Arrhythmias are conditions characterized by irregular heart rhythms, varying from benign to potentially life-threatening (Antzelevitch and Burashnikov, 2011; Çınar and Tuncer, 2021; Attia et al., 2019). These anomalies arise from electrical disturbances that disrupt the normal rhythm of the heart, potentially leading to severe health complications such as stroke, heart failure, or sudden cardiac arrest. Over time, advances in medical technology and understanding of cardiovascular diseases have highlighted the critical need for accurate detection and timely treatment of arrhythmias, underscoring their significance in public health (Ferreira and Zanesco, 2016; Organization, 2020; Mc Namara et al., 2019).

The study focuses on five primary types of arrhythmias: Atrial Fibrillation (AFib), Supraventricular Tachycardia (SVT), Bradycardia, Premature Ventricular Contractions (PVCs), and Atrial Flutter (Luo et al., 2017). Each type presents unique challenges in detection and management, largely due to their distinct manifestations and impacts on cardiac function (Saini et al., 2013). For example, AFib is known for the rapid and irregular beating of the atrial chambers, while Bradycardia is characterized by abnormally slow heart rates (Saini et al., 2015; Islam et al., 2022; Behl et al., 2020; Councils, 2015). Accurate classification of these conditions is crucial for implementing appropriate therapeutic strategies (Nurmaini et al., 2020; Tenepalli and Navamani, 2024) proposed the ERFEX framework using XGBoost, RFE, and SHAP, achieving 98.23% accuracy. However, it still faces challenges such as overfitting and limited generalizability to diverse datasets.

Kumar et al. (2022) developed a deep learning model for detecting heart arrhythmias using data from IoT-connected ECG devices. Based on a modified Grey Wolf Optimizer (GWO) algorithm, this model achieved 95% accuracy. However, it faces challenges with computational cost and requires more sophisticated feature extraction techniques for improved performance (Singh and Mahapatra, 2024) investigates the use of Recurrent Convolutional Neural Networks (RCNN) optimized by a GWO algorithm for classifying heart arrhythmias. This model achieved an accuracy of 98% on standard datasets, surpassing traditional machine learning methods. However, it presents challenges, including high computational demands, the need for large amounts of labeled data, the risk of overfitting, and potential difficulties in practical implementation (Sinnoor and Janardhan, 2024). Proposed an LSTM-based arrhythmia classification system with optimized hyperparameters using the Runge-Kutta (RUN) optimizer, achieving 99. 87% precision by combining the preprocessing of the ECG signal (Liu et al., 2024) developed a lightweight deep learning model to detect heart arrhythmias using 12-lead ECG data. This model, optimized through techniques such as compact convolutions and identifying critical accuracy segments, achieves high accuracy (95.532%) while requiring fewer parameters. However, oversimplifying the model may lead to losing important information, potentially impacting its performance.

In recent years, optimization algorithms and machine learning models have become pivotal in enhancing the diagnosis and classification of medical conditions, including arrhythmias (Kohli and Verma, 2011). These technologies leverage vast amounts of data to uncover patterns undetectable by human experts, offering significant improvements in diagnostic accuracy and efficiency (Brenyo and Aktas, 2014; Karpagachelvi et al., 2010; Hagiwara et al., 2018). Integrating optimization techniques helps to fine-tune the model parameters and select relevant features, which is crucial for building robust predictive models (Hammad et al., 2022).

Electrocardiogram (ECG), a straightforward, inconspicuous, harmless method that measures the heart’s electrical activity, is a frequently used test to identify arrhythmias (Kıymaç and Kaya, 2023). Manual ECG interpretation is challenging due to waveform variability and subjective human assessment, prompting the development of automated machine-learning-based Cardio Vascular Disease (CVD) systems. Midani et al. (2023) developed a hybrid Deep Neural Network (DNN) model that effectively identifies arrhythmias in ECG data. It combines recurrent and contextual DNNs, achieving high accuracy, but may have computational limitations for real-time applications in portable devices or systems with restricted processing power. Kıymaç and Kaya (2023) proposed a memory-enhanced artificial hummingbird algorithm to optimize deep learning models for classifying cardiac arrhythmias from ECG data. It is fully automated, efficient, and robust but requires significant computational resources and large datasets for training. The 1D Convolutional Neural Network (CNN) method developed by Ahmed et al. (2023) has performed well in extracting important characteristics from ECG data. However, it had a compromised sensitivity, resulting in a greater occurrence of false negatives, a crucial issue in medical diagnosis. Furthermore, the approaches that use traditional Convolutional Neural Networks (CNNs) for image analysis, as introduced by Liu et al. (2022), showed effectiveness in extracting features. However, they faced challenges in terms of sensitivity, which could result in omitting important positive information necessary for clinical decision-making. Ojha et al. (2022) proposed a novel approach that combines autoencoders and SVM for feature extraction and classification. This method has demonstrated remarkable accuracy. However, it necessitates significant processing resources and meticulous tuning to prevent overfitting, particularly when dealing with smaller or imbalanced datasets.

The diverse range of approaches employed for arrhythmia detection, each with its advantages and limitations. While ongoing advancements in these methods show promise for improving heart health diagnoses, challenges such as computational requirements, mathematical complexity, and applicability to various datasets remain significant obstacles. Previous studies often encounter challenges such as limited accuracy due to inefficient feature optimization and the inability to capture ECG signal complexities fully. These constraints underscore the need for more robust and efficient models, creating opportunities for innovative solutions like the proposed EGOLF-Net model. This model aims to address these deficiencies by combining advanced optimization techniques with deep learning. The unique methodologies emphasized, such as improvements achieved through meta-heuristic algorithms and attention-based models, demonstrate the ongoing development in this field, emphasizing the need for further research and refinement to optimize patient care and clinical outcomes.

The main contributions of this research work are as follows:

• Proposed a novel fusion EGOLF-Net model that combines Enhanced Grey Wolf Optimization (EGOLF) with Long Short-Term Memory (LSTM) networks to predict arrhythmia.

The structure of this paper is organized as follows. The Related Works section discusses the existing methodologies in arrhythmia detection. The methodology section discusses the proposed methodology, introducing the EGOLF-Net model and describing its unique components and operational framework. The Result Analysis section presents the results of applying the EGOLF-Net model on ECG. Finally, the Conclusion and Future Scope are discussed.

In recent years, advancements in deep learning have significantly improved arrhythmia diagnosis. Researchers have explored various methods combining neural networks with advanced data processing to enhance accuracy. This section provides an overview of these methods, highlighting their strategies and performance in ECG data analysis.

Kumar et al. (2023) explored combining deep learning with fuzzy clustering to improve medical image diagnosis. Fuzzy clustering helped handle uncertainties in medical data, enhancing the model’s performance. However, using fuzzy logic can make the model less understandable due to complex rules and membership functions. To analyze Electrocardiogram (ECG) signals, Daydulo et al. (2023) used a time-frequency approach with deep learning models (ResNet 50 and AlexNet). Their approach achieved 99.2% overall accuracy, with high sensitivity and specificity. This method effectively captured the temporal and frequency aspects of ECG signals for accurate heart disease diagnosis. However, combining multiple deep learning models can make the model more computationally demanding, limiting its applicability in resource-constrained settings.

Tripathi et al. (2022) proposed a deep learning-based method for classifying cardiac arrhythmias using ECG signals, achieving high accuracy with reduced hardware needs. However, it requires large datasets and prolonged training. Soman and Sarath (2024) proposed a Chameleon-Sparrow Search Algorithm-based Deep Convolutional Neural Network (CsSA-based Deep CNN) for effectively classifying arrhythmia in ECG signals using wavelet-based preprocessing and feature extraction. However, further improvements are necessary to enhance its accuracy and address computational complexity concerns, particularly for real-time clinical applications. Cai et al. (2020) proposed Multi-ECGNet, a deep learning-based approach for multi-label ECG classification, effectively identifying multiple heart diseases simultaneously with high accuracy, surpassing human experts. However, its computational complexity and potential for misclassification may limit its real-world application. Ojha et al. (2022) extracted and classified features using an autoencoder and a Support Vector Machine (SVM). This combination made effective reduction of dimensionality and reliable classification possible. Autoencoder training can be computationally demanding, though, and it might need fine-tuning to prevent overfitting especially when working with smaller or imbalanced datasets. Haq et al. (2023) proposed the Reseek-Arrhythmia model. It effectively detects and classifies heart arrhythmias with high accuracy, but it may struggle with handling noisy data or distinguishing subtle arrhythmia patterns.

Mandala et al. (2024) proposed a new ensemble learning approach for arrhythmia detection using multi-lead ECG data, which achieves high sensitivity, specificity, and accuracy for detecting various arrhythmia classes. Jha (2024) provided a comprehensive review of existing machine-learning and deep-learning-based techniques for arrhythmia detection using ECG signals, highlighting their strengths, limitations, and potential areas for improvement. Begum and Singh (2024) proposed a CNN-based continual normalization classifier for arrhythmia detection, achieving 99.2% accuracy, outperforming existing methods. Glaser et al. (2024) systematically reviewed ML algorithms for predicting and detecting New-Onset Atrial Fibrillation (NOAF) in ICU patients, finding CatBoost and SVM as promising methods and highlighting their potential to improve clinical decision-making. Sposato et al. (2024) presented the management of atrial fibrillation (AF) in ischemic stroke/TIA patients, detailing the risk assessment and monitoring strategies based on different AF detection methods.

Yu et al. (2024) provided a comprehensive review of the pathophysiology and etiology of arrhythmia in COVID-19 patients, offering valuable insights for understanding the disease. Still, it may not detail the specific mechanisms underlying arrhythmia development. An automatic CNN arrhythmia classifier improved by a memory-enhanced artificial hummingbird algorithm was proposed by Kıymaç and Kaya (2023). Their methodology utilizes sophisticated algorithms to enhance the accuracy of classification. Nevertheless, the efficacy of the study may be limited by the dependence on memory-enhancement methods, which may not exhibit strong generalization capabilities across various datasets. Admass and Bogale (2024) Proposed a metaheuristic enhancement, combining optimal weighted feature integration with an attention-based hybrid deep learning model to categorize ECG signals. The objective of this approach is to improve the accuracy of categorization by combining different features and attention methodologies. Nevertheless, the intricacy of the model may result in higher computational requirements and possible constraints in handling various ECG signal specifications.

The existing literature review comprehensively explores contemporary methods for arrhythmia detection, focusing on various techniques and their effectiveness in utilizing electrocardiogram (ECG) data. Recurrent Deep Neural Networks (DNN), convolutional models like ResNet 50 and AlexNet Daydulo et al. (2023), and autoencoders integrated with Support Vector Machines (SVM) Ojha et al. (2022) have been widely adopted to enhance arrhythmia identification accuracy. While these approaches demonstrate excellent diagnostic performance, they often face computational complexity, interpretability issues, and resource requirements. For instance, the hybrid DNN model proposed by Midani et al. (2023) although achieving remarkable precision through temporal dynamics and contextual learning, is constrained by its high computational demands, limiting its suitability for real-time applications on portable devices. Similarly, Kumar et al. (2023) proposed a model that integrates deep learning with fuzzy clustering, improving diagnostic accuracy, but introduced interpretability challenges due to the complex nature of fuzzy logic rules and membership functions.

Recent artificial intelligence (AI) breakthroughs have markedly enhanced electrocardiogram (ECG)-based cardiac diagnostics, especially in arrhythmia diagnosis. Multiple studies have used machine learning and deep learning algorithms to automate ECG interpretation, exhibiting elevated accuracy levels. Muzammil et al. (2024) examined the efficacy of AI-enhanced ECG analysis in clinical diagnoses, highlighting that deep learning, especially convolutional neural networks (CNNs), may surpass conventional interpretation techniques by identifying patterns undetectable by human observers. Nevertheless, they emphasized significant limitations like dataset bias and interpretability challenges. Telmoud et al. (2024) assessed traditional machine learning models, including Decision Trees, Support Vector Machines, and Random Forests, to categorize ECG data. Their trials with the MIT-BIH dataset achieved accuracies ranging from 97% to 99%, especially with ensemble models. Nonetheless, these approaches often rely on manually crafted features and frequently fail to capture the temporal dynamics of ECG data, essential for detecting arrhythmic patterns.

Notwithstanding the substantial advancements in ECG-based arrhythmia classification, two significant deficiencies persist: the absence of cohesive optimization for feature selection and temporal modeling. Current models often depend on comprehensive feature sets lacking optimum selection or disregard temporal dynamics by concentrating on static ECG segments or images—insufficient resilience and adaptability. Numerous models are tailored to particular datasets, have inadequate noise filtering or generalization abilities, and often do not sustain high accuracy when used in clinical environments. Hence, we proposed EGOLF-Net (Enhanced Gray Wolf Optimization with LSTM Fusion Network), which addresses these gaps by introducing a hybrid framework that integrates Enhanced Gray Wolf Optimization (EGWO) for optimal feature selection, effectively reducing noise and redundancy in high-dimensional ECG data. LSTM layers to capture sequential dependencies inherent in ECG signals, enabling robust temporal pattern recognition, a fully integrated pipeline combining preprocessing, feature engineering, and learning to boost generalizability and diagnostic reliability.

The methodology for detecting and classifying arrhythmias using Electrocardiogram (ECG) data begins with collecting ECG signals from individuals showcasing a variety of cardiac conditions, encompassing both normal and abnormal rhythms as shown in Figure 1. These signals undergo preprocessing steps to refine data quality, including noise filtering, signal amplitude normalization, and segmentation to isolate specific features within the ECG trace. The signal fusion and optimization phase in the EGOLF-Net system combines multiple segments or different aspects of the ECG signals into a comprehensive signal. The optimized signal undergoes detailed feature extraction, involving statistical measures and frequency domain analysis, to capture the essential characteristics of the ECG data for accurate classification. Linear Discriminant Analysis (LDA) is applied for dimensionality reduction, emphasizing the most informative features for subsequent classification. This technique helps reduce the model’s computational complexity while preserving its discriminative power. Features extracted from the ECG signals serve as inputs to the proposed model EGOLF-Net model. After following the mechanism of EGWO, the selected features are fed to LSTM. The processed data is fed through LSTM layers, which learn complex patterns in the ECG signals. Finally, a SoftMax layer classifies the signals into different Arrhythmia types based on the learned patterns. The EGOLF-Net architecture effectively combines signal fusion, deep learning, and optimization techniques to enhance classification accuracy and overcome challenges associated with ECG data analysis.

The experiments were conducted in a MATLAB R2023a environment using the system shown in Table 1. All ECG signal preprocessing, feature extraction, EGWO optimization, and LSTM training were performed on the same platform. To enhance the performance of the EGOLF-Net model in categorizing arrhythmias, the training parameters are carefully set as shown in Table 2. The learning rate is configured at 0.01, enabling the model to significantly adjust weights during each iteration while maintaining a balance between convergence speed and stability. To achieve a suitable level of learning without overfitting, the model is trained for 30 epochs with a batch size of 100, effectively balancing training speed and memory efficiency. An L2 regularization factor of 0.001 punishes big weights to improve the model’s generalization capabilities further. A total of 50 LSTM units are configured to ensure the network can accurately capture the temporal relationships in ECG data.

The normal initialization is employed for the weights, specifically designed for ReLU activation functions, to ensure a consistent signal variance across the network. To avoid overfitting, the Adam optimizer is selected for its adjustable learning rate, effective management of sparse gradients, and a dropout rate of 0.2. The feature selection process is guided by Enhanced Gray Wolf Optimization (EGWO), which iteratively improves the input features to guarantee that the most pertinent data is included in the classification procedure. By meticulously choosing these parameters, the EGOLF-Net model’s resilience and effectiveness in real-time arrhythmia detection are significantly improved.

The diagnostic system’s performance is assessed using accuracy, sensitivity, specificity, predictive values, and likelihood ratios, as defined in Equations 18–21, ensuring its reliability and efficacy for clinical application.

Accuracy A c c u r a c y = T P + T N T P + T N + F P + F N (18)

Sensitivity (True Positive Rate): S e n s i t i v i t y = T P T P + F N (19)

Specificity (True Negative Rate): S p e c i f i c i t y = T N T N + F P (20)

Area Under the ROC Curve (AUC): A U C = ∫ 0 1 T P R d ′ d F P R (21) where TPR is depicted as the True Positive Rate whereas FPR is the False Positive Rate as a function of threshold t.

Figure 4 presents the fused ECG signals. This fusion process integrates signals from multiple subjects into a single, comprehensive dataset, shown here with overlaid signals that enhance the representation for analysis. Figure 5 shows the optimized fused signals from Figure 4, where optimization processes have been applied to enhance diagnostic features, reduce noise, and suppress irrelevant information. The result is a cleaner, more diagnostic-friendly signal, highlighting key features crucial for effective arrhythmia detection.

Figure 6 illustrates the clustering of optimized ECG signals, with various symbols and colors denoting different clusters. These clusters are formed based on characteristics identified in the ECG signals that correspond to distinct arrhythmic conditions or signal patterns.

Table 3 lists the dynamic features extracted from the ECG signals of ten subjects (S-1 to S-10). It includes maximum and minimum values, mean, median, and RMS (Root Mean Square). These features capture the essential statistical properties of the ECG signals which are critical for training the EGOLF-Net in pattern recognition associated with normal and abnormal cardiac functions.

Figure 7 is a confusion matrix that categorizes the classification performance of the EGOLF-Net model across five types of arrhythmias. It quantifies the model’s accuracy in predicting each arrhythmia type, with true positives highlighted on the diagonal and classification errors in the off-diagonal cells. Figure 8 specifically demonstrates a Supraventricular Tachycardia (SVT) event detected in the ECG data for Subject 2. It highlights how the model isolates and identifies specific types of arrhythmias based on the learned patterns from the training data.

The convergence curves shown in Figure 9 illustrate the performance of the Enhanced Gray Wolf Optimization (EGWO) and Particle Swarm Optimization (PSO) over 100 iterations. The curves show how the fitness values improve as the number of iterations increases. EGWO demonstrates a more rapid convergence than PSO, highlighting its effectiveness in optimizing feature selection in the EGOLF-Net model. Figure 10 charts the training and validation loss over 28 epochs, marking the epochs where the validation performance was optimal. This point, where the validation loss is minimized to approximately 0.19854 at epoch 22, indicates the most effective balance between learning and generalization.

Figure 11 tracks the gradient magnitude and validation checks across the training epochs. The gradient plot shows the model’s learning rate, with a notable gradient value of about 0.064519 at epoch 28, suggesting effective learning steps without instability. The validation check graph illustrates the model’s performance improvement stops after six validation checks without further gain, indicating a possible point for halting training to prevent overfitting.

Figure 12 shows the Error Histogram with 20 Bins. It depicts the distribution of prediction errors for both training and validation phases in the EGOLF-Net model. Each bin on the x-axis represents a range of error values between the predicted and actual outputs, with the y-axis displaying the number of instances in each bin. The zero-error line represents perfect prediction accuracy. The histogram shows a concentration of instances near the zero error, particularly for the validation set, indicating a good model fit. However, there are also instances with higher errors, highlighting areas where the model may need further refinement or indicating the presence of more complex arrhythmic patterns.

Figure 13 shows the Receiver Operating Characteristic (ROC) curves from several datasets for five kinds of arrhythmias. The one-vs-rest ROC curves for Classes 1 through 5 (Class 1: Atrial Fibrillation, Class 2: Supraventricular Tachycardia, Class 3: Bradycardia, Class 4: Premature Ventricular Contractions, Class 5: Atrial Flutter) are displayed in each subplot in the following datasets: (a) Training, (b) Validation, (c) Testing, and (d) Combined. Strong generalization of the suggested EGOLF-Net model is indicated by the curves’ consistent performance throughout all stages, with AUCs well above the random guessing baseline (diagonal).

The baseline CNN-SVM model, which combines Convolutional Neural Networks (CNN) for feature extraction and Support Vector Machines (SVM) for classification, serves as the comparative model. This model includes a common strategy for categorizing arrhythmias, however, it lacks the complex optimization and temporal pattern recognition skills in EGOLF-net.

The statistical significance analysis in Table 4 demonstrates that the EGOLF-Net model invariably surpasses the baseline CNN-SVM model in all assessed measures, such as accuracy, sensitivity, specificity, precision, F1-score, and ROC-AUC. Since the p-values for all comparisons are below 0.01, the performance enhancements of EGOLF-Net are statistically significant. This strong performance highlights the efficacy of combining Enhanced Gray Wolf Optimization with LSTM networks, confirming the model’s superior diagnostic abilities in detecting arrhythmias.

The present work involved developing and validating the EGOLF-Net model as a reliable instrument for the real-time identification of arrhythmias using electrocardiogram results. The proposed model combines Enhanced Gray Wolf Optimization (EGWO) with Long Short-Term Memory (LSTM) networks, utilizing sophisticated optimization and deep learning methods to improve the precision and dependability of diagnostics. EGOLF-Net tackles significant obstacles in interpreting ECG signals, including noise, high dimensionality, and the intricacy of arrhythmic patterns through successfully integrating feature optimization and temporal pattern recognition. The model exhibited outstanding performance, attaining an accuracy of 99.61%, sensitivity of 99.4%, and specificity of 99.64%, well surpassing the performance of current approaches in the literature. The demonstrated results highlight the potential of EGOLF-Net as a very efficient and dependable instrument for categorizing arrhythmias, thereby providing substantial enhancements in clinical diagnostics. This work emphasizes the benefits of combining metaheuristic optimization algorithms such as EGWO with deep learning models, offering a comprehensive methodology that can be tailored to different medical diagnostic applications.