Received 16 June 2025, accepted 3 July 2025, date of publication 16 July 2025, date of current version 11 August 2025.

Digital Object Identifier 10.1109/ACCESS.2025.3589529

Automated Cardiac Disease Prediction Using

Composite GAN and DeepLab Model
SOHAIL JABBAR 1 , UMAR RAZA 2 , (Senior Member, IEEE), MUHAMMAD ASIF HABIB 1,

MUHAMMAD FARHAN 3 , AND SAQIB SAEED 4

1 Collegeof Computer and Information Sciences, Imam Mohammad Ibn Saud Islamic University (IMSIU), Riyadh 11432, Saudi Arabia
2 Department of Engineering, Manchester Metropolitan University, M15 6BX Manchester, U.K.
3 Department of Computer Science, COMSATS University Islamabad, Sahiwal Campus, Sahiwal 57000, Pakistan
4 Department of Computer Information Systems, College of Computer Science and Information Technology, Imam Abdulrahman Bin Faisal University,

Dammam 31441, Saudi Arabia

Corresponding author: Umar Raza ([email protected])
This work was supported and funded by the Deanship of Scientific Research at Imam Mohammad Ibn Saud Islamic University (IMSIU)
(grant number IMSIU-DDRSP2501).

ABSTRACT Cardiovascular diseases remain the leading global cause of mortality, resulting in
over 17 million deaths annually. Manual cardiac image interpretation is often subjective and varies
significantly among clinicians. However, constraints like limited annotation and model generalization persist.
We introduce GenDeep, a novel framework integrating an unsupervised Generative Adversarial Network
(GAN) and DeepLab model for robust cardiac pathology classification from cine-MRI scans. The GAN
component performs data augmentation to synthesize realistic pathological imagery, overcoming dataset
constraints. Meanwhile, the DeepLab segmentation network exploits inter-slice spatial contexts for precise
anatomical quantification. GenDeep is trained on over 4000 expert-annotated scans from the ACDC dataset,
leveraging Apache Spark and Hadoop for efficient parallel data loading and preprocessing. The Generator
maps noise vectors to synthetic MRIs while the Discriminator predicts disease labels and classifies images as
real/fake. Weights are updated through backpropagation to refine image realism and classification accuracy.
Once trained, the Generator produces additional pathological data to boost model generalization. The
Discriminator then serves as the diagnostic classifier based on ventricular morphology from DeepLab
segmentation. Extensive comparative testing on a held-out test set achieves 97% accuracy and 93% F1
Score, significantly exceeding benchmarks. Smooth convergence is verified with a low 2.21 MSE. These
results highlight the effective integration of generative learning and segmentation for automated and reliable
cardiac diagnosis.

INDEX TERMS Cardiovascular diseases, generative adversarial networks, DeepLab, semantic segmenta-
tion, deep learning, cardiac MRI analysis, disease detection, heart disease.

I. INTRODUCTION complex, tedious, and prone to subjective variability and

Cardiovascular diseases (CVDs) remain the leading cause human error, thus impeding timely detection [4]. Conse-
of mortality worldwide, accounting for over 17 million quently, more than 50% of CVD cases remain undiagnosed
deaths annually [1]. CVDs encompass various cardiac until significant cardiac damage has occurred [5]. Machine
conditions, including coronary artery disease, heart failure, learning has surfaced as a promising solution for automated
cardiomyopathy, and arrhythmias [2]. Early intervention is cardiac disease prediction through discerning analysis of
essential, as initial cardiac events or strokes are often fatal [3]. medical images [6]. Computational diagnostics systems
Manual examination of cardiac scans for diagnosis remains leveraging deep neural networks bear immense potential to
refine screening, elevate detection rates in pre-symptomatic
The associate editor coordinating the review of this manuscript and individuals, and mitigate mortality. These systems offer
approving it for publication was Jiachen Yang . an objective, precise, and efficient means of extracting
2025 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
VOLUME 13, 2025 For more information, see https://creativecommons.org/licenses/by/4.0/ 138313
 S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

insights from large-scale historical data to derive data-driven Even though GAN and semantic segmentation have
disease patterns reliably [7]. Specifically, CNN architectures been explored separately for medical imaging tasks, their
have displayed unparalleled effectiveness on cardiac image cooperation in a single cardiac disease prediction system
segmentation and diagnosis classification tasks [8]. However, from cine-MRI is still unproven. GenDeep is unique in
research to elevate generalization across diverse patient merging different approaches, by first using a GAN to
demographics remains imperative. Key technical challenges create fake MRI images that are then used to support both
include limited annotated samples for training and variability segmentation and classification in heart scans. In contrast
across multi-center image captures, alongside interpretability to other data augmentation methods, the generator is taught
constraints hindering clinical adoption [9]. Tackling said to conserve the body’s correct form, making the model
hurdles while benchmarking real-world data holds the key to more reliable when faced with variability in medical images.
expediting the translation of machine learning innovations to Furthermore, the diagnostic label set is expanded from five to
transform cardiology. ten ACDC conditions, so the disease classification benefits
Recent studies have employed supervised deep learning from even better segmentation. Merging these techniques
models for automated analysis of cardiac imaging data and produces a special architecture that can be used for true
disease diagnosis [10, 11]. Convolutional neural networks cardiac screening applications. The main contributions of this
(CNNs), in particular, have achieved high efficiency by paper are as follows:
automatically learning discriminative features from medical
• We propose a new deep learning framework combining
images [10]. However, typical CNN architectures struggle to
an unsupervised GAN for medical image generation and
accurately segment irregular anatomical structures and fail to
a DeepLab convolutional neural network for semantic
sufficiently leverage all available supervision signals beyond
segmentation. This composite architecture aims to syn-
pixel-labels [11]. This has impelled the advancement of
ergize the complementary strengths of GANs for robust
sophisticated deep learning architectures such as Generative
feature learning and DeepLabs for precise anatomical
Adversarial Networks (GANs) and DeepLab models to
delineation.
transcend said limitations. This paper introduces GenDeep,
• Our model automatically learns discriminative visual
a novel framework that integrates an unsupervised Composite
features from cardiac MRIs and accurately segments the
GAN for realistic medical image synthesis and a DeepLab
left and right ventricle structures. Accurate segmentation
neural network for precise semantic segmentation to enable
forms the basis for subsequent pathology classification
automated cardiac disease prediction. The model undergoes
based on ventricular morphology.
training on an expansive dataset of 4000+ cardiac MR images
• The research demonstrates GenDeep’s ability to predict
from the ACDC challenge [12], leveraging Apache Spark ten clinically significant cardiac conditions, including
and Hadoop for swift parallel data loading and filtering. coronary artery disease, myocardial infarction, heart
The Composite GAN encompasses interlinked generator failure, arrhythmias, valvular heart disease, cardiomy-
and discriminator components to produce realistic cardiac opathy, hypertension, peripheral artery disease, congen-
MRIs. Meanwhile, DeepLab employs atrous convolutions ital heart disease, and rheumatic heart disease, even
and Conditional Random Fields to accurately segment left when trained on a sparsely annotated dataset. Transfer
and right ventricles. The engineered GenDeep framework learning and semi-supervised approaches are leveraged
aims to boost automated feature learning and cardiac to compensate for limited labeling.
structure demarcation. Post-training, it can analyze novel • The ACDC dataset, comprising more than 4000 cine-
MR scans to identify likely cardiac defects. Rigorous MRI scans, is utilized for rigorous evaluation. Compar-
benchmarking on a held-out test set and comparisons to ative testing using 5-fold cross-validation demonstrates
cutting-edge approaches across segmentation and diagnos- the model’s generalization capability to unseen data.
tic accuracy, sensitivity, and specificity gauge GenDeep’s Segmentation and multi-class classification metrics
efficacy. showcase GenDeep’s strengths.
Applications of automated cardiac diagnosis systems span
across policy screening initiatives, mobile & rural health The remainder of this paper is organized as follows.
services, and smart wearable devices. Population screening Section II reviews related works on deep learning for cardiac
enables health authorities to systematically evaluate CVD image analysis and diagnosis, focusing on CNNs, GANs,
risk in communities to guide interventions. Diagnostic and DeepLab models. Section III describes the ACDC
wearables allow continuous monitoring for patients in remote cardiac cine-MRI dataset and preprocessing pipeline used
areas or with mobility constraints. Smartphone integrations in our experiments and introduces our proposed GenDeep
serve those lacking expensive infrastructure. Ultimately by methodology, integrating Composite GAN and DeepLab
democratizing screening access, these solutions aim to boost for end-to-end cardiac disease prediction, elaborating the
early detection rates and positively impact CVD outcomes underlying architectures and training. Section IV benchmarks
globally. Realizing this vision hinges on extending robustness GenDeep’s effectiveness for medical image synthesis, cardiac
and accessibility of machine learning innovations through ventricle segmentation, and pathology classification against
multi-disciplinary collaboration [13]. state-of-the-art techniques. Finally, Section V presents the

138314 VOLUME 13, 2025

S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

key conclusions and scientific contributions of our work quantum machine learning in cardiovascular diagnostics. The
alongside future research directions for advancing computa- examination emphasizes the significance of this quantum leap
tional diagnosis with deep learning. and the success of the bagging ensemble learning technique
in improving prediction accuracy. Examining heart disease
II. LITERATURE REVIEW through deep learning and traditional methods, compares
Several studies have applied machine learning and data UCI and real-time datasets. Introducing the cluster-based
mining techniques to address the global issue of heart bi-directional long-short term memory (C-BiLSTM) for
disease. For example, Bertsimas et al. [14] used the Cleveland improved accuracy, the inquisition removes duplicate data
dataset with 14 key attributes and found that the K-Nearest using K-Means clustering. The output shows C-BiLSTM
Neighbors (KNN) algorithm achieved the highest classifi- outperforming Regression Tree, SVM, Logistic Regression,
cation accuracy. The findings contribute valuable insights KNN, Gated Recurrent Unit, and Ensemble, achieving
into data-driven approaches for early diagnosis and effective 92.84% accuracy for real-time datasets. Sk et al. [20]
management of cardiovascular health [14]. In the realm of underscore the importance of integrating advanced machine
clinical data analysis, Mohan et al. [13] proposed a Hybrid learning techniques in healthcare for predicting chronic
Random Forest and Linear Model (HRFLM), which achieved diseases, particularly heart disease. They presented a hybrid
an accuracy of 88.7% in predicting cardiovascular disease, machine learning model using Decision Tree and AdaBoost
contributing valuable insights to enhance predictions in for coronary heart disease prediction, emphasizing key eval-
cardiovascular health [15]. Bhatt et al. [16] applied K-modes uation metrics. Examining the urgency of disease diagnosis,
clustering combined with traditional classifiers (e.g., Random this study focuses on heart disease and emphasizes the vital
Forest, Decision Tree, MLP, XGBoost) and found the role of machine learning classification methods in providing
Multilayer Perceptron model achieved the best performance reliable and immediate assistance to healthcare professionals.
at 87.28% accuracy. Emphasizing the importance of precise The brief overview highlights the current advancements in
classification for effective treatment in cardiovascular health. utilizing these techniques, stressing their potential to enhance
Exploring early prediction for life-threatening conditions the efficiency of disease identification [21]. In the realm of
like heart disease, this analysis utilizes machine learning healthcare challenges, the focus on early-stage heart disease
to enhance accuracy and simplicity. Testing various algo- detection, particularly crucial in underserved areas lacking
rithms on three datasets, Logistic Regression consistently specialized medical professionals. The developed hybrid
performs well, achieving accuracies of 91.6% and 90.8%. decision support system, leveraging clinical parameters,
The experimentation underscores machine learning’s efficacy exhibits remarkable accuracy (86.6%) with the random forest
in predicting heart disease, emphasizing the potential for classifier. Tested on the UCI machine learning repository’s
improvement through collaboration with medical profession- Cleveland heart disease dataset, the system surpasses existing
als and further exploration [17]. Explores data mining for prediction models, showcasing its effectiveness in enhancing
healthcare, specifically addressing the challenge of predicting early detection of heart disease [22].
heart disease in individuals with diabetes. Despite existing In a recent study, the authors Dwivedi [23] proposed a
classification algorithms, there is a notable gap in diabetic- heart disease prediction tool based on 14 clinical features for
related data. Arumugam et al. [17] highlights the decision tree predicting heart disease vulnerability based on 14 essential
model’s consistent superiority and its fine-tuning for optimal symptoms, exceeding the 10 features typically considered.
forecasting of heart disease likelihood in diabetic individuals, Valuable for doctors, it employs a comparative analysis of
contributing valuable insights to healthcare data mining machine learning techniques, revealing Random Forest as
applications. In the realm of healthcare challenges, Kre- the most accurate and reliable algorithm. The system not
soja et al. [18] focus on early-stage heart disease detection, only aids in classification but also explores the nuanced
particularly in underserved areas lacking specialized medical relationship between diabetes and its impact on heart
professionals. Leveraging machine learning, the analysis disease, providing essential insights for medical profession-
explores active learning methods to enhance classification als. Dwivedi [23] evaluate the potential of six machine
quality with sparsely labeled data. Applying five selection learning techniques for heart disease prediction, assessing
strategies, the experimentation optimizes hyperparameters their efficiency on eight diverse classification indices and
and evaluates accuracy and F-score for heart disease pre- the receiver operating characteristic (ROC) curve. Logistic
diction. Yields highlight the label ranking model’s superior regression emerges with the highest classification accuracy
accuracy, emphasizing its potential for generalization beyond at 85%, accompanied by a sensitivity of 89% and specificity
existing data. This work contributes to advancing preventive of 81%. Contributing valuable insights, this investigation
healthcare through optimized labeling strategies in machine into the effectiveness of machine learning methods aims
learning applications for heart disease prediction. for prompt and accurate identification of heart disease.
Abdulsalam et al. [19] introduced a Bagging-QSVC Bertsimas et al. [14] address the global challenge of heart-
model, which integrates a quantum support vector classifier. related anomalies, emphasizing the asymptomatic nature of
Their comparative experiments on the Cleveland dataset patients until critical events. Leveraging Machine Learning
achieved 90.16% accuracy, demonstrating the promise of and digital Electrocardiograms (ECG), the study proposes a

VOLUME 13, 2025 138315

 S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

real-time prediction methodology with remarkable accuracy TABLE 1. Summary of literature review on machine learning for heart
disease prediction.
in less than 30 milliseconds. Using a dataset of 40 thousand
labeled ECGs, the models achieve high outcomes, detecting
seven types of signals with F1 Scores of 0.93. This work
represents a pioneering effort in achieving accuracy across
diverse settings in ECG anomaly detection.
A summary of the literature review is shown in Table 1.
Deep learning has recently improved medical image analysis,
especially when it comes to separating different body parts
and finding diseases. They put forward a UDA framework
by combining CycleGAN with VAMCEI to resolve the
domain shift in myocardial segmentation, resulting in leading
outcomes among various types of cardiac MRI scans [1].
Chen et al. showed TransUNet, a method that unifies the
capacity of Transformers to see the whole image with U-Net’s
ability to pinpoint details. It has worked very well in many
medical segmentation areas, for example, by finding both
large and small parts of a cardiac MRI scan [2]. Besides car-
diac imaging, these networks have demonstrated potential in
improving diagnostic skills. Singh and colleagues introduced
a preprocessing model with CycleGAN to reduce distortions
in chest X-rays, significantly increasing how well lung
disease was identified in ChestX-Ray14 [3]. They introduced
nnU-Net, a self-configuring neural network that performs
better than individual, specialized models in 23 medical
segmentation challenges, allowing anyone to use it without
dependence on experts [4]. Because of these improvements,
it’s now clear that generative and attention-based mod-
els are important in medical image processing, proving
the necessity and usefulness of our proposed GenDeep
framework for reliable cardiac disease recognition from
cine-MRI.
The literature review provides an overview of relevant
prior exploration using machine learning approaches for
automated heart disease prediction and cardiovascular risk
assessment. Key techniques studied are clustering, multilayer
perceptrons, LSTM networks, quantum ML, decision trees,
and gradient boosting. The review discusses recent advances
in deep learning for cardiac image analysis, especially CNNs
for detection, segmentation, and diagnosis. Highlights the
potential of AI innovations in modern cardiology, both cardiomyopathy, heart failure with infarction, and right
to enhance clinical decisions through decision support ventricular abnormality. For each patient, 28 to 40 short-
systems, as well as enable large-scale screening initiatives axis cine-MRI frames are available, with expert annotations
for early disease identification. However, also outlines exist- for end-diastolic and end-systolic phases, identifying the
ing limitations regarding real-world adoption, underscoring left ventricle (LV) and right ventricle (RV) structures. Each
the need for further multidisciplinary theory addressing 2D slice measures 235 × 263 voxels, with a resolution
generalizability gaps. Overall, transformational potential is ranging from 1.37 to 1.68 mm and a thickness of 5-10 mm.
conveyed but balancing improved predictive outcomes with Scans were acquired using both 1.5T and 3.0T MRI scanners
translational barriers. over six years at the Hospital of Dijon, resulting in over
4000 annotated cine-MRI images as shown in 2. The
III. GENDEEP: A COMPOSITE FRAMEWORK FOR ROBUST cine-MRI scan image samples from the ACDC dataset are
CARDIAC DISEASE PREDICTION AND DIAGNOSIS illustrated in 1.
A. DATASET DESCRIPTION Although the original ACDC dataset was built for five main
This study utilizes the ACDC dataset, consisting of cine-MRI heart conditions, we have broadened its classification to ten
scans from 100 patients. Each patient belongs to one of five types that are clinically significant. The need for the model
categories: normal, dilated cardiomyopathy, hypertrophic to focus on and handle important cardiac pathologies seen in

138316 VOLUME 13, 2025

S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

FIGURE 1. Cine-MRI scans images samples from ACDC dataset.

daily clinical work led to this extension. Our classification

task concerns ten diseases, including: Coronary Artery
Disease (CAD), Myocardial Infarction (Heart Attack), Heart
Failure, Arrhythmias, Valvular Heart Diseases, Cardiomy-
opathy, Hypertension (High Blood Pressure), Peripheral
Artery Disease (PAD), Congenital Heart Diseases, and
Rheumatic Heart Disease. These classes were formed by
both relabeling in the clinic and morphological analysis
using segmentation from the original data. The results from FIGURE 2. Workflow diagram of the GenDeep model integrating
DeepLab were used to spot important patterns (such as composite GAN and DeepLab for automated cardiac disease prediction
through robust feature learning and image synthesis.
thinning walls, large ventricles, valve problems) and, together
with clinical information, helped allocate the original cases to
one of the ten expanded categories. We split the dataset using to find differences between authentic and fake images. The
stratified sampling so that 80% was used for training, 10% results from the GAN adversarial loss and the DeepLab
for validation, and 10% for testing, ensuring each of the ten segmentation loss are mixed by a weighted sum, with
disease categories remained equally represented. the balance coefficient λ assigned a value of 0.4 after
testing different values. Within DeepLab, the DeepLabV3+
TABLE 2. Statistics and characteristics of the ACDC Cine-MRI images architecture is used on top of a ResNet-101 backbone.
dataset. The Atrous Spatial Pyramid Pooling (ASPP) module takes
advantage of dilated convolutions at rates of 6, 12, and
18. This technique allows the model to analyze features at
multiple scales without down-sampling the image, thereby
preserving crucial spatial resolution for precise segmentation.
After getting the segmentation, CRFs are run to improve the
boundaries. The model segments both ventricles, and both
contributions are judged by the appearance and structure of
the ventricles. After receiving adversarial training, GenDeep
can differentiate between multiple cardiac issues such as
cardiomyopathy, hypertrophy, infarction, and others. GANs
help generate data strong enough for any situation, and
DeepLab aids precise separation of structures in medical
images. Testing the model with unseen data proves its strong
ability to detect disease. Strong potential for cardiac disease
B. GENDEEP ARCHITECTURE OVERVIEW detection and clinical decision support is shown in GenDeep
The proposed methodology introduces a new GenDeep by using synthetic augmentation to fix the problem of limited
model that combines an unsupervised Composite GAN data. The overall workflow diagram of the GenDeep model is
with DeepLab to allow automated prediction of cardiac presented in the 2.
diseases directly from cine-MRI scans. Images from the
ACDC dataset (over 4000) are preprocessed using Apache C. DATA AUGMENTATION AND FEATURE FUSION IN THE
Spark on Hadoop to ensure data is processed efficiently and DEEPLAB MODEL FOR ENHANCED CARDIAC
accurately. Similar to modified DCGAN, in the Composite SEGMENTATION
GAN, the generator is made of transposed convolutional To improve segmentation performance and model gen-
layers, with batch normalization and LeakyReLU activations, eralization, both traditional augmentation techniques and
to synthesize cardiac MR images that resemble real images. GAN-based synthetic image generation were applied. Tradi-
Convolution layers make up the discriminator, and it learns tional data augmentation included random rotations (±15◦ ),

VOLUME 13, 2025 138317

 S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

1) ENCODER MODULE FOR FEATURE EXTRACTION

Fk = ReLU (BN (Wk ∗ Fk−1 + bk )) (1)

where Fk is the k th feature map, ReLU applies rectified linear
unit activation, BN refers to batch normalization, Wk and
bk are learned weight and bias parameters. This equation
defines the sequence of convolutional and pooling layers in
the Encoder module to hierarchically extract visual features
from the input cardiac MRI while reducing spatial resolution.
Batch normalization and ReLU activation facilitate training
this CNN pipeline.

2) DECODER MODULE FOR RESOLUTION RECOVERY

Ok = U ReLU BN Wk′ ∗ Ok+1 + b′k

FIGURE 3. DeepLab model Workflow, data augmentation, (2)
encoder-decoder feature extraction, and fusion for improved
generalization in cardiac segmentation. Here, Ok refers to the k th decoder output, U denotes upsam-
pling operation, Wk′ and b′k are decoder weights and biases.
This equation specifies the Decoder module’s workflow,
horizontal and vertical flipping, zoom scaling in the range
applying interleaved upsampling, convolution, and batch
of 90% to 110%, contrast modulation, and Gaussian noise
normalization to incrementally recover the original input
addition. These transformations were applied on-the-fly dur-
resolution from the encoder feature maps. Skip connections
ing training using PyTorch’s built-in augmentation pipeline.
transfer encodings across the modules as well.
In addition to this, a trained composite GAN was employed
to generate synthetic MRI images that reflect diverse
3) INPUT FUSION FOR AUGMENTATION
but clinically plausible cardiac morphologies. For each
original image, two GAN-generated variants were created, J = Fm ⊚X (3)
effectively expanding the dataset size from approximately
4,000 to 12,000 images. These augmentations enriched the This fuses the final decoder output Fm with the original input
training set and provided the DeepLab model with enhanced image X through element-wise multiplication, represented by
morphological variability, which in turn improved its ability ⊚, to yield the augmented version J . The equation captures
to segment cardiac structures across pathological cases. The how the autoencoder architecture, combining complementary
DeepLab model depicted performs data augmentation on the encoder and decoder pathways, processes cardiac MRIs to
original input images before feeding them into the actual output augmented variants with enhanced morphological
DeepLab segmentation network. First, the original images diversity.
are passed through an Encoder module comprising repeated
convolution and max pooling layers to extract feature maps. 4) DEEPLAB SEGMENTATION
These feature maps encode hierarchical visual information Y = σ (Wd ∗ A (X ; 2) + bd ) (4)
but at significantly reduced spatial resolution. They are then
processed by the Decoder module, which employs upsam- Here, Wd , bd denote DeepLab model parameters, σ is
pling and convolution operations to recover the original the sigmoid activation, and A represents the core DeepLab
input resolution, as shown in Figure 3. Skip connections architecture comprising backbone, atrous spatial pyramid
between the encoder and decoder transfer features across pooling, and convolution heads. Overall, this equation defines
and concatenate them to retain spatial details. The decoder the full DeepLab pipeline that takes augmented cardiac MRI
output is fused with the original image through element-wise X as input to produce a segmentation mask Y , leveraging data
multiplication to produce an augmented version of the input. augmentation through feature extraction and fusion to boost
This augmented data better captures aspects that may be accuracy.
underrepresented in the original dataset. Additionally, batch
normalization and leaky ReLU activation are utilized between D. GAN-BASED AUTOMATED CARDIAC DIAGNOSIS WITH
convolution layers to facilitate training. The augmented AUGMENTED MRI IMAGES
images are finally passed into the main DeepLab architecture, This GAN framework leverages the augmented cardiac
comprising the backbone, atrous spatial pyramid pooling, and MR images produced by the DeepLab model to enable
convolution heads for semantic segmentation. By enhancing automated diagnosis of cardiac conditions. It comprises a
diversity through data augmentation, the model is able Generator and Discriminator network trained adversarially.
to achieve better generalization for cardiac segmentation The Generator receives an input vector, which is transformed
across morphological variability in the presentation of heart through an expanding sequence of convolutional layers into a
diseases. synthetic fake MRI image. This is fed into the Discriminator

138318 VOLUME 13, 2025

S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

along with a real cardiac MRI image from the original

dataset. A scalar probability indicating whether the input
image is real or fake is produced by the discriminator.
The weights of both networks are rationalized based on
this feedback through backpropagation. Concluded iterative
training, the Generator absorbs to produce progressively
representative fake MRIs while the Discriminator becomes
an increasingly better detector. Once trained, the Generator
can synthesize pathological imagery, and corresponding
label vectors can be passed through the Discriminator, now
serving as an investigative classifier, to predict potential
cardiac abnormalities. For instance, morphological patterns
in the augmented left and right ventricle segmentations
can indicate conditions like cardiomyopathy, ventricular FIGURE 4. GAN framework for cardiac diagnosis leveraging
DeepLab-generated augmented images for adversarial training, realistic
hypertrophy, or infarction. This composite GAN framework image synthesis, and diagnostic classification.
capitalizes on the representational power of generative
models to construct robust classifiers despite scarce patient
5) DISCRIMINATOR MODEL FOR CLASSIFICATION
data, conquering restrictions permeating therapeutic imaging
datasets. The augmentation module boosts diversity to elevate y = σ (Wd ∗ X + bd ) (6)
generalization across heterogeneous demographics. Figure 4
delineates the GAN architecture for automated cardiac Here, y mentions the predicted label, X embodies the
disease prediction. involvement image, real or fake, Wd , bd constitute Discrimi-
nator weights and biases, σ applies sigmoid activation. This
1) INTERPRETATION OF MORPHOLOGICAL PATTERNS equation stipulates the Discriminator component that pro-
A thorough examination discloses the morphological shapes cesses input cardiac scans to output a pathology classification
in the augmented segmentations of the left and right label forecasting disease status, in totaling to categorizing
ventricles. By running the synthetic imagery through the real/fake.
diagnostic classifier, probable cardiac abnormalities can be
identified. The precision of the model’s estimate is improved 6) ADVERSARIAL LOSS FUNCTION
by the methodical interpretation of variations suggestive of
Ladv = Ex pdata [logD (x)] + Ex p(z) log (1 − D (G (z)))
 
cardiomyopathy, ventricular hypertrophy, or infarction.
(7)
2) ENHANCING DIAGNOSTIC PRECISION
The adversarial loss function proves instrumental in training
Including GAN-generated scans progresses diagnostic accu- the composite Generative Adversarial Network (GAN),
racy while also elevating the dataset. When the classifier is orchestrating the Generator G and Discriminator D com-
uncovered to a variety of synthetic examples, it becomes more ponents. The original term of the loss function penalizes
adept at identifying and predicting subtle variations in cardiac the Generator for constructing unrealistic synthetic images
structure under a wide range of conditions. by assessing how well the Discriminator discriminates them
from real ones. Simultaneously, the second term incentivizes
3) ROBUSTNESS ACROSS DEMOGRAPHICS the Generator to craft synthetic images resounding enough to
Besides, augmentation improves the robustness of the model deceive the Discriminator into classifying them as authentic.
in the face of miscellaneous demographics. Synthesis intro- This adversarial tug-of-war ensures that the Generator refines
duces variability, enabling powerful generalization across its ability to generate more authentic and diagnostically
numerous patient populations and clinical settings. relevant cardiac MRI images over time. The optimization
of this composite objective function not only enhances the
4) GENERATOR MODEL FOR SYNTHESIS realism of generated medical images but also augments the
Discriminator’s proficiency in accurately classifying pathol-
If = σ Wg ∗ z + bg


(5) ogy, contributing to the overall efficacy of the automated
where If denotes the produced fake image, z represents cardiac diagnosis system.
the input noise vector, Wg and bg indicate Generator
7) COMPOSITE MODEL OBJECTIVE
prototypical parameters. The σ activation function applies a
Tanh transformation, and ∗ is the convolution operation. This minG max D Ladv + λLseg (8)
convolution operation maps random noise trajectories into
synthetic cardiac MRIs through an intensifying CNN Gen- Here, Lseg refers to the segmentation loss between predicted
erator pipeline to enable realistic medical image generation. and ground truth masks. The compound objective balances

VOLUME 13, 2025 138319

 S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

adversarial diagnosis training alongside precise anatomical Algorithm 1 Automated Cardiac Disease Prediction
delineation. Tuning the λ hyperparameter controls the Input:
tradeoff, enabling the model to learn both robust classification Dataset: Heart disease dataset with patient records
features as well as accurate segmentation simultaneously. Features: Key attributes (age, cholesterol, BP, etc.)
ML_Models: Machine learning models (KNN, RF, Logistic
E. DISTRIBUTED BIG DATA PIPELINE Regression, etc.)
This big data pipeline leverages a Hadoop cluster for Hyperparameters: Model-specific hyperparameters
scalable storage and parallel processing of the large ACDC Preprocessing_Config: Data preprocessing configuration
cardiac MRI dataset, enabling efficient preprocessing. The Training_Epochs: Number of iterations for training
original dataset is divided into splits and stored across the
Hadoop Distributed File System (HDFS), which replicates Output:
partitions across cluster nodes to provide redundancy and Trained_Model: Optimized model for heart disease
fault-tolerance. Spark modules are deployed on the cluster to prediction
take advantage of in-memory processing for accelerated data Predictions: Heart disease classification results
transformations. The workflow begins with the dataset splits
sorted and grouped by common keys to rearrange related Procedure: Automated_Cardiac_Prediction
entries as shown in Figure 5. This facilitates subsequent (Dataset, Features, ML_Models, Hyperparameters,
aggregation operations. The mapped & reduced outputs Preprocessing_Config, Training_Epochs)
across nodes are shuffled to route datasets to respective 1) Data Preprocessing:
Spark operations and jobs. Finally, the intermediate outputs Normalize numerical features, handle missing
from the distributed preprocess are merged to yield the final values, and encode categorical attributes using
aggregated preprocessed dataset. This architecture provides Preprocessing_Config.
a robust and time-efficient pipeline to handle terabyte-scale 2) Feature Selection and Engineering:
medical imaging repositories. The hybrid ecosystem marries Identify the most significant predictors using feature
Spark’s speed with Hadoop’s storage capabilities. Automated importance techniques (e.g., mutual information, cor-
cardiac diagnosis systems stand to gain immense scalability relation matrix).
and achievement benefits from such large-scale distributed 3) Model Training with Iteration:
computing frameworks in order to intake clinical imaging For each model in ML_Models do:
data from diverse hospitals and scanners. The ability to
a) Initialize model with Hyperparameters
efficiently manage voluminous health records is key to
b) For each epoch in Training_Epochs do:
training more holistic machine learning models for enhanced
generalization. An algorithm of the said methodology is i) Train the model on the dataset
shown in 1. ii) Compute loss function and update weights
Using a pipeline based on Hadoop and Spark for the iii) Evaluate model performance on validation
ACDC database may seem like overkill, but it follows the data
infrastructure at the institution designed for analyzing multi- c) Store trained model and performance metrics
center images, many of which have hundreds of thousands. 4) Evaluate Model Performance:
This pipeline was built to work with this dataset and also grow For each trained model, compute performance metrics
with future federated and hospital deployments. Yet, when the (Accuracy, Precision, Recall, F1-score, AUC) on
work is limited to the ACDC data alone, data management validation data.
can also be done efficiently using regular data loaders from Select the best-performing model based on evaluation
PyTorch or TensorFlow, with augmentation carried out as scores.
images are fed. Similar results were noticed once we used 5) Final Prediction:
PyTorch’s internal pipeline for checking data and training. Use the selected model to predict heart disease
As a result, the proposed framework can be used consistently probability on test data. Store Predictions for further
across many different deployment settings, regardless of their analysis.
size. 6) Post-Evaluation and Refinement:
While model performance is below the desired
IV. RESULTS AND DISCUSSION threshold:
This section systematically evaluates the performance of the a) Tune hyperparameters and retrain the model
proposed GenDeep framework for cardiac MRI analysis and b) Re-evaluate model performance
automated pathology classification tasks. Extensive compar-
Return: Trained_Model, Predictions
ative evaluations are performed on a held-out test set from
the ACDC dataset. Quantitative results are reported across
diverse evaluation metrics spanning segmentation accu- operating characteristics, and computational efficiency. Key
racy, multi-class classification precision and recall, receiver empirical The analysis highlights GenDeep’s capabilities

138320 VOLUME 13, 2025

S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

FIGURE 6. Training and validation accuracy trends of the GenDeep model

reflecting incremental learning and overfitting dynamics in automated
cardiac pathology prediction.
FIGURE 5. Big data pipeline using hadoop and spark for scalable storage,
parallel processing, and efficient preprocessing of large-scale cardiac MRI
datasets.

in learning discriminative features, synthesizing realistic

pathological images, and precisely segmenting cardiac
anatomical structures. Comparative discussions weigh the
improvements against state-of-the-art approaches regarding
diagnostic precision under sparse data constraints. Ablation
studies assess the effects of parameter tuning, such as
weight decay, dataset splitting, and batch size, on model
generalization. Both quantitative metrics and qualitative
visualization provide multi-faceted insights into the efficacy,
robustness, and accessibility of the model. The empirical FIGURE 7. Training and validation loss dynamics illustrate the learning
evidence converges to highlight GenDeep’s transformational and generalization challenges in automated cardiac pathology prediction
potential to augment automated analysis in cardiology, using gendeep model.

guiding time-critical screening and interventions.

datasets, reinforcing the need for data augmentation. These
A. TRAINING AND VALIDATION ACCURACY AND LOSS results support future work on larger cardiac MRI repositories
METRICS OF COMPOSITE GAN-DEEPLAB FOR CARDIAC to enhance generalizability to real-world clinical scenarios.
PATHOLOGY PREDICTION The training and validation accuracy trends are shown in
For the training process, 1000 epochs were used to guarantee Figure 6.
that the GAN and DeepLab parts both converge well. Due to Loss values for the training and validation processes
augmentation methods and composite GAN, our final dataset decreased from 2.5 to 0.2 and from 3.5 to 0.5, suggesting
was more than triple its original size, with approximately that the models learned well. It highlights the challenge
12,000 total samples from 4,000 real images. Considering of generalizing from limited labeled data and handling
that GAN training is challenging because it quickly becomes morphological variability. These findings suggest that better
unstable and must be optimized for long periods, we used a regularization is required, and studying multi-center cardiac
longer training plan. The researchers discovered that using MRI data can create more reliable diagnostic AI systems.
a small number of epochs caused the program to fail to The GenDeep model achieved strong performance in
learn well and delivered inconsistent segmentation results in cardiac pathology classification, with 97% accuracy, 91%
early-stage or minimally changed disease instances. To avoid precision, 94% recall, and 93% F1-score. AUC reached
overfitting, early stopping was applied, stopping the training 96%, and low error values (MSE 2.21, RMSE 2.77, MAE
after performance reached a stable or downward trend. 1.8) confirmed smooth convergence. These results validate
The GAN and DeepLab model was trained and validated GenDeep’s ability to learn complex cardiac patterns using
on over 4000 ACDC cine-MRI scans. Training accuracy GAN-based augmentation and DeepLab segmentation. The
improved from 45% to 97% by the 1000th epoch, while automated pipeline offers potential for accurate, scalable
validation accuracy reached 96%, indicating successful learn- MRI-based screening and early disease detection in clinical
ing with minor overfitting. The early plateau in validation settings. These results validate GenDeep’s ability to learn
performance highlights the challenges of small annotated complex cardiac patterns using GAN-based augmentation

VOLUME 13, 2025 138321

 S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

FIGURE 10. Representation of the receiver operating characteristic(ROC)

for each cardiac disease prediction class.

FIGURE 8. Showcasing the overall model’s power in automated cardiac

MRI analysis.

FIGURE 11. Evaluating the effects of weight decay on GenDeep model

metrics for automated cardiac pathology prediction.

98.9% accuracy, while valvular and peripheral artery diseases

showed slightly lower precision around 96.2-97.1%, likely
FIGURE 9. The confusion matrix offers a comprehensive view of the
due to morphological similarity. Overall, per-class accuracy
model’s accurate classification across diverse cardiac disease categories, exceeded 97%, demonstrating the model’s robustness and
showcasing distinct diagnostic patterns and strong efficiency metrics. potential as a reliable AI-assisted diagnostic tool in clinical
cardiology.
and DeepLab segmentation. In essence, the model demon- The ROC curves Figure 10 demonstrate GenDeep’s strong
strates a high degree of reliability in distinguishing between diagnostic capability across 10 cardiac conditions, with
healthy and diseased hearts based on MRI scans. AUC values ranging from 0.92 to 0.95. These consistently
To prove the computational savings of the proposed high scores reflect excellent sensitivity and specificity. Most
architecture, additional performance metrics are offered. classes achieved 93-95 accuracy, validating the model’s
There are about 42 million trainable parameters in the ability to distinguish complex cardiac patterns. The empir-
DeepLabV3+ component, which uses a ResNet-101 back- ical evidence confirms the generalization strength of the
bone. The resulting GenDeep framework controls about GAN-DeepLab framework and its clinical potential for
54 million parameters, since its composite GAN part includes AI-assisted cardiac screening.
both generator and discriminator networks. The model can
execute about 62 GFLOPs each time it is run forward. The B. IMPACT OF DIFFERENT WEIGHT DECAY, DATASET
training process averages 2.7 minutes on an NVIDIA RTX SPLITTING, AND BATCH SIZE IN GENDEEP ARCHITECTURE
3090 card. For the full training to be done across 1000 epochs, FOR CARDIAC PATHOLOGY PREDICTION
it took almost 45 hours. It is clear from these metrics that the Table 3 highlights the effect of weight decay tuning on Gen-
model offers a good balance between design complexity and Deep’s performance. A decay of 0.1 yielded 89.22 accuracy
how easy it is to apply in practice. with higher loss values (MSE 3.47), while reducing it to
The confusion matrix, see Figure 9, confirms GenDeep’s 0.01 improved all metrics. Optimal results were achieved at
strong classification performance across 10 cardiac condi- 0.001, with 97.32 accuracy, 91.31 precision, 92.64 recall, and
tions, with true positives ranging from 919 to 958 out of the lowest MSE of 2.37 as shown in Figure 11. These results
960 cases per class. Coronary artery disease, myocardial confirm that smaller weight decay enhances convergence and
infarction, arrhythmias, and hypertension achieved over diagnostic accuracy by better controlling regularization.

138322 VOLUME 13, 2025

S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

TABLE 3. Weight decay tuning’s influence on composite GAN-DeepLab TABLE 4. Exploring the influence of dataset splitting on GenDeep model
model capability in automated cardiac pathology classification. metrics for cardiac pathology classification.

Table 4 evaluates the impact of dataset splitting on the

predictive outcome of the GenDeep architecture comprising
composite GAN and DeepLab on the ACDC cardiac MRI
dataset. A shallow base model with training data 50% and
testing data 50% achieves accuracy of 85.43%, recall of
84.23%, F1 score of 86.36%, and precision of 83.65%.
However, the loss metrics of RMSE at 6.87, MAE at 7.55, and
MSE at 8.24 are comparatively high. Increasing training data
by 60% and testing data by 40% leads to clear improvements
with accuracy rising to 89.35%, recall at 85.44%, F1 score
hitting 87.41%, and precision reaching 84.58%. The RMSE,
MAE, and MSE values also drop, indicating more robust
learning. Further training data 70% and testing data 30% FIGURE 12. Exploring and evaluating the influence of the dataset
pushes accuracy above 90% to 92.78% and achieves the splitting on GenDeep model metrics.
highest recall of all models at 93.88% while maintaining F1
score at 90.66% and precision at 87.35%. Finally, the training
data 80% and testing data 20% variant maximizes functioning
across the board, achieving 97% accuracy, 93.88% recall,
92.65% F1 score, and 91.77% precision as depicted in
Figure 12 while minimizing losses to RMSE of 2.11, MAE
of 1.25, and MSE of 2.22 as demonstrated in Figure 13. The
consistent gains with increasing depth verify the importance
of model capacity to capture the complex morphological
patterns and inter-class variances in cardiac pathologies. The
results motivate leveraging larger datasets and computational
resources to train deeper models for reliable automated
diagnosis systems.
We systematically investigated the impact of varying the FIGURE 13. Evaluation and measuring the error rate of the dataset
splitting on GenDeep model metrics.
batch size on the proficiency metrics of our automated cardiac
disease prediction model, which leverages a Composite GAN
and DeepLab architecture within the realm of machine superior performance with an accuracy of 97.45%, precision
learning. The model was trained and evaluated using three of 90.51%, recall of 90.88%, and F1-Score of 89.65%.
different batch sizes, namely 8, 32, and 64. Our findings Moreover, the mean squared error (MSE) score was notably
revealed notable disparities in the predictive capabilities low at 2.21%, and the root mean squared error (RMSE) and
of the model across the evaluated parameters. Specifically, mean absolute error (MAE) scores were equally minimized
when utilizing a batch size of 8, the model demonstrated at 2.77% and 1.8%, respectively. Contrastingly, employing a

VOLUME 13, 2025 138323

 S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

F1 score reaching 93%. This surpasses prior benchmarks

across all key metrics, highlighting the synergistic benefit
of integrating GAN-based data generation with DeepLab’s
segmentation capabilities. Key advantages include capturing
complex morphological patterns, synthesizing augmented
training data to overcome dataset constraints, and encoding
inter-slice spatial contexts for whole heart quantification.
Smooth convergence is verified with a low RMSE of 2.77,
MAE of 1.8, and MSE of 2.21. Comparatively, hybrid ML
approaches in literature achieved top accuracy around 91%,
while precisions spanned 83-90% and F1 scores peaked at
89%. Loss metrics also trended higher by factors of 1.5x
FIGURE 14. Unraveling the influence of different batch sizes on GenDeep to 3x. The consistent margin illustrates the robustness of
model metrics in automated cardiac disease prediction. the design methodology. The fully automated end-to-end
pipeline could accelerate adoption in clinical settings.
larger batch size of 32 led to a reduction in results metrics,
with accuracy, precision, recall, and F1-Score registering TABLE 7. Comparison of automated cardiac disease prediction results
with different deep learning models.
at 92.23%, 81.23%, 86.58%, and 82.35%, respectively,
portrayed in the 14. The corresponding MSE, RMSE, and
MAE values demonstrated higher magnitudes of 4.35%,
6.35%, and 4.98%, respectively, as shown in 5. This high-
lights the model’s sensitivity to batch size variations. These
observations accentuate the essence of meticulous parameter
tuning for elevating predictive efficiency and accuracy, with
smaller batch configurations exhibiting optimal performance.
Table 1 portrays the influence of batch dimension on
GenDeep metrics.
TABLE 5. Effect of batch size on model performance metrics.

The bar graph presents a comprehensive comparison of

the GenDeep Model with several literature models, including
Hybrid Decision Support, HRFLM, ML Algorithms, Logistic
Regression, and C-BiLSTM. Key execution metrics such as
Table 6 illustrates the impact of GAN-based augmenta- accuracy, precision, recall, F1-Score, AUC, RMSE, MAE,
tion on model performance. Significant improvements are and MSE are evaluated. The GenDeep Model dependably
observed across all metrics when GAN is applied, confirming outclasses its counterparts, showcasing greater accuracy,
its effectiveness in enhancing generalization and robustness. precision, recall, and overall model execution. Notably, C-
BiLSTM also demonstrates strong proficiency, particularly in
TABLE 6. Effect of GAN-Based augmentation on model performance
(Batch size = 8).
accuracy, recall, and AUC. Logistic Regression models reveal
competitive results, emphasizing their adaptability. HRFLM
and ML Algorithms offer balanced proficiency across
multiple metrics. This visual representation highlights the
GenDeep Model’s superiority, positioning it as an advanced
and promising solution in comparison to established models
in the literature.
The line graph demonstrates a detailed comparison of
specific evaluation metrics, namely RMSE (Root Mean
Squared Error), MAE (Mean Absolute Error), and MSE
C. COMPARING OUR METHODOLOGY WITH (Mean Squared Error), across different models with a primary
SOPHISTICATED EXISTING SYSTEMS emphasis on the proposed GenDeep Model. Graphical
The GenDeep model achieves state-of-the-art outcomes visualization suggests significantly lower deviations between
for automated cardiac pathology classification, with an GenDeep’s predicted and ground truth targets. This implies
accuracy of 97%, precision of 91%, recall hitting 94%, and superior accuracy and precision in the GenDeep Model’s

138324 VOLUME 13, 2025

S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

contribution is the use of a GAN framework for data

augmentation, enabling the synthesis of realistic pathological
images and improving model generalization across diverse
patient groups. The DeepLab component leverages inter-slice
spatial context to perform precise whole-heart anatomical
segmentation. Extensive ablation studies on weight decay,
dataset partitioning, and batch size tuning offered insights
into optimizing deep learning pipelines for cardiology
applications. Our diagnostic system is designed to seamlessly
integrate into clinical workflows to expedite the screening
and early detection of cardiac diseases, thereby guiding
timely intervention. The high classification accuracy suggests
FIGURE 15. Accomplishment and error rate contrast with state-of-the-art GenDeep can serve as a powerful decision support tool.
models. For instance, it could be integrated with a hospital’s Picture
Archiving and Communication System (PACS), automati-
cally analyzing incoming cine-MRI scans to flag high-risk
cases for priority review by a cardiologist. This approach
does not replace expert clinical judgment but rather augments
it, helping to reduce the manual interpretation burden and
mitigate the subjective variability that can exist between
clinicians. By providing an objective and quantitative
assessment of ventricular morphology, the framework can
enhance diagnostic confidence and consistency. Furthermore,
population-level screening programs could leverage this
automated tool to identify at-risk individuals in a more
efficient and scalable manner, directly contributing to more
effective preventative cardiovascular healthcare policies.
FIGURE 16. Comparison of RMSE, MAE, and MSE metrics across models, Overall, this study highlighted the transformative potential of
highlighting the superior performance of the proposed GenDeep model.
AI to enhance cardiovascular outcomes.
Future work should focus on improving the robustness
predictions, with the lowest errors in guesstimating the and accessibility of automated cardiac diagnosis systems
differences between predicted and actual values. The graph to support clinical adoption. Priorities include aggregating
provides a succinct visual representation, emphasizing the larger multi-center cardiac MRI datasets and boosting
higher achievement of the GenDeep Model in terms of these model generalizability across heterogeneous scanners and
critical metrics, showcasing its potential as a highly accurate acquisition protocols. Incorporating self-supervised and few-
and steadfast predictive model. shot learning techniques may reduce reliance on large
annotated datasets, particularly for rare pathologies. Testing
V. CONCLUSION AND FUTURE WORK model effectiveness on diverse ethnic demographics is also
This study introduced GenDeep, a novel composite deep vital to ensure equity and fairness. Longitudinal evaluations
learning framework that integrates an unsupervised Condi- across multiple patient visits may support the development of
tional GAN and a DeepLab network for automated cardiac prognostic models for long-term cardiac monitoring. From an
pathology classification from cine-MRI scans. The model implementation perspective, optimized model compression
was rigorously trained and tested on the ACDC dataset and porting onto cloud, mobile, and edge devices could
comprising over 4000 annotated MRI images positive for broaden access and enable personalized telecardiology. User-
conditions like cardiomyopathy, ventricular hypertrophy, centric studies assessing interpretability would also build
and infarction. The proposed methodology achieved strong trust amongst the clinician community. Filling these gaps
performance, with 97% classification accuracy. 91% preci- through interdisciplinary collaboration is essential to fully
sion, 94% recall and 93% F1-score, significantly exceeding realizing the potential of AI in cardiology while navigating
state-of-the-art benchmarks. Convergence was confirmed practical implementation challenges.
by low error values: RMSE = 2.77, MAE = 1.8, and
MSE = 2.21. Comparative experiments demonstrated con- VI. DATA AVAILABILITY
sistent performance gains over hybrid machine learning Inquiries about data availability should be addressed to the
techniques reported in the literature, owing to the comple- authors.
mentary representational strengths of GANs and DeepLabs
in learning robust feature encodings from medical images VII. CONFLICT OF INTEREST
while precisely delineating anatomical structures. A key No conflict of interest has been declared by the authors.

VOLUME 13, 2025 138325

 S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

REFERENCES [21] P. Rani, R. Kumar, N. M. O. S. Ahmed, and A. Jain, ‘‘A decision

[1] F. Bray, M. Laversanne, E. Weiderpass, and I. Soerjomataram, ‘‘The ever- support system for heart disease prediction based upon machine learning,’’
increasing importance of cancer as a leading cause of premature death J. Reliable Intell. Environments, vol. 7, no. 3, pp. 263–275, Sep. 2021.
worldwide,’’ Cancer, vol. 127, no. 16, pp. 3029–3030, Aug. 2021. [22] P. Rubini, C. Subasini, A. Katharine, V. Kumaresan, S. Kumar, and
[2] G. Mensah, G. Roth, and V. Fuster, ‘‘The global burden of cardiovascular T. Nithya, ‘‘A cardiovascular disease prediction using machine learning
diseases and risk factors: 2020 and beyond,’’ J. Amer. College Cardiol., algorithms,’’ Ann. Romanian Soc. Cell Biol., vol. 25, no. 2, pp. 904–912,
vol. 74, no. 20, pp. 2529–2532, 2019. 2021.
[23] A. K. Dwivedi, ‘‘Performance evaluation of different machine learning
[3] M. F. Faienza, F. Urbano, G. Lassandro, F. Valente, G. D’Amato,
techniques for prediction of heart disease,’’ Neural Comput. Appl., vol. 29,
P. Portincasa, and P. Giordano, ‘‘The cardiovascular disease (CVD) risk
no. 10, pp. 685–693, May 2018.
continuum from prenatal life to adulthood: A literature review,’’ Int.
J. Environ. Res. Public Health, vol. 19, no. 14, p. 8282, Jul. 2022. [24] P. Dileep, K. N. Rao, P. Bodapati, S. Gokuruboyina, R. Peddi, A. Grover,
and A. Sheetal, ‘‘An automatic heart disease prediction using cluster-based
[4] K. K. L. Wong, G. Fortino, and D. Abbott, ‘‘Deep learning-based
bi-directional LSTM (C-BiLSTM) algorithm,’’ Neural Comput. Appl.,
cardiovascular image diagnosis: A promising challenge,’’ Future Gener.
vol. 35, no. 10, pp. 7253–7266, Apr. 2023.
Comput. Syst., vol. 110, pp. 802–811, Sep. 2020.
[5] M. Ullah, S. Hamayun, A. Wahab, S. U. Khan, M. U. Rehman, Z. U. Haq,
K. U. Rehman, A. Ullah, A. Mehreen, U. A. Awan, M. Qayum, and
M. Naeem, ‘‘Smart technologies used as smart tools in the management
of cardiovascular disease and their future perspective,’’ Current Problems
Cardiology, vol. 48, no. 11, Nov. 2023, Art. no. 101922.
[6] N. Ghaffar Nia, E. Kaplanoglu, and A. Nasab, ‘‘Evaluation of artificial
intelligence techniques in disease diagnosis and prediction,’’ Discover
Artif. Intell., vol. 3, no. 1, p. 5, Jan. 2023.
[7] D. Hassan, H. I. Hussein, and M. M. Hassan, ‘‘Heart disease prediction
based on pre-trained deep neural networks combined with principal
component analysis,’’ Biomed. Signal Process. Control, vol. 79, Jan. 2023,
Art. no. 104019. SOHAIL JABBAR received the M.S. degree in
[8] A. Jain, A. Chandra Sekhara Rao, P. Kumar Jain, and Y.-C. Hu, ‘‘Optimized telecommunication and networks and the Ph.D.
levy flight model for heart disease prediction using CNN framework in big
degree in computer science (wireless commu-
data application,’’ Expert Syst. Appl., vol. 223, Aug. 2023, Art. no. 119859.
nication and networks) from Bahria University,
[9] R. Abbas and N. Gu, ‘‘Improving deep learning-based image super-
Islamabad, in 2009 and 2015, respectively. He was
resolution with residual learning and perceptual loss using SRGAN
model,’’ Soft Comput., vol. 27, no. 21, pp. 1–17, Nov. 2023. a Professor at the Department of Computational
[10] V. Shankar, V. Kumar, U. Devagade, V. Karanth, and K. Rohitaksha, ‘‘Heart Sciences, the Associate Dean of the Faculty of
disease prediction using CNN algorithm,’’ Social Netw. Comput. Sci., Information Technology, and the Director of the
vol. 1, no. 3, p. 170, May 2020. Office of Research, Innovation, and Commercial-
[11] M. Jafari, A. Shoeibi, M. Khodatars, N. Ghassemi, P. Moridian, ization (ORIC), The University of Faisalabad,
R. Alizadehsani, A. Khosravi, S. H. Ling, N. Delfan, Y.-D. Zhang, Pakistan. He served in various academic and managerial positions in
S.-H. Wang, J. M. Gorriz, H. Alinejad-Rokny, and U. R. Acharya, different universities. Previously, he was a Research Associate with the
‘‘Automated diagnosis of cardiovascular diseases from cardiac magnetic CfACS IoT Laboratory, Manchester Metropolitan University, U.K., in 2017;
resonance imaging using deep learning models: A review,’’ Comput. Biol. and a Postdoctoral Fellow with the Network Laboratory, Kyungpook
Med., vol. 160, Jun. 2023, Art. no. 106998. National University, Daegu, South Korea, in 2020. He is currently an
[12] A. Janik, J. Dodd, G. Ifrim, K. Sankaran, and K. Curran, ‘‘Interpretability Associate Professor of computer science with the College of Computer
of a deep learning model in the application of cardiac MRI segmentation and Information Sciences, Imam Mohammad Ibn Saud Islamic University
with an ACDC challenge dataset,’’ Proc. SPIE, vol. 11596, pp. 861–872, (IMSIU), Riyadh, Saudi Arabia. He has been involved in many national
Aug. 2021. and international projects. He is involved in collaborative research with
[13] S. Mohan, C. Thirumalai, and G. Srivastava, ‘‘Effective heart disease renowned research centers and institutes worldwide on various issues in
prediction using hybrid machine learning techniques,’’ IEEE Access, vol. 7, the Internet of Things, data science, and blockchain domains. He has
pp. 81542–81554, 2019.
authored four book chapters and has published over 150 research articles in
[14] D. Bertsimas, L. Mingardi, and B. Stellato, ‘‘Machine learning for real-
prestigious journals and conferences. His research work has been published
time heart disease prediction,’’ IEEE J. Biomed. Health Informat., vol. 25,
in various renowned journals and magazines of IEEE, Springer, Elsevier,
no. 9, pp. 3627–3637, Sep. 2021.
[15] C. Bhatt, P. Patel, T. Ghetia, and P. Mazzeo, ‘‘Effective heart disease
MDPI, Old City Publication, Hindawi, and the conference proceedings
prediction using machine learning techniques,’’ Algorithms, vol. 16, no. 2, organized by IEEE, ACM, and IAENG. He is a TPC member/the Chair/the
p. 88, 2023. Co-Chair of many conferences. He received many awards and honors
[16] D. AbdElminaam, M. Radwan, N. M. Abdelrahman, H. W. Kamal, from the Higher Education Commission of Pakistan, Bahria University,
A. K. A. Elewa, and A. M. Mohamed, ‘‘MLHeartDisPrediction: Heart COMSATS University Islamabad, and Korean Government. He received the
disease prediction using machine learning,’’ J. Comput. Commun., vol. 2, M.S. leading to Ph.D. Scholarship, Magna Cum Laude, the Best Student
no. 1, pp. 50–65, Jan. 2023. Research Awards of the Year, and the BK-21 Plus Post-Doctoral Fellowship.
[17] K. Arumugam, M. Naved, P. Shinde, O. Leiva-Chauca, A. Huaman-Osorio, He also received the Research Productivity Award from COMSATS
and T. Gonzales-Yanac, ‘‘Multiple disease prediction using machine University Islamabad, from 2014 to 2017. In a list representing the top 2
learning algorithms,’’ Mater. Today, Proc., vol. 80, pp. 3682–3685, % of the most-cited scientists in various disciplines released by Stanford
Jun. 2023. University, he is included in the Top 2 % percent of authors based on the
[18] K.-P. Kresoja, M. Unterhuber, R. Wachter, H. Thiele, and P. Lurz, citation impact, from 2019 to 2023. He has been a Reviewer of leading
‘‘A cardiologist’s guide to machine learning in cardiovascular disease journals, including ACM TOSN, JoS, MTAP, AHSWN, and ATECS, and
prognosis prediction,’’ Basic Res. Cardiol., vol. 118, no. 1, p. 10, conferences. He is the Guest Editor of special issues in leading journals
Mar. 2023. of his domain, including Concurrency and Computation Practice and
[19] G. Abdulsalam, S. Meshoul, and H. Shaiba, ‘‘Explainable heart disease Experience (Wiley), Future Generation Computer Systems (Elsevier), Peer-
prediction using ensemble-quantum machine learning approach,’’ Intell.
to-Peer Networking and Applications (Springer), Journal of Information
Autom. Soft Comput., vol. 36, no. 1, pp. 761–779, Sep. 2022.
and Processing System (KIPS), Cyber-Physical System (Taylor & Francis),
[20] S. K. K. Basha, D. Roja, S. S. Priya, L. Dalavi, S. S. Vellela, and
IEEE WIRELESS COMMUNICATIONS (IEEE Communication Society), and IEEE
B. V. Reddy, ‘‘Coronary heart disease prediction and classification using
hybrid machine learning algorithms,’’ in Proc. Int. Conf. Innov. Data Internet of Things Magazine. He is an Associate Editor of IEEE ACCESS and
Commun. Technol. Appl. (ICIDCA), Mar. 2023, pp. 1–7. Telecommunication Systems (Springer).

138326 VOLUME 13, 2025

S. Jabbar et al.: Automated Cardiac Disease Prediction Using Composite GAN and DeepLab Model

UMAR RAZA (Senior Member, IEEE) received MUHAMMAD FARHAN received the Ph.D.
the B.Sc. degree in engineering electronics from degree in computer science, in 2017. He is cur-
De Montfort University, Leicester, the M.Sc. rently an Assistant Professor with the Department
degree in electronics and computer-based systems of Computing Science, COMSATS University
design from the University of Huddersfield, and Islamabad, Sahiwal Campus. He has published
the Ph.D. degree in the application of wireless in high-impact journals and conferences and has
sensor networks (WSN) to the plastics industry actively reviewed for IEEE and other leading
from the Polymer IRC Laboratory, University of publishers. He has contributed to various research
Bradford. He has worked in the industry for seven projects involving federated learning, explainable
years as a Software Engineer and then moved back AI, and bio-inspired methodologies for smart envi-
into academia as a Lecturer in robotics, networking, and computing at ronments. His research interests include multi-agent systems, reinforcement
Staffordshire University. He is currently a Senior Lecturer in networking learning, cyber-physical systems security, and AI-driven optimization.
IoT and smart systems with Manchester Metropolitan University and the
Manager/the Coordinator of the Cisco Network Academy, Department
of Engineering. His research interests include the application of the
Internet of Things (IoT) concept, machine learning (ML), service-oriented
architecture (SOA), WSN, robotics in industrial, smart cities, and agricultural
environments. His current research projects include the IoT and ML for
remote monitoring in domiciliary care settings, the IoT and ML for helping
people with cognitive impairment, wireless sensor networks for early
warning and identification of natural disasters, and smart farming using the
IoT and ML in Pakistan.

MUHAMMAD ASIF HABIB is currently a

Professor of computer science with the College SAQIB SAEED received the B.Sc. degree (Hons.)
of Computer and Information Sciences, Imam in computer science from International Islamic
Mohammad Ibn Saud Islamic University (IMSIU), University Islamabad, Pakistan, in 2001, the MSc.
Riyadh, Saudi Arabia. He has been working degree in software technology from Stuttgart
professionally, since 2004. He also worked in the University of Applied Sciences, Germany, in 2003,
industry in various network-related posts. Then, and the Ph.D. degree in information systems from
he joined teaching and research. He has experience the University of Siegen, Germany, in 2012. He is
working as a Faculty Member and a Researcher at a Certified Software Quality Engineer at American
various international universities. He also served Society of Quality. He is currently an Associate
at the Department of Computer Science, National Textile University, Professor with the Department of Computer Infor-
Faisalabad, Pakistan, as the Chairman/an Associate Professor, and the mation Systems, Imam Abdulrahman Bin Faisal University, Dammam,
Director IT of National Textile University, Faisalabad, Pakistan. He has Saudi Arabia. His research interests include digital transformation, user-
been teaching network-relevant and other core computing courses at Ph.D., centric security, and human-centered information systems. He is a member
master’s, and B.S. levels. His research interests include, but are not limited to, of the advisory boards of several international journals, in addition to being
information and network security, digital image processing, and the Internet a guest editor of several special issues.
of Things.

VOLUME 13, 2025 138327