dimensional reconstruction of
coronary arteries and noninvasive estimation of brilliance, with classification relying on intricate
computational models for arterial plaque growth.
Uses of Cardiovascular Electrocardiography(ECG)
The prominence of cardiovascular diseases (CVD) as the benchmark has drawn the attention of
numerous researchers to conduct studies[12]. Classification of cardiovascular disease refers to
categorizing or grouping different types or classes of cardiovascular diseases based on specific
criteria or characteristics [27]. ML or data analysis may involve developing models or algorithms
to classify individuals into specific cardiovascular disease categories, often using features or
patterns extracted from medical data such as ECG, imaging, or clinical records.
Numerous researchers have endeavored to delineate boundaries through various techniques. The
extraction of features from electrocardiograms, proven crucial for detecting cardiovascular
diseases, has garnered significant attention from the research community.
Nevertheless, due to the intrinsic challenges of ECG CVD, Diverse ECG patterns challenge
universal classification models due to inconsistent data quality, noise, and recording variations
across sources, impacting algorithm accuracy and reliability. Developing a comprehensive model
for accurate CVD classification faces challenges, particularly with imbalanced datasets. There is
still room for enhancement in this area of analysis. This classification process can aid in
prognosis, diagnosis, and treatment planning for patients with heart conditions. Recently, there
has been a growing inclination toward utilizing ML algorithms to classify cardiovascular
disease[2]. These methods have demonstrated promising outcomes by enhancing diagnostic
accuracy and efficiency and identifying individuals at high risk of developing CVD. However, it
is crucial to acknowledge that incorporating machine learning algorithms into clinical practice
comes with challenges. These challenges encompass the intricacies of feature extraction
procedures and the requirement for domain expertise in crafting adequate inputs for the
classifier[6].
Evaluating and validating machine learning (ML) approaches with extensive clinical datasets is
crucial for ensuring their reliability and applicability across diverse cases. While ML methods
hold promise in classifying and analyzing cardiovascular disease, further research and validation
are essential for unlocking their benefits and seamlessly integrating them into clinical practices.
A systematic literature review on CVD classification using ML techniques would contribute
valuable insights to healthcare, offering an overview of current research and guiding future
directions. Medical imaging, along with diagnostic electrocardiography, significantly benefits
from ML integration. The ECG, crucial for detecting heart electrical abnormalities, undergoes
streamlined anomaly identification through ML models, particularly in deep learning (DL). This

8
reduces interpretation time and lessens reliance on individual variability. Noteworthy progress
includes Hannum et al.'s 34-layer deep neural network successfully classifying 12 arrhythmia
types, outperforming a dataset annotated by certified cardiologists. Ongoing automation efforts
encompass diagnosis, prognosis, drug design, and testing, using abundant data from medical
imaging and electronic medical records [28]. ML and DL in medicine aim to customize medical
decisions, health practices, and therapies for individual patients[10].
Nevertheless, the existing status of AI in the medical domain is considered promising but needs
more substantial data and evidence [12]. Concerns about bias, privacy, security, transparency,
causality, transferability, informativeness, fairness, and confidence accompany machine learning
in medicine. The direct impact of these systems on human health underscores the urgent need to
comprehend the decision-making processes involved[29]. particularly in areas where disease
diagnosis leads to life-changing outcomes[30]. Precision medicine, where experts seek nuanced
information beyond binary predictions to support diagnoses [29].Represents another crucial area.
Addressing these challenges requires clear explanations for how and why a model generates its
outputs [30]. Thus, issues of explainability and related concepts like interpretability and
transparency have gained significant attention in machine learning in medicine in recent
years[31]. Despite substantial evidence supporting their utility, widespread adoption of machine
learning-based systems in routine medical practice beyond specialized applications is likely only
if these challenges are effectively tackled. A potential solution is having these systems offer
satisfactory explanations for their decisions [8]. To achieve widespread adoption beyond
specialized applications, machine-learning-based systems in routine medical practice require
effective solutions to challenges. A potential remedy is for these systems to explain their
decisions satisfactorily[31].
ResNet led MI detection at 99.99%, followed by CNN at 99.95%. Papers consistently reported
accuracies above 97%, averaging over 93%, with CNN recording the lowest MI detection
accuracy at 78%. This section details the highest, average, and lowest accuracies for each DL
technique in MI localization. CNN achieved the highest accuracy at 99.87%. Limited research on
LSTM for MI localization exists, with only one study each for AE and GRU models. AE and
GRU models share identical maximum and average accuracies, though the minimum for AE
remains unspecified. The remaining three models in MI localization exhibit accuracies below
65%.
Additionally, this section explores the maximum accuracy of DL techniques trained on diverse
ECG datasets with different leads. Generally, networks trained on 12-lead ECG data from the
PTB database outperformed those with fewer leads. Specifically, using 12-lead ECG data from
the PTB database, ResNet achieved the highest accuracy at 99.99% in MI localization. Using
only lead II ECG data yielded favorable outcomes in CRNN, CNN, and AE methods on the PTB
database. However, it is essential to acknowledge that this result may be influenced by three
studies achieving high performance solely in MI detection using lead II ECG data without
achieving comparable results in MI localization. This paper examines various algorithms in CVD
classification, reviewing studies based on utilized algorithms, databases, challenges, and
evaluation metrics, providing a comprehensive explanation for each research question and
corresponding discoveries. A systematic literature review (SLR) involves formulating questions
and employing systematic and explicit procedures to discover and critically assess relevant
research. The objective is to collect and assess data derived from the studies incorporated in the
review[32]. This technique is favored for its accuracy and dependability in consolidating
academic literature, gaining broad acceptance in CVD classification.
RQ1: What challenges exist in the domain of CVD classification?
In summary, the response to this inquiry is that challenges are present in CVD classification.

9
Incorporates different research challenges related to patient profile diversity, data quality
inconsistency, and ECG recording variations. The complexity of cardiovascular conditions poses
a hurdle to creating comprehensive classification models, and imbalanced datasets may lead to
biased models. Additional challenges include ensuring the interpretability of model decisions,
achieving generalization across diverse populations, and integrating models into clinical practice.
Rigorous validation, addressing ethical concerns, and keeping up with technological
advancements further contribute to the complexities in this research area. Table 2 outlines the
drawbacks identified in each paper. The literature review emphasizes a recognized limitation in
the current machine-learning approach for cardiovascular disease (CVD).
Sample of Data Set:Systems often advise small sample sizes for high accuracy. An investigation
reveals that ML algorithms perform poorly with large sample sizes. Conversely, the same
classifier excels with exclusive attributes in compact and efficient techniques, significantly
impacting estimated accuracy.
Clinical Aspect: Some research needs to include monitoring, which leads to insufficient
observation of clinical features, hindering patient data provision for informed treatment decisions.
Data-Clinical Correlation:Inconsistent findings with healthcare practices make interpretation
challenging. Meta-analysis enhances study effectiveness by applying consistent techniques and
addressing the limitations of ML approaches in handling insufficient and inconsistent clinical
data.
Data Distribution: There are no standardized rules for data usage; joint random splitting is
80%:20% or 70%:30%. Limited data sizes for each CVD type affect the combined results.
Feature Selection:Many studies overlook the impact of feature selection on data splitting. A
good approach improves ML model performance and reduces processing costs from
inappropriate input data.
Hyper-parameter optimization: Existing research lacks parameter tuning, hindering system
improvement. Anonymous hyper-parameter customization leads to substantial statistical diversity.
Algorithmic optimization approach: Researchers often present heuristic ML frameworks for
predicting cardiovascular disease, improving accuracy at a minimal computational expense.
Ensemble Technique: Studies overlook ensemble methods for enhancing ML models in
predicting cardiovascular diseases, combining findings into a single score. Researchers and
practitioners increasingly favor DL algorithms over traditional ML for MLBHDD model
development, with 27 out of 49 studies adopting DL-based approaches. For example, Awan et al.
used MLP, achieving 48% sensitivity and 70% specificity for predicting CVD patient
readmission within 30 days[34]. Li et al[13]. Introduced the craftNet DNN model, achieving
86.82% to 89.25% accuracy in recognizing handcrafted features for cardiovascular disease
detection. Kala and Dixit[42]. Employed a 1D CNN model to detect early-stage heart disease,
achieving 93% accuracy using data from 300 actual patients.The second most utilized algorithm
is SVM, with Liu et al. employing it in the CAB approach and achieving 99.71% accuracy for
arrhythmia patients[46]. Rath et al[18]. Proposed a model that combines GAN and LSTM,
accurately detecting heart disease patients with up to 99.4% accuracy Shah et al[44]. Also tested
SVM along with RF, Ordinal Regression, NB, and LR on the Cleveland dataset, where SVM
exhibited the best performance with 95% accuracy.Other algorithms were also explored,
including ensemble learning[38]and linear methods. This section aids scholars in identifying gaps
in recent literature. Recognition of cardiovascular disease (CVD) is crucial due to its life-

10
threatening nature. Table 1 details the limitations of each research paper. Based on the literature
review, the existing machine learning approach for CVD is characterized by one of the following
constraints[33]. Deep learning models enhance risk stratification for cardiovascular diseases
using electronic medical record data and medical images. However, assessing these models goes
beyond statistical success measures, necessitating a framework for clinical utility evaluation[1].
These models should be accurate and transparent about limitations, ensuring understand ability
for healthcare providers. The existing literature lacks powerful, interpretable models for medical
practice, urging future efforts in these aspects for improved patient outcomes. Evaluating deep
learning models in clinical practice requires scrutiny by healthcare providers to ascertain their
clinical relevance [2]. Unlike widely observed trends in machine learning research, the paper
stresses the importance of explainability and interpretability for deep learning models[10]. Other
regression models, like the MAGIC score and the Seattle Heart Failure Model, quantify patient
risk for cardiovascular diseases [6]. Deep learning models show potential for improving risk
stratification for cardiovascular diseases, particularly in medical electronic health records and
imaging. However, evaluating the clinical utility of these models takes time and effort, requiring
accuracy, transparency, and interpretability for healthcare providers. Healthcare providers must
critically assess deep learning models to determine their clinical usefulness and accurate
application to specific patient demographics. The broader literature needs more powerful,
explainable models for medical practice, necessitating future work to develop impactful models.
Implementing deep learning models in clinical practice necessitates a comprehensive strategy,
including assessing failure scenarios and explaining model learning. This paper explores the
utilization of various datasets, including the unfiltered original dataset (DS2,1) and a database
processed through wavelet de-noising (DS0,2;1), for comparison and experimentation. Cross-
validation experiments are conducted to evaluate the system's robustness[5]. The outcome
showcases the high accuracy of the proposed CNN model in detecting arrhythmia assessed across
different heartbeats [2].Performance metrics such as accurate upbeat beats, false negative beats,
and false positive beats are employed for evaluation[10].
Table 1: Summarizing Challenges and Future Directions in Cardiovascular Disease (CVD) Classification
The challenges that could be tackled in future research within the CVD image analysis domain
are condensed and presented in Table 1. These challenges include the complexity of feature
extraction procedures and the need for domain expertise in creating suitable inputs for the
classifier[6].As an illustration, Daraei and Hamidi introduced a myocardial infarction (MI). A
prediction model based on machine learning, utilizing J48 algorithms, achieved an accuracy of
82.57%[11]. Bashir et al. employed machine learning approaches, including naive Bayes,
decision trees based on the Gini index and information gain, and SVM, to create an ensemble
model for predicting and analyzing heart patients, achieving an accuracy of 87.37%[34].In recent
Challenges Future Directions
Diversity in patient profiles Enhanced model adaptability to individual traits
Imbalanced datasets Strategies for handling imbalances and biases
Variations in ECG recordings Advancements in signal processing techniques
Inconsistent data quality Implementation of standardized data protocols
Complexity of CVD conditions Development of more nuanced classification models
Model interpretability Research on transparent and interpretable models
Applicability across populations Incorporation of diverse demographic factors
Integration into clinical practice User-friendly interfaces and workflow integration
Rigorous validation Continuous improvement in validation methodologies
Ethical considerations Development of ethical guidelines for CVD models
Continuous technological progress Adoption of state-of-the-art technologies

11
developments, deep learning (DL) has introduced an additional layer, emphasizing data-driven
approaches for heart disease diagnosis with Accuracy approaching 100%. Dutta implemented a
convolutional neural network for diagnosing coronary heart disease[28], and Li et al[27].
Introduced a deep neural network model named CraftNet. Despite the multiple hidden layers in
deep learning models, understanding the specific contribution of each layer in the final prediction
often poses challenges[35]. Additionally, a potential challenge is associated with the biased
performance of machine learning algorithms toward the majority class.
RQ2: What are the commonly employed performance metrics in CVD classification?
Various metrics gauge classification task performance. In cardiovascular disease (CVD)
classification, such as Accuracy, Accuracy assesses correctness, precision focuses on optimistic
predictions, and recall echoes the model's ability to recall relevant instances. Specificity discerns
the model's skill in identifying negative instances. F1 Score blends precision and recall for
balanced performance. The ROC-AUC assesses the model's capability to distinguish between
positive and negative cases; the Confusion Matrix furnishes details regarding true positives, true
negatives, false positives, and false negatives. Matthews Correlation Crescendo (MCC) balances
elements of the confusion matrix. Kappa Cadence gauges model alignment with expectations.
The Precision-Recall the nuances of optimistic predictions. Researchers use these metaphorical
notes for a thorough CVD classification model evaluation. They comprehensively evaluate the
model's performance in detecting and classifying cardiovascular diseases [18] .
RQ3: What are the various CVD classification algorithms?
Machine learning is increasingly used to define subtypes and predict risks in cardiovascular
diseases. Despite its growing application, using regularly employed ML models in cardiovascular
disease management remains to be determined. ML offers a unique AI method, categorizing
cardiac disease datasets with fewer features and higher performance indicators[1]. Moreover,
physicians are assisted in predicting outcomes by identifying complex patterns[36]. ML plays a
pivotal role in addressing challenges posed by the scale and complexity of medical data, aiming
to recommend early detection and treatment of cardiac conditions[6]. For example, Hamidi and
Daraei's ML-based model achieved 82.57% accuracy in early Myocardial Infarction (MI)
prediction using J48 algorithms[11]. Bashir et al. employed various ML methods, creating an
ensemble model for heart patient analysis with 87.37% accuracy[37]. Deep Learning (DL)
achieves close to 100% accuracy in heart disease diagnosis, as seen in Dutta's CNN-based model
for coronary heart disease[12], and Li et al.'s DNN model named CraftNet[27].and Li et al.'s
DNN model named CraftNet[27]. However, discerning the specific contribution of each layer in
DL models remains challenging[35]. This study used a UCI dataset with 76 columns, but only 14
encompassing heart disease attributes and risk factors were chosen for the experiment. Specific
attributes or risk factors used are not mentioned. It compares ML algorithms for heart disease
classification, evaluating metrics such as Accuracy, RMSE, MAE, precision, and recall[6].
Lacking details about the dataset's size or characteristics[2].It does not discuss potential biases or
limitations of selected ML algorithms, and there is no consideration of demographic factors'
impact on CVD prediction [6]. Generalizability to diverse populations or healthcare settings must
be addressed[10]. The authors do not explore potential ethical or privacy concerns related to
using medical data for machine learning prediction models[4]. This study proposes an ML model
for CVD classification. Krittanawong et al. highlight the paper's key points[38]. This study
assesses predictive efficacy for coronary artery disease, heart failure, stroke, and cardiac
arrhythmias. Of 344 studies, 103 cohorts (3,377,318 individuals) met the criteria. Boosting
algorithms predicted coronary artery disease (pooled AUC 0.88, 95% CI 0.84–0.91), and custom-
built algorithms had a pooled AUC of 0.93 (95% CI 0.85–0.97). For stroke prediction, SVM
algorithms had a pooled AUC of 0.92 (95% CI 0.81–0.97), boosting algorithms had a pooled
AUC of 0.91 (95% CI 0.81–0.96), and CNN algorithms had a pooled AUC of 0.90 (95% CI
0.83–0.95). Insufficient studies hinder meta-analysis in heart failure and cardiac arrhythmias,

12
suggesting no significant difference, though SVM may excel. The promising predictive ability of
ML algorithms in cardiovascular diseases, especially SVM and boosting algorithms, includes
heterogeneity among ML algorithms concerning multiple parameters, aiding clinicians in data
interpretation and optimal algorithm implementation. In Table 2, the summary highlights the
advantages and challenges associated with machine learning and traditional statistics in terms of
explainability, flexibility, variable selection, identifying interactions, and handling different data
sizes.
Table 2. A comparison between traditional statistics and machine learning in general terms[39]
Machine Learning Traditional statistics
Explainability Due to the hidden layer's
complexity, clinicians may
struggle to interpret machine
learning outcomes, especially
in neural networks.
Clinicians easily comprehend odds
ratios, hazard ratios, and confidence
intervals generated by conventional
statistical methods.
Flexibility
.
Highly flexible due to no
assumptions.
The necessary assumptions for
various models limit their
applications.
Variable selection ML utilizes diverse data,
including numerical data,
reports, and imaging, to discern
valuable variables from noise,
especially in genomics, where
predictors outnumber
outcomes.
Statistical methods, like LASSO,
use past calculations to distinguish
genuine predictor variables from
noise by regularising and selecting
variables for model generation.
Identifying interactions Unsupervised ML excels at
discovering new predictor
interactions compared to
traditional methods.
Variables must be collectively
chosen to identify interactions,
which may not occur if the
prevailing belief assumes none due
to a lack of evidence.
Variables must be jointly
chosen to identify Interplay it is
unlikely if the prevailing belief
assumes no interaction due to
insufficient evidence.
Incorporating various data
types into the calculation is
more Plainthan traditional
statistics.
It is not feasible to analyze multiple
data types with traditional statistics.
Size of data sets
.
ML excels with large datasets
but could be more dependable
with smaller ones.
Typically more dependable than
ML with smaller datasets.
Ganga[40]Explored machine learning for cardiovascular disease prediction, emphasizing early
detection. The authors highlight its potential for precise outcomes through insights from
healthcare datasets. The study covers genetic algorithms, random forests, naïve Bayes ANN,
SVM, particle swarm optimization, and ant colony optimization, achieving 99.65% maximum
accuracy with the Cleveland dataset. Various classification techniques like NB, DT, SVM, RF,
ANN, and ensemble learning are widely used. The study suggests that future research could
integrate diverse optimization techniques with machine learning for improved predictive
accuracy. R.K. Halder and Uddin, M.N [41]. Introduced an ensemble method-based multiplayer
dynamic system (MLDS) for cardiovascular disease prediction using machine learning, achieving
high accuracy (88.84% to 94.16%) across various train-test data ratios. Feature selection
techniques like CAE, GRAE, and IGAE are employed, and classifiers, including RF, NB, and GB,

13
are integrated into the model. The KNN algorithm is utilized when base classifiers fail. Data
preprocessing involves converting gender values, handling missing data, and outlier removal
using DBSCAN. The MLDS is evaluated on Kaggle and UCI datasets, outperforming five
comparison models. However, the paper lacks discussions on limitations, generalizability, biases
in datasets, computational resources, and model interpretability.
Machine Learning Techniques:
Arthur Samuel coined In 1959 'machine learning' was coined, defining it as a field where
computers autonomously learn without explicit instructions[42]. It involves developing
algorithms based on past data and experience, which is the essence of machine learning. In 1997,
Tom Mitchell further characterized machine learning as a process where a computer program
learns from experience E concerning tasks T and performance measure P, improving its
performance in T as measured by P with experience[43]. As a subset of artificial intelligence,
machine learning enables systems to learn from data, avoiding explicit programming. It employs
various algorithms to enhance, interpret, and predict outcomes through continuous learning from
data[44]. Data can be automated user training sets or other records obtained through interactions
with the environment. In all scenarios, the effectiveness of a learner's prediction depends on the
quality and size of the data. Figure 4 illustrates the results of a machine learning (ML) model
trained and tested with additional data for prediction. A classification approach is a model
providing predictions based on previously trained data[12]. Machine learning is pivotal in
developing analytics models that are applied to accurate diagnoses, prescription medicines,
pathological tests, early disease detection, and informed decision-making. Medical diagnosis, in
particular, is crucial as it influences treatment. Therefore, the widespread use of machine learning
techniques in healthcare, as outlined in[13]. Is vital for swift and accurate treatment decisions.
Leveraging computer-based decision support systems plays a significant role in ensuring accurate
diagnoses and cost-effective treatments.
Figure 4. A general framework for cardiovascular disease classification.
Machine learning aims for increasingly positive outcomes through precise predictions. The
classifications of machine learning techniques are detailed as follows:
Support Vector Machines (SVM): SVM serves as a supervised learning algorithm for
classification tasks, identifying the optimal hyperplane that separates different classes within a
feature space.

14
Random Forests: This ensemble learning method constructs numerous decision trees and
integrates their outputs to enhance accuracy and mitigate overfitting.
K-Nearest Neighbour (k-NN): A straightforward and effective classification algorithm, k-NN
assigns a data point it is class based on the majority class among its k-nearest neighbors.
Logistic Regression: Utilized for binary classification problems, logistic regression models the
probability of a specific outcome.
Naive Bayes: A probabilistic algorithm, Naive Bayes calculates the probability of a data point
belonging to a particular class based on its features.
Artificial neural networks, also known as multilayer perceptron (MLP), simulate complex non-
linear functions inspired by biology and are essential tools in machine learning [27]. With hidden
input, neural networks utilize the hidden layer to transform inputs into patterns controlled by the
output layer, excelling at instructing computers to recognize intricate patterns [28]. In use since
the 1940s, neural networks, especially with backpropagation, have gained significance in AI by
allowing adjustments to hidden layers when outputs deviate from user expectations[44].
A decision tree a tree-like structure in predictive models, is widely used across various fields. It
employs an algorithmic approach to partition data into different scenarios and is a practical
strategy for supervised learning techniques applied in classification and regression[16]. The
Decision Tree method starts by calculating attribute entropy, dividing the dataset based on
predictors or variables, and repeating the process with the remaining features.
Random Forest is a well-known and effective machine learning technique proficient in
classification and regression, excelling particularly in classification tasks
[20]. Bagging Aggregation, a machine learning algorithm, utilizes the bootstrap method to
approximate values such as the mean from many data samples [21]. The average of mean values
estimates the actual average value. Bagging follows a similar procedure but predominantly
employs decision trees. It generates individual models for each sample by considering the
number of data samples for training and providing predictions when needed. These predictions
are averaged to evaluate the output value more effectively[22].
Ensemble Methods: Techniques like AdaBoost and Gradient Boosting amalgamate predictions
from multiple weak learners to create a more robust classifier.
Neural Networks:Simple NN with one or more hidden layers can also be employed for CVD
ECG classification, especially in cases where deep learning may not be necessary.
Deep Learning Techniques:
Gated Recurrent Units (GRU): GRU are a type of recurrent neural network similar to LSTM but
with a simplified architecture. They are effective for sequence modeling.
Capsule Networks:Capsule networks aim to overcome gaps in traditional neural networks by
better handling hierarchical relationships within data.
Generative Adversarial Networks (GAN):While GAN are primarily used for generating data,
they can also be applied to anomaly detection and improve the robustness of classifiers.
Transfer Learning: Pte-trained models, such as those trained on ImageNet, can be fine-tuned for
CVD ECG classification tasks. This is especially useful when dealing with limited datasets.

15
Self-Supervised Learning: This learning paradigm involves creating supervisory signals from
the data, often by defining pretext tasks. It has shown promise in specific medical image analysis
tasks.
Convolutional Neural Networks (CNN): Specifically designed for tasks involving images,
CNN excel in ECG signal classification, automatically acquiring hierarchical features from raw
data.
Long Short-Term Memory (LSTM): LSTM are a type of RNN designed to handle data
sequences. They are effective in capturing temporal dependencies in time-series data like ECG
signals.
Bidirectional LSTM (Bi-LSTM): Bi-LSTMs process sequences in both forward and backward
directions, providing better context understanding and capturing long-term dependencies.
Deep Residual Networks (ResNet): ResNet architectures utilize residual learning to address the
vanishing gradient problem in deep networks, making them suitable for intense neural networks.
Attention mechanisms concentrate on particular segments of the input sequence, enabling the
model to emphasize pertinent aspects while undertaking classification.
Autoencoders are used for unsupervised learning and feature extraction. They can be applied in
preprocessing stages to enhance the representation of ECG signals before classification.
Graph Neural Networks (GNNs): GNNs are designed to work with graph-structured data. They
can be applied when the relationships between different components in the ECG signal are
crucial. Table 3 outlines the advantages and disadvantages of some deep learning models.
Table 3. Pros and cons of the six distinct deep learning approaches
Models Pros Cons
Convolutional Neural
Network
Sharing weights reduces training
parameters
.
1. Valuable information is lost
in the pooling layer.
2. Limited interpretability.
Long Short-Term Memory
(LSTM)
1. Ideal for handling sequential
signals.
2. Addresses vanishing gradient in
temporal sequences
The LSTM model encounters
complexity with prolonged
training and prediction times.
Convolutional Recurrent
Neural Network
It combines CNN and RNN to
analyze ECG morphology and
beat-to-beat variations in MI
research.
Significant computational
expense
ResNet 1. Deepening the network during
training.
2. Addressing adverse effects of
increased depth, like degradation.
3. Mitigating information loss
compared to CNN, but training
duration extends
The duration of training is
extended.

16
Table 4. Evaluating Machine Learning Methods for Predicting Cardiovascular Disease.
Author Proposed Study Dataset
Used
Proposed
Algorithm
Limitation/Gaps Accuracy
Chintan
M. et al.
[45]
Improvements to
enhance classification
accuracy.
70,000
patient
records
RF, DT, MP,
and
XGBoost.
Single database
dependency variable
scope gaps.
87.28%
Jilbab et
al. [46]
Use six standard tools
and six machine
learning algorithms
for classifying cardiac
disease.
UCI LR, SVM,
KNN, A NN,
NBand, RF.
Small dataset, and the
algorithms may not
encompass all available
techniques.
84.48%
Siddiket
al. [47]
Predict HD using ML
algorithms and
improve prediction
accuracy.
LR, DT,
KNN, NB,
SVM
91%
Sindhu[48] CVD UCI RF,
DT,SVM,
KNN
Insufficient dataset. 85.71%
Nissa[49] ML algorithms for
CVD prediction and
assess their
performance..
1329
instances
and 14
attributes
.
ANN,RF,LR
,SVM,NB
Limited insights for
heart disease diagnosis,
incomplete details on
sample size and dataset
in machine learning
evaluations.
97.29%
Uddin, M.
N[41]
Ensemble method-
based multilayer
dynamic system
70,000
from
Kaggle.
(RF), (NB),
and Gradient
Boosting
Lack of
modelcomparison.
94.16%
Nisa, N. et
al.[35]
Comparative analysis
of techniques for
predicting heart
disease.
UCI RF, SVM,
DT,
The time complexity
remains undetermined,
and the framework is
also unspecified.
99.35%
Patil,[5] Developing a machine
learning-based
algorithm for risk
prediction of CVD
NN,KNN,N
B
98.71%
GRU Simplifying structure match to
LSTM while maintaining
effectiveness.
GRU performs less active on
extensive datasets than LSTM.
AE Conducting feature dimension
reduction aids data visualization
analysis.
Compression applies only to
samples resembling the training
data

17
Hau Tran,
et al.[51]
Early detection of
CAD deep learning
framework with TL
and XGBoost of
CAD.
1250
samples
SPECT
MPI
XGBoost need for thorough
evaluation and detailed
dataset discussion.
81.5%
Sajja,
T[52]
Proposed a CNN to
classification the
disease at an
early stage
UCI CNN Enhancing Risk Factor
Dataset Inclusivity.
94%
Ketuet
al.[23]
predict HD using the
pickled model in
Flask's ML models.
UCI DT,
KNN,RF,
LR, ANN.
Small dataset used. 100%
This study strongly recommends further enhancements by shifting the focus from theoretical
frameworks and classification to practical datasets[11]. It suggests exploring improved
forecasting techniques and developing innovative feature selection methods to deepen
understanding essential elements and enhance precision in heart disease prediction. Deep learning,
leveraging CVD data, has the potential to enhance comprehension of heart disease subgroups and
automate ECG workflows. Encouraging research engagement and employing advanced
representation learning methods on echocardiographic data requires standardized benchmarks due
to the limited availability of publicly accessible datasets[63]. Applying machine learning and
deep learning in cardiovascular fields presents extensive possibilities, enabling personalized care.
Cardiology, especially in cardiac imaging, is transforming, and physicians must be prepared for
these changes. Instead of resisting the integration of AI into cardiology, physicians should
embrace it, acknowledging that their expert knowledge will remain essential. These techniques
can be tailored or specialized based on the specific requirements of cardiovascular disease (CVD)
ECG classification issues and the characteristics of available data, striking a balance between
model complexity, interpretability, and computational resources[65]. With ML advancements,
researchers utilize data-driven methods for diagnosing heart-related diseases through ECG
signals. ML-based approaches provide early identification, enabling routine self-assessments with
affordable sensors, unlike traditional tests that cause delayed diagnoses[10]. Analyzing heartbeat
rhythms can detect heart disease early, categorizing beats into five types. Abnormal heartbeats,
known as arrhythmia, were a focus in 13 out of 49 selected papers due to their severe
consequences, contributing significantly to morbidity and mortality among cardiac patients
[11].The recent discussion on advancements in modeling, imaging, and monitoring CVD with
machine learning explores its application in analyzing, visualizing, and monitoring these diseases.
The paper emphasizes machine learning's potential to address challenges like computational cost,
limited resolution, and data analysis obstacles. It specifically focuses on applying deep learning
methods to expedite flow modeling, improve resolution, reduce noise in blood flow imaging, and
accurately detect cardiovascular diseases using wearable sensor data. The authors stress the need
for collaboration among experts in computational fluid dynamics, blood flow imaging, cardiology,
wearable sensors, and machine learning to leverage cardiovascular research[2]. further. Deep
learning techniques and profound neural networks were utilized to accelerate flow modeling,
enhance resolution, diminish noise, and reduce scanning time in blood flow imaging methods.
Machine learning algorithms, including deep learning, were applied to accurately detect and
classify cardiovascular diseases using data gathered from wearable sensors[10].Physics-informed
neural networks were integrated to surrogate model the flow field with minimal or absent
simulated data, reducing the necessary training samples. Geometric details of blood vessels, such
as the aorta and coronary artery branches, were acquired through computed tomography
angiography (CTA) scanning to construct training datasets for predicting flow dynamics using

18
deep learning[4].It is crucial to foster collaboration among experts in computational fluid
dynamics, blood flow imaging, cardiology, wearable sensors, and machine learning for further
cardiovascular disease analysis and monitoring advancements. Increasingly utilizing machine
learning techniques and intense learning can significantly enhance cardiovascular monitoring,
care, and management. Continuous research and development are imperative to decrease the
computational cost and time required for CFD simulations using deep learning methods, making
them more applicable for clinical applications. Efforts should also improve the resolution and
quality of images obtained from 4D-flow MRI by applying deep learning techniques to reduce
noise and scanning time.Additionally, it is advisable to explore further applications of machine
learning for early detection and risk stratification of cardiovascular diseases using data from
wearable sensors[66]. Precision medicine elevates care through early disease risk prediction and
personalized treatments, leveraging advanced diagnostics and genomic medicine. AI and ML
play pivotal roles by providing insights and enabling predictive models for physicians through
high-throughput data [2]. ML algorithms, like decision trees and random forests, are applied in
real-time healthcare monitoring, therapeutic decision support, and disease prediction[6]. They
assist cardiologists in predicting cardiovascular risk, detecting cardiac events, and extracting
phenotypic information from the EHR for rare diseases and cancer phenotyping. Specific ML
algorithms, e.g., linear regression, naive Bayes, and KNN, find use in computational analyses,
clinical decision support, and drug response anticipation[10]. Despite the potential, challenges
include insufficient accuracy for widespread medical use, limitations in understanding gene
function, and unaddressed ethical and privacy concerns in precision and genomic medicine. The
future holds promise for ML in precision and genomic medicine, but overcoming obstacles like
the research-to-clinic divide and addressing limitations are crucial for realizing its full potential[6]
3. RESULTS
In the study's initial goal of identifying heart disease-related literature addressing imbalanced
datasets, it was observed that all 65 articles utilized imbalanced datasets, either from actual or
open sources. Although not all studies recognize different data as a challenge, recent literature
increasingly acknowledges its impact on model performance. Authors employ various strategies
to address imbalanced data; some utilize data-level solutions[68], while others prefer algorithm-
level solutions [69] for lower time complexity. Standard data-level solutions include the
widespread application of SMOTE, and a few studies adopt other methods like under-sampling.
DL-based methods are prevalent among algorithm-level solutions, utilizing GAN to augment
ECG data and enhance classification accuracy. Cost-sensitive approaches are also considered.
Daraei and Hamidi utilized the Metacost classifier with varying cost ratios, achieving a
sensitivity of 86.67%, an F1-score of 80%, and an accuracy of 82.67% with a cost ratio of 1:200.
Gan et al. proposed AdaC-TANBN, integrating TANBN with a cost-sensitive classification
algorithm, showcasing the effectiveness of cost-sensitive approaches in handling class
imbalances with the Cleveland dataset. While arrhythmia is extensively studied in machine-
learning-based heart disease diagnosis, cardiac arrest remains a significant concern in intensive
care units. Screening encounters challenges due to low sensitivity and high false alarm rates,
suggesting a need for researchers to broaden their focus to encompass all heart diseases, not
solely arrhythmia. ML models primarily utilize Cleveland data and MIT-BIH arrhythmia, with
limited attention given to real-world data.
Clinical decision support systems encounter instability, necessitating continuous model
adjustment based on new patient data. ML algorithms, including SVM, KNN, ANN, CNN, and
GAN, are employed, with CNN and GAN garnering attention for their robust performance.
Imbalanced data issues are tackled through methods like SMOTE and CNN-based solutions.
However, CNN needs more interpretability. Performance is superior under the intra-patient
paradigm but poor under the inter-patient paradigm. Distinguishing heartbeats from noisy ECGs

19
proves challenging. Models claiming success with imbalanced data should incorporate
explainable AI. Conventional classification models prioritize accuracy without considering
fluctuations in misclassification costs. Literature often involves multiple steps, making real-world
implementation challenging.
A standardized process and presentation requirements are necessary for cross-disciplinary
research. Studies vary in their emphasis on accuracy, AUC, ROC, sensitivity, and specificity.
Table 5 summarises research papers focusing on cardiovascular disease (CVD) diagnosis, each
with a distinct study emphasis and employing techniques. Notable topics include data preparation,
effective machine learning classification, and evaluating diagnostic performance for congenital
heart disease. Techniques range from content analysis to machine learning methods, emphasizing
features, algorithms, and samples. The comprehensive overview aids in understanding diverse
approaches in CVD research.
Table 4: Related Research.
Title
of
the
Papers
Study
Emphasis
Techniques
Content
Analysis
Imbalance
Issues
Meta
Analysis
SLR
Performance
Metrics
Data preparation for CVD
diagnosis: A systematic review
[13]
Data preparation
Effective classification of
CVD using machine learning
methods: a review[14]
Machine learning
techniques
Machine learning-based
cardiac disease prediction: a
meta-analysiss[15]
Features, algorithms,
and samples
Evaluation of the effectiveness
of machine learning methods
in detecting congenital heart
disease: a review[4]
Estimating the
diagnostic
performance of ML.
An exhaustive scope review of
machine learning and deep
learning-based classification of
cardiac disease[3]
Test data, patient
signals, dataset
origins, and feature
reduction/extraction
methods.
Our study Machine learning
based on CVD
classification
4. DISCUSSION
Selected papers undergo comprehensive evaluation based on predefined criteria, including the
research problem, proposed approach, machine learning classification, and accuracy. Criteria

20
assign the highest scores to models used in evaluating studies, typically applied to heart disease
datasets from various sources, not limited to UCI. Researchers often use reduction strategies or
feature selection when providing data to ML classifiers, focusing on key characteristics like chest
discomfort, blood pressure, fasting blood glucose, cholesterol, ECG readings, pulse rate, exercise,
age, and gender. While over 75% of systems rely on UCI heart disease data, multiple datasets and
ML techniques yield similar results. Despite variations, Random Forest (RF) showed the highest
accuracy frequency among the identified classifiers, with Support Vector Machine (SVM)
ranking second. Seven classifiers addressed CVD prediction: decision trees (DT), RF, naive
Bayes (NB), logistic regression (LR), artificial neural networks (ANN), k-nearest neighbors
(KNN), and SVM. RF, SVM, and KNN demonstrated reasonably high accuracy, but reliability
clarification on predictions was needed in literature sources. Constructing a machine-learning-
based clinical diagnosis system requires unique models due to variations in training data. The
stability of the model when parameters shift or update is critical for reliable diagnoses. The
results from the literature sources are summarized in Table 5, providing a comprehensive
understanding of machine learning-based heart disease diagnosis. Challenges arise from complex
feature extraction and the need for domain expertise in clinical applications. A thorough
evaluation of extensive clinical datasets is essential for the reliability of machine learning
methods. Integrating machine learning into cardiovascular disease analysis can improve patient
outcomes and significantly guide future research in cardiology[2].
5. LIMITATIONS
The study must thoroughly address biases and challenges related to imbalanced datasets in
cardiovascular disease (CVD) classification, potentially affecting the reliability of machine
learning models. Additionally, there needs to be more discussion on the generalization of
machine learning models to diverse populations and external validation, which is essential for
evaluating their robustness across various patient groups. The study briefly mentions the
interpretability of CNN models. However, it does not sufficiently emphasize their critical
importance in clinical settings, impacting healthcare professionals' understanding and trust in
machine learning decisions for cardiovascular disease.
6. CONCLUSION AND FUTURE WORK
This thorough literature review provides a nuanced synthesis of the current research on machine
learning applications for classifying cardiovascular diseases. Its primary focus is elucidating
recent trends and techniques for diagnosing heart disease using machine learning (ML) and data-
driven approaches. The central conclusion drawn from the analysis underscores the importance of
broadening ML-based experiments to incorporate real-time patient data and providing a clear
interpretation of the final predictions through interpretable machine learning. The amalgamation
of findings highlights the potential of ML models to improve diagnostic accuracy and risk
prediction substantially. Nevertheless, the expectation is that ML-based heart disease diagnosis
with imbalanced data still involves unexplored aspects and holds significant potential for further
exploration in the years to come.
For future work, This study recommends shifting focus from theoretical frameworks to practical
datasets, promoting research engagement with advanced representation learning on CVD data.
Collaboration among experts in computational fluid dynamics, cardiology, wearable sensors, and
machine learning is crucial for advancing cardiovascular disease analysis.

21
REFERENCES
[1] Ali, M.M., et al., Heart disease prediction using supervised machine learning algorithms:
Performance analysis and comparison. Computers in Biology and Medicine, 2021. 136: p. 104672.
[2] Rajdhan, A., et al., Heart disease prediction using machine learning. INTERNATIONAL JOURNAL
OF ENGINEERINGRESEARCH & TECHNOLOGY (IJERT), 2020. 9(O4).
[3] Rath, A., et al., An exhaustive review of machine and deep learning based diagnosis of heart
diseases. Multimedia Tools and Applications, 2022. 81(25): p. 36069-36127.
[4] Hoodbhoy, Z., et al., Diagnostic accuracy of machine learning models to identify congenital heart
disease: a meta-analysis. Frontiers in artificial intelligence, 2021. 4: p. 708365.
[5] Patil, P.B., P.M. Shastry, and P. Ashokumar, Machine learning based algorithm for risk prediction
of cardio vascular disease (Cvd). Journal of critical reviews, 2020. 7(9): p. 836-844.
[6] Li, T., et al., Combining machine learning and logical reasoning to improve requirements
traceability recovery. Applied Sciences, 2020. 10(20): p. 7253.
[7] Ahsan, M.M. and Z. Siddique, Machine learning-based heart disease diagnosis: A systematic
literature review. Artificial Intelligence in Medicine, 2022. 128: p. 102289.
[8] Muhtaseb, R. and M. Yaqub. EchoCoTr: Estimation of the Left Ventricular Ejection Fraction from
Spatiotemporal Echocardiography. in International Conference on Medical Image Computing and
Computer-Assisted Intervention. 2022. Springer.
[9] Bhavekar, G.S. and A.D. Goswami, A hybrid model for heart disease prediction using recurrent
neural network and long short term memory. International Journal of Information Technology, 2022.
14(4): p. 1781-1789.
[10] Müller, H., et al., Explainability and causability for artificial intelligence-supported medical image
analysis in the context of the European In Vitro Diagnostic Regulation. New Biotechnology, 2022.
70: p. 67-72.
[11] Zhang, D., et al., An ECG heartbeat classification method based on deep convolutional neural
network. Journal of Healthcare Engineering, 2021. 2021: p. 1-9.
[12] Collins, C., et al., Artificial intelligence in information systems research: A systematic literature
review and research agenda. International Journal of Information Management, 2021. 60: p. 102383.
[13] Benhar, H., A. Idri, and J. Fernández-Alemán, Data preprocessing for heart disease classification: A
systematic literature review. Computer Methods and Programs in Biomedicine, 2020. 195: p.
105635.
[14] Verma, S. and A. Gupta. Effective prediction of heart disease using data mining and machine
learning: A review. in 2021 International Conference on Artificial Intelligence and Smart Systems
(ICAIS). 2021. IEEE.
[15] Kumar, N. and D. Kumar. Machine learning based heart disease diagnosis using non-invasive
methods: A review. in Journal of Physics: Conference Series. 2021. IOP Publishing.
[16] Lu, Y., et al., Automated arrhythmia classification using depthwise separable convolutional neural
network with focal loss. Biomedical Signal Processing and Control, 2021. 69: p. 102843.
[17] Yang, P., et al., Ensemble of kernel extreme learning machine based random forest classifiers for
automatic heartbeat classification. Biomedical Signal Processing and Control, 2021. 63: p. 102138.
[18] Katarya, R. and S.K. Meena, Machine learning techniques for heart disease prediction: a
comparative study and analysis. Health and Technology, 2021. 11: p. 87-97.
[19] Plati, D.K., et al., A machine learning approach for chronic heart failure diagnosis. Diagnostics,
2021. 11(10): p. 1863.
[20] Dell'Angela, L. and G.L. Nicolosi, Artificial intelligence applied to cardiovascular imaging, a
critical focus on echocardiography: The point‐of‐view from “the other side of the coin”. Journal of
Clinical Ultrasound, 2022. 50(6): p. 772-780.
[21] Chang, V., et al., An artificial intelligence model for heart disease detection using machine learning
algorithms. Healthcare Analytics, 2022. 2: p. 100016.
[22] Xiong, P., S.M.-Y. Lee, and G. Chan, Deep learning for detecting and locating myocardial infarction
by electrocardiogram: A literature review. Frontiers in cardiovascular medicine, 2022. 9: p. 860032.
[23] Ketu, S. and P.K. Mishra, Empirical analysis of machine learning algorithms on imbalance
electrocardiogram based arrhythmia dataset for heart disease detection. Arabian Journal for Science
and Engineering, 2022: p. 1-23.
[24] Zhu, H., et al., Segmentation of coronary arteries images using spatio-temporal feature fusion
network with combo loss. Cardiovascular Engineering and Technology, 2022: p. 1-12.

22
[25] Sharmila, V., “Multi-Class Arrhythmia Detection using a Hybrid Spatial-Temporal Feature
Extraction Method and Stacked Auto Encoder. Int. J. Intell. Eng. Syst, 2020. 14(2): p. 82-94.
[26] Arrieta, A.B., et al., Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities
and challenges toward responsible AI. Information fusion, 2020. 58: p. 82-115.
[27] Shankar, V., et al., Heart disease prediction using CNN algorithm. SN Computer Science, 2020. 1(3):
p. 170.
[28] Kusunose, K., Steps to use artificial intelligence in echocardiography. Journal of echocardiography,
2021. 19(1): p. 21-27.
[29] Wahlang, I., et al., Deep Learning methods for classification of certain abnormalities in
Echocardiography. Electronics, 2021. 10(4): p. 495.
[30] Manur, M., A.K. Pani, and P. Kumar, A prediction technique for heart disease based on long Short
term memory recurrent neural network. International Journal of Intelligent Engineering and Systems,
2020. 13(2): p. 31-39.
[31] Chen, C.-W., et al., Outpatient text classification using attention-based bidirectional LSTM for
robot-assisted servicing in hospital. Information, 2020. 11(2): p. 106.
[32] Siddique Ibrahim, S. and M. Sivabalakrishnan, An evolutionary memetic weighted associative
classification algorithm for heart disease prediction. Recent advances on memetic algorithms and its
applications in image processing, 2020: p. 183-199.
[33] Schlesinger, D.E. and C.M. Stultz, Deep learning for cardiovascular risk stratification. Current
Treatment Options in Cardiovascular Medicine, 2020. 22: p. 1-14.
[34] Yazdani, A., et al., A novel approach for heart disease prediction using strength scores with
significant predictors. BMC medical informatics and decision making, 2021. 21(1): p. 194.
[35] Hughes, J.W., et al., Deep learning prediction of biomarkers from echocardiogram videos. Medrxiv,
2021: p. 2021.02. 03.21251080.
[36] Banerjee, A., et al., Machine learning for subtype definition and risk prediction in heart failure, acute
coronary syndromes and atrial fibrillation: systematic review of validity and clinical utility. BMC
medicine, 2021. 19: p. 1-14.
[37] Thura, S., et al. Analysis of the Machine Learning Classification of Cardiac Disease on Embedded
Systems. in 2023 Systems and Information Engineering Design Symposium (SIEDS). 2023. IEEE.
[38] Krittanawong, C., et al., Machine learning prediction in cardiovascular diseases: a meta-analysis.
Scientific reports, 2020. 10(1): p. 16057.
[39] Allan, S., R. Olaiya, and R. Burhan, Reviewing the use and quality of machine learning in
developing clinical prediction models for cardiovascular disease. Postgraduate Medical Journal,
2022. 98(1161): p. 551-558.
[40] Sharma, G., G. Rani, and V.S. Dhaka. A Review on Machine Learning Techniques for Prediction of
Cardiovascular Diseases. in 2020 Sixth International Conference on Parallel, Distributed and Grid
Computing (PDGC). 2020. IEEE.
[41] Uddin, M.N. and R.K. Halder, An ensemble method based multilayer dynamic system to predict
cardiovascular disease using machine learning approach. Informatics in Medicine Unlocked, 2021.
24: p. 100584.
[42] Popic, J., B. Boskovic, and J. Brest. Deep Learning and the Game of Checkers. in MENDEL. 2021.
[43] Steck, H., et al., Deep learning for recommender systems: A Netflix case study. AI Magazine, 2021.
42(3): p. 7-18.
[44] Yadav, D., Machine Learning: Trends, Perspective, and Prospects. 2020.
[45] Bhatt, C.M., et al., Effective heart disease prediction using machine learning techniques. Algorithms,
2023. 16(2): p. 88.
[46] Tougui, I., A. Jilbab, and J. El Mhamdi, Heart disease classification using data mining tools and
machine learning techniques. Health and Technology, 2020. 10: p. 1137-1144.
[47] Chowdhury, M.N.R., et al. Heart disease prognosis using machine learning classification techniques.
in 2021 6th International Conference for Convergence in Technology (I2CT). 2021. IEEE.
[48] Kumar, N.K., et al. Analysis and prediction of cardio vascular disease using machine learning
classifiers. in 2020 6th International Conference on Advanced Computing and Communication
Systems (ICACCS). 2020. IEEE.
[49] Nissa, N., S. Jamwal, and S. Mohammad, Early detection of cardiovascular disease using machine
learning techniques an experimental study. Int. J. Recent Technol. Eng, 2020. 9(3): p. 635-641.

23
[50] Triki, N., M. Karray, and M. Ksantini, A real-time traffic sign recognition method using a new
attention-based deep convolutional neural network for smart vehicles. Applied Sciences, 2023. 13(8):
p. 4793.
[51] Thi, H.T., et al., A Deep Learning framework with transfer learning and XGBoost for multi-label
classification of Coronary Artery Disease from SPECT polar maps.
[52] Sajja, T.K. and H.K. Kalluri, A Deep Learning Method for Prediction of Cardiovascular Disease
Using Convolutional Neural Network. Rev. d'Intelligence Artif., 2020. 34(5): p. 601-606.
[53] Yang, H., et al., Efficient and robust instrument segmentation in 3D ultrasound using patch-of-
interest-FuseNet with hybrid loss. Medical Image Analysis, 2021. 67: p. 101842.
[54] Danu, M., C.F. Ciuşdel, and L.M. Itu. Deep learning models based on automatic labeling with
application in echocardiography. in 2020 24th International Conference on System Theory, Control
and Computing (ICSTCC). 2020. IEEE.
[55] Kazemi Esfeh, M.M., et al. A deep Bayesian video analysis framework: towards a more robust
estimation of ejection fraction. in International Conference on Medical Image Computing and
Computer-Assisted Intervention. 2020. Springer.
[56] Ahamed, M.A., et al. ECG heartbeat classification using ensemble of efficient machine learning
approaches on imbalanced datasets. in 2020 2nd International Conference on Advanced Information
and Communication Technology (ICAICT). 2020. IEEE.
[57] Chen, Y., et al. Assessing the generalizability of temporally coherent echocardiography video
segmentation. in Medical Imaging 2021: Image Processing. 2021. SPIE.
[58] Mehmood, A., et al., Prediction of heart disease using deep convolutional neural networks. Arabian
Journal for Science and Engineering, 2021. 46(4): p. 3409-3422.
[59] Chandra, V., P.G. Sarkar, and V. Singh, Mitral valve leaflet tracking in echocardiography using
custom Yolo3. Procedia Computer Science, 2020. 171: p. 820-828.
[60] Liu, X., et al., Deep learning-based automated left ventricular ejection fraction assessment using 2-D
echocardiography. American Journal of Physiology-Heart and Circulatory Physiology, 2021. 321(2):
p. H390-H399.
[61] Dutta, A., et al., An efficient convolutional neural network for coronary heart disease prediction.
Expert Systems with Applications, 2020. 159: p. 113408.
[62] Deng, K., et al. TransBridge: A lightweight transformer for left ventricle segmentation in
echocardiography. in Simplifying Medical Ultrasound: Second International Workshop, ASMUS
2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021,
Proceedings 2. 2021. Springer.
[63] Baccouche, A., et al., Ensemble deep learning models for heart disease classification: A case study
from Mexico. Information, 2020. 11(4): p. 207.
[64] Hwang, I.-C., et al., Differential diagnosis of common etiologies of left ventricular hypertrophy
using a hybrid CNN-LSTM model. Scientific Reports, 2022. 12(1): p. 20998.
[65] Moradi, H., et al., Recent developments in modeling, imaging, and monitoring of cardiovascular
diseases using machine learning. Biophysical Reviews, 2023. 15(1): p. 19-33.
[66] Quazi, S., Artificial intelligence and machine learning in precision and genomic medicine. Medical
Oncology, 2022. 39(8): p. 120.

IMBALANCED DATASET EFFECT ON CNN-BASED CLASSIFIER PERFORMANCE FOR FACE RECOGNITION

More Related Content

Similar to IMBALANCED DATASET EFFECT ON CNN-BASED CLASSIFIER PERFORMANCE FOR FACE RECOGNITION (20)

Recently uploaded (20)

IMBALANCED DATASET EFFECT ON CNN-BASED CLASSIFIER PERFORMANCE FOR FACE RECOGNITION