GAN-SkipNet: A Solution for Data Imbalance in Cardiac Arrhythmia Detection Using Electrocardiogram Signals from a Benchmark Dataset

Abstract

Electrocardiography (ECG) plays a pivotal role in monitoring cardiac health, yet the manual analysis of ECG signals is challenging due to the complex task of identifying and categorizing various waveforms and morphologies within the data. Additionally, ECG datasets often suffer from a significant class imbalance issue, which can lead to inaccuracies in detecting minority class samples. To address these challenges and enhance the effectiveness and efficiency of cardiac arrhythmia detection from imbalanced ECG datasets, this study proposes a novel approach. This research leverages the MIT-BIH arrhythmia dataset, encompassing a total of 109,446 ECG beats distributed across five classes following the Association for the Advancement of Medical Instrumentation (AAMI) standard. Given the dataset’s inherent class imbalance, a 1D generative adversarial network (GAN) model is introduced, incorporating the Bi-LSTM model to synthetically generate the two minority signal classes, which represent a mere 0.73% fusion (F) and 2.54% supraventricular (S) of the data. The generated signals are rigorously evaluated for similarity to real ECG data using three key metrics: mean squared error (MSE), structural similarity index (SSIM), and Pearson correlation coefficient (r). In addition to addressing data imbalance, the work presents three deep learning models tailored for ECG classification: SkipCNN (a convolutional neural network with skip connections), SkipCNN+LSTM, and SkipCNN+LSTM+Attention mechanisms. To further enhance efficiency and accuracy, the test dataset is rigorously assessed using an ensemble model, which consistently outperforms the individual models. The performance evaluation employs standard metrics such as precision, recall, and F1-score, along with their average, macro average, and weighted average counterparts. Notably, the SkipCNN+LSTM model emerges as the most promising, achieving remarkable precision, recall, and F1-scores of 99.3%, which were further elevated to an impressive 99.60% through ensemble techniques. Consequently, with this innovative combination of data balancing techniques, the GAN-SkipNet model not only resolves the challenges posed by imbalanced data but also provides a robust and reliable solution for cardiac arrhythmia detection. This model stands poised for clinical applications, offering the potential to be deployed in hospitals for real-time cardiac arrhythmia detection, thereby benefiting patients and healthcare practitioners alike.

1. Introduction

Cardiac arrhythmias are irregular heart rhythms that can be caused by various factors. For instance, coronary artery disease (CAD) occurs when blood vessels supplying the heart become narrowed or blocked, disrupting the heart’s electrical signaling system [1,2]. Heart failure, resulting from a weakened heart muscle, can also lead to arrhythmias due to changes in the heart’s structure and electrical properties [3,4]. Valvular heart disease disrupts the heart’s electrical signals by causing turbulence in blood flow [5]. Additionally, arrhythmias can occur independently, with genetic factors making some individuals more susceptible to irregular heart rhythms. Imbalances in essential electrolytes, lifestyle choices like excessive caffeine or alcohol consumption, smoking, and certain medications can all contribute to arrhythmias [6,7]. A wide array of arrhythmias can affect the human heart, often tied to various underlying conditions. These arrhythmias encompass fast and slow heart rhythms and regular and irregular heartbeat patterns and can be either persistent or paroxysmal, as well as classified as malignant or non-malignant. Moreover, diverse therapeutic options are available to manage these conditions effectively. When categorizing arrhythmias based on heart rate, they can be divided into two primary types: fast arrhythmias, known as tachyarrhythmias, and slow arrhythmias, or bradyarrhythmias. Tachyarrhythmias, like atrial fibrillation, atrial flutter, and ventricular tachycardia, lead to a rapid and often irregular heartbeat. Conversely, bradyarrhythmias, including sinus bradycardia and heart blocks, result in a heart rate that is slower than the normal pace [8]. Arrhythmias can also be classified based on the regularity of the heartbeat. Those with a consistent and predictable heartbeat pattern, such as atrial flutter and supraventricular tachycardia (SVT), are termed regular arrhythmias. In contrast, irregular arrhythmias, like atrial fibrillation (AFib) [8,9,10], ventricular fibrillation (VFib), and premature contractions, exhibit erratic and unpredictable heartbeat patterns. Moreover, the persistence of arrhythmias plays a vital role in classification. Persistent arrhythmias, like persistent atrial fibrillation (AFib) or atrial flutter, endure for an extended period without self-termination, necessitating active management and treatment. On the other hand, paroxysmal arrhythmias, such as paroxysmal supraventricular tachycardia (PSVT), paroxysmal AFib, and paroxysmal ventricular tachycardia (VT), involve sudden and intermittent episodes of irregular heart rhythms that spontaneously start and stop [11,12]. Lastly, the severity of arrhythmias distinguishes them as malignant or non-malignant. Malignant arrhythmias, including ventricular tachycardia (VT) and ventricular fibrillation (VFib), exhibit fast and chaotic rhythms originating in the ventricles and can potentially lead to sudden cardiac arrest. Conversely, non-malignant arrhythmias, such as atrial fibrillation (AFib), bradycardia (slow heart rate), and premature contractions (PACs and PVCs), are generally less immediately life-threatening, although they can still pose health risks and complications [13,14].

The diagnosis of arrhythmias, which are irregular heart rhythms, is primarily accomplished through the use of an electrocardiogram (ECG). This diagnostic tool is indispensable for understanding and characterizing the heart’s electrical activity [15,16]. During an ECG, electrodes are placed on specific locations of the chest, arms, and legs to record the heart’s electrical signals. The resulting ECG waveform illustrates various phases of the cardiac cycle, encompassing atrial and ventricular depolarization and repolarization. These waveforms serve as vital indicators for diagnosing arrhythmias, as deviations in their shape and timing can precisely identify the specific type of arrhythmia and any associated underlying conditions. Morphology in an ECG refers to the shape, duration, and characteristics of the recorded waveforms [17,18]. In a normal ECG, waveforms follow a specific and well-defined pattern, including the P wave representing atrial depolarization, the QRS complex indicating ventricular depolarization, the T wave signifying ventricular repolarization, and sometimes the U wave [19,20]. Abnormal morphology may present as alterations in the shape, duration, or regularity of these waveforms, which can reveal the presence of arrhythmias or underlying heart conditions [21,22]. Arrhythmias are diagnosed through various ECG monitoring methods, each serving a specific purpose. A standard ECG, typically conducted in clinical settings, briefly records the heart’s electrical activity. Short-term monitoring utilizes portable devices like Holter monitors and event recorders, which are crucial for capturing intermittent arrhythmias that may not appear during standard ECGs. Long-term monitoring, on the other hand, involves the use of wearable or implantable devices, ensuring the detection of sporadic arrhythmias over extended periods, often spanning weeks or months. These diverse monitoring methods are essential for diagnosing a wide range of arrhythmias, enabling accurate assessments and effective treatment [23,24,25]. ECG monitoring devices come in various forms, with standard ECGs and short-term monitoring taking place in clinical offices. Wearable ECG monitors, including options like chest straps and wristwatches, provide real-time monitoring in daily life, enhancing the detection of arrhythmias that may not manifest during clinical visits. Implantable ECG monitors, surgically positioned beneath the skin, offer continuous recording over several years.

The ECG signal remains the primary diagnostic modality for arrhythmia diagnosis. However, interpreting ECGs, especially for extended monitoring, demands considerable experience and time. It can be a complex task, as arrhythmias come in various forms and may not always present clear patterns to the naked eye. Consequently, the development of computer-aided diagnostic systems that can enhance the accuracy and efficiency of ECG interpretation is an active and critical area of research. One significant limitation in training diagnostic systems for arrhythmia detection is the issue of imbalanced data. Imbalanced datasets occur when one class of arrhythmia is significantly more prevalent than others, leading to an inherent bias in the model’s learning [26,27]. In real-world ECG datasets, certain types of arrhythmias are rare, while others are common. The limitations of training datasets with imbalanced data can result in machine learning models that are biased toward the majority class, leading to reduced sensitivity, overfitting, and misleading evaluation metrics. These models may perform poorly in detecting rare arrhythmias, making them less applicable in real-world clinical settings. To address these limitations, various techniques used by the researchers, such as oversampling, undersampling, and the use of specialized evaluation metrics, are employed to improve the accuracy and reliability of arrhythmia detection systems [28,29,30].

Indeed, the challenges posed by imbalanced data have spurred a significant push toward the development of systems designed to overcome this issue. These systems are driven by the goal of ensuring the accurate and comprehensive detection of arrhythmias, irrespective of the rarity of certain types in the training data. To tackle data imbalance, an innovative solution has been introduced in the form of a generative adversarial network (GAN). This GAN is capable of generating a synthetic dataset for the minority class, resolving the imbalance problem. The GAN-generated data, which include sufficient ECG beats, mitigate the imbalance issue, enabling deep learning models to be trained effectively on a balanced dataset. The result is improved efficiency in arrhythmia detection, addressing a significant challenge in this critical area of healthcare.

1.1. Literature Review

We performed a nonsystematic literature review of machine learning methods for arrhythmia detection, focusing on strategies to mitigate data imbalance induced by minority classes of uncommon ECG beats in extant training ECG databases, the established and popular being the Massachusetts Institute of Technology–Beth Israel Hospital (MIT-BIH) arrhythmia database [31,32,33]. Our findings are summarized in

Table 1. Comparative review of data imbalance techniques utilizing ML/DL models for arrhythmia detection.

Study	Model	Dataset	Findings	Limitations
Rajesh and Dhuli, 2018 [35]	AdaBoost	MIT-BIH arrhythmia database	Improved complete empirical mode decomposition-based model for classifying ECG beats. Higher-order statistics and sample entropy computed from intrinsic mode functions. Pre-processing balances heartbeat class distribution.	Limited ECG beat sample size. No comparison with traditional machine learning classifiers. Good classifier performance without feature extraction raises interpretability concerns.
Jiang et al., 2019 [28]	MMNNS	MIT-BIH arrhythmia database, EDB, European ST-T Database, European ST-T Database	MMNNS system addresses ECG imbalance. Four submodules with combined resampling, features, and algorithms. Denoising autoencoders and CNN are used to extract features, followed by softmax regression. Attains accuracy of 97.3% for SVEB and 98.8% for VEB.	Adopts Association for the Advancement of Medical Instrumentation normal (“N”), supraventricular (“S”), ventricular (“V”), and fusion (“F”) ECG beat classes. Classification is limited to “S” and “V” classes in case of inter-patient scheme. Model validation on two additional datasets; morphology variation unaddressed. Generalization to other arrhythmias and data sources not investigated.
Pandey and Janghel, 2019 [30]	CNN	MIT-BIH arrhythmia database	End-to-end CNN structure eliminates the need for ECG QRS segmentation. SMOTE addresses ECG class imbalance. Attains high classification performance.	Oversampling all 87,542 ECG beats risks overfitting. With significant oversampling of total training data, imbalance concerns. Impact of oversampling on model robustness and generalizability is unclear.
Gao et al., 2019 [36]	LSTM, focal loss	MIT-BIH arrhythmia database	Addresses data imbalance: LSTM disentangles timing features, and focal loss handles category imbalance by down weighting. Attains reliable ECG beat classification with 99.26% accuracy.	Validation is based only on the MIT-BIH database. Real-world applicability needs further investigation. Computational and resource requirements pose limitations to practical implementation.
Shaker et al., 2020 [37]	Two-stage deep CNN	MIT-BIH arrhythmia database	GAN generates heartbeats, outperforming standard data augmentation and leading to improved performance. Deep learning obviates hand engineering, enhancing accuracy by over 98%.	Analysis covers 15 ECG beat classes, some of which are uncommon and can be challenging to detect. Validation is solely based on the MIT-BIH arrhythmia dataset.
Petmezas et al., 2021 [38]	CNN, LSTM	MIT-BIH AF database	Focal loss addresses training data imbalance. Hybrid neural model attained 97.87% sensitivity and 99.29% specificity.	Generalization to other datasets and patient populations needs further validation. Future research is needed for other arrhythmia types. Hardware requirements and real-time processing pose limitations to practical implementation.
Zhou et al., 2021 [39]	ACE-GAN	MIT-BIH arrhythmia database	Augments data by creating varied coupling matrix inputs. Attains sensitivity of 87% for supraventricular beats and 93% for ventricular beats; improves F1 score for supraventricular beats by up to 10%.	Performance may vary with different dataset characteristics. Real-world applicability needs further investigation. Computational demands of GAN-based method pose limitations to practical implementation.
Rai et al., 2022 [40]	Ensemble	MIT-BIH, Physikalisch-Technische Bundesanstalt databases	New sequential ensemble technique. SMOTE+Tomek hybrid resampling addresses ECG data imbalance. Attains 99.02% accuracy on the balanced dataset. Improves minority class accuracy by 20%.	Dataset-specific findings limit generalization. Hybrid model performance may vary across datasets. Complex arrhythmia types may impact performance.
Fan et al., 2022 [41]	Majority voting	MIT-BIH arrhythmia database	Combined active training subset selection and modified broad learning system address ECG class imbalance. Iterative and dynamic training subset selection improves accuracy. Attains superior beat classification performance over standard approaches.	Focuses on beat classification within the MIT-BIH arrhythmia database. Generalization to different datasets and real-world applicability need further investigation. Computational complexity and requirements pose limitations to practical implementation.
Ma et al., 2022 [42]	CBAM-ResNet	MIT-BIH arrhythmia database	Gramian angular summation field image transformation is used to represent ECG features. CWGAN-GP addresses imbalanced data, aiding in classification performance. Attains high performance on MIT-BIH arrhythmia database.	Limited to MIT-BIH arrhythmia database. Real-world applicability needs further investigation. Computational requirements pose limitations to practical implementation.
Qin et al., 2022 [43]	Squeeze-and-excitation ResNet1D	MIT-BIH arrhythmia database	WGAN-GP generates a balanced dataset. Attains 95.80% precision, 96.75% recall, and 96.27% F1 score; outperforming VGGNet, DenseNet, and CNN+bidirectional LSTM.	Computational demands of WGAN-GP and squeeze-and-excitation ResNet1D pose limitations to practical implementation. Susceptibility to noise in real-world ECG signals may degrade accuracy. Fine-tuning for clinical applications may delay deployment.
Asadi et al., 2023 [11]	CNN	PhysioNet paroxysmal AF prediction challenge	Neural architecture search customizes CNN to classify paroxysmal AF. GAN generates certified synthetic paroxysmal AF ECGs to address the class imbalance. Attains 99.0% accuracy.	Limited to two-class classification of paroxysmal AF versus no AF. No multi-class capability for detecting other arrhythmia. Generated synthetic ECG signals are not shown or compared.
Qin et al., 2023 [34]	ECG anomaly detection using GAN	MIT-BIH arrhythmia database	One-class classification GAN with bidirectional LSTM and mini-batch discrimination, which generates ECG samples to match healthy data distribution to facilitate anomaly detection. Attains 95.5% accuracy and 95.9% area under the curve.	Data imbalance is not prioritized. Model performance may vary across datasets. Real-world applicability needs further investigation. Computational and resource requirements pose limitations to practical implementation.

ACE-GAN, generative adversarial network with auxiliary classifier for electrocardiogram; AF, atrial fibrillation; CBAM, convolutional block attention modules; CNN, convolutional neural network; CWGAN-GP, conditional Wasserstein generative adversarial network with gradient penalty; GAN, generative adversarial network; LSTM, long short-term memory; MBLS, modified broad learning system; MIT-BIH, Massachusetts Institute of Technology–Beth Israel Hospital; MMNNS, multilayer multiset neuronal networks; ResNet, residual neural network; SMOTE, synthetic minority oversampling technique; SVEB, supraventricular ectopic beat; VEB, ventricular ectopic beats.

. Most models in this review involved a multi-class classification of different morphologies of ECG beats, except for two studies. Using the PhysioNet paroxysmal AF prediction challenge dataset, Asadi et al. (2023) [11] developed a generative adversarial network (GAN) that synthesized ECGs of the paroxysmal AF minority class and designed a neural architecture search that customized the convolutional neural network (CNN) for the binary classification of paroxysmal AF versus no AF. Qin et al. (2023) [34] proposed a model that detected ECG anomalies using a one-class classification GAN with bidirectional long short-term memory (LSTM) and mini-batch discrimination, attaining high accuracy and discrimination for the binary classification of normal versus abnormal ECGs in the MIT-BIH arrhythmia database. Overall, the reviewed studies support the utility of machine learning for arrhythmia detection. Challenges in terms of variations in performance and model robustness, generalizability and real-world applicability beyond the study databases, and computational demands with practical feasibility in clinical settings were emphasized. Importantly, specific approaches to mitigating data imbalance were highlighted.

Apart from arrhythmia detection, researchers have explored the application of data imbalance techniques in the ECG-based detection of myocardial infarction (MI). Sharma and Sunkaria (2020) [44] addressed data imbalance in the Physikalisch-Technische Bundesanstalt (PTB) database using adaptive synthetic sampling. Their proposed model based on stationary wavelet transform and k-nearest neighbors attained consistent accuracy using top-ranked features from multiple ECG leads. Li et al. (2022) [45] used a single-lead convolutional GAN to synthesize minority-class myocardial infarct ECGs on the PTB dataset. Their model attained 99.06% accuracy. Working with both the MIT-BIH and PTB databases, Rai and Chatterjee (2021) [46] used the synthetic minority oversampling technique (SMOTE)-Tomek Link to balance the combined study dataset, while acknowledging that data combination might introduce signal morphology variations. Using CNN, hybrid CNN-LSTM, and ensemble techniques, they attained 99.89% accuracy for myocardial infarction detection.

1.2. Novelties and Contributions

In this work, we present three DL models, SkipCNN, SkipCNN+LSTM, and SkipCNN+LSTM+Attention, for the accurate and automatic detection of cardiac arrhythmias. The design structure of these models is unique and different from that of the other models in the literature as it includes a reduced number (optimized) of layers, resulting in enhanced computational efficiency and improved generalization performance. The integration of skip connections with a CNN facilitates the simplified flow of the data through the layers, providing deeper learning without the risk of a vanishing gradient. The addition of LSTM units enables the model to capture the temporal dependencies very effectively, as the LSTM is well suited for handling time series data such as ECG or EEG signals. The utilization of an attention mechanism in our proposed model permits us to emphasize the most significant segments of the input sequence, boosting the capability to extract the complex and discriminative features.

In addition, the proposed structures make use of the Swish activation function rather than the conventional ReLU. Swish was selected because it can provide more balanced gradients and enhance training dynamics, which will enhance the performance of the model. Furthermore, the employed GAN structure provides a lower computational cost than others described in the literature. Both the generator and discriminator sections of our GAN model utilize Bi-LSTM to increase the quality of synthetic data and boost the model’s overall performance. The detailed description of the layer-wise structure of the proposed DL models and GAN architecture is presented in Section 3. The contributions and novelties of this research are as follows:

• Innovative GAN model with bidirectional LSTM (Bi-LSTM) to generate two minority class signals to address data imbalance.
• A large ECG dataset comprising 109,446 ECG beats divided into five arrhythmia classes to promote model robustness and real-world applicability.
• Similarity matching assessment using key metrics to quantify similarities between real and synthetic ECG signals, providing insights into the quality of synthetic data.
• Innovative convolutional neural network with skip connections (SkipCNN) model to enrich model diversity.
• Diverse model architectures encompassing three distinct models—SkipCNN, SkipCNN+LSTM, and SkipCNN+LSTM+Attention—which enhance versatility in studying different arrhythmia scenarios.
• Ensemble model integration to combine the strengths of individual models and enhance system accuracy.
• Utilization of multiple evaluation metrics for a comprehensive evaluation of model performance.
• Data imbalance solution based on synthetic ECG signal generation using GAN.
• Adequate ECG beat generation by proposed GAN model of 2400 premature atrial contraction (PAC) ECG signals from a training dataset comprising 2085 samples.
• Combined dataset utilization MIT-BIH and GAN-generated synthetic ECG data to enrich the training data.

The rest of the manuscript is organized in the subsequent manner. Section 2 covers the materials and techniques employed in the research. Section 3 introduces the proposed methodology for detecting arrhythmias in five categories. Section 4 provides a comprehensive account of the experiments, including the experimental setup and results. Section 5 engages in a detailed discussion, drawing comparisons with state-of-the-art techniques. Finally, Section 6 offers closing remarks and outlines future research directions.

2. Materials and Methods

2.1. ECG Dataset

Our study dataset is primarily centered on the open-access MIT-BIH arrhythmia dataset, which comprises 48 30-minute ECG recordings of 47 individuals [31]. The 109,446 ECG beats in the recordings have been manually annotated beat-by-beat by experts into multiple distinct arrhythmia classes [47]. We downloaded the ECG beats from Kaggle [48], which have been segmented and individually mapped to the five classes of the Association for the Advancement of Medical Instrumentation (AAMI) standard classification framework—normal (N), supraventricular (S), ventricular (V), fusion (F), and beats of unknown etiology (Q)—(

Table 2. Distribution of ECG beats in MIT-BIH database mapped to AAMI standard.

AAMI Label	MIT-BIH Label	Beats, n	Beats, %
Normal (N)	Normal, nodal escape, atrial escape, right bundle branch block, left bundle branch block	90,589	82.77%
Supraventricular (S)	Supraventricular premature, atrial premature, nodal premature, aberrant atrial premature	2779	2.54%
Ventricular (V)	Ventricular escape, premature ventricular contraction	7236	6.61%
Fusion (F)	Fusion of ventricular and normal	803	0.73%
Beats of unknown etiology (Q)	Unclassifiable, paced, fusion of paced and normal	8039	7.35%
Total		109,446	100%

) [47,49]. Due to variations in the lengths of the extracted ECG beats, we employed zero padding to ensure a uniform fixed length of 187 for the analyzed ECG segments (Figure 1), as previously described [40,46]. These pre-processed ECG segments constituted our study dataset.

The MIT-BIH ECG database and the derived AAMI five-class dataset are imbalanced datasets, in which the collected data are distributed unevenly among different classes [15]. In the context of multi-class classification, data imbalance refers to the unequal distribution of samples among multiple classes in the training dataset [50]. From Table 2, the majority, second lowest minority, and lowest minority AAMI classes comprise N, S, and F ECG beats, which constitute 82.77%, 2.54%, and 0.73% of the dataset, respectively, indicating severe data imbalance. Under such conditions, machine learning models struggle to learn from and accurately predict minority class samples, often resulting in poor minority class accuracy. The challenges imposed by data imbalance in multi-class classification closely resemble those encountered in binary classification tasks—encompassing issues like model bias, suboptimal generalization, and the potential for misleading model evaluation—but are exacerbated due to the presence of multiple imbalanced classes. Techniques commonly applied in binary classification problems—including oversampling, undersampling, and cost-sensitive learning—can be extended and adapted to balance sample distributions across all classes to address data imbalance in multi-class settings and enhance the overall classification performance and fairness of the machine learning models.

2.2. CNN Model

CNNs are versatile and effective deep learning models for automatic feature extraction from intricate data [51,52,53,54]. CNNs comprise several standard layers: convolutional layers for feature extraction, batch normalization for data standardization, rectified linear unit (ReLU) activation for introducing non-linearity, and pooling layers to reduce dimensionality, while fully connected layers, also known as a feed-forward network, are utilized for detection and classification tasks [55,56]. Convolutional layers employ filters for extracting pertinent information called features from the input information; their size and stride can be tailored to the specific task, defined by mathematical Equation (1) (in all equations, all vectors, matrices, and tensors are annotated in bold).

$${\mathit{Y}}_{i,j,k}={\left(\mathit{X}\ast \mathit{F}\right)}_{i,j,k}+b_k$$

(1)

where $\ast$ represents the convolution operation, $i \mathrm{a}\mathrm{n}\mathrm{d} j$ are spatial indices, $k$ is the feature map index, and $b_k$ is the bias term. Batch normalization layers standardize input characteristics, speeding up the learning process and improving the neural network’s stability. For a mini-batch $\mathcal{B}=\{{\mathit{x}}_1,x_2,\dots ,{\mathit{x}}_m\},$ the normalized output $\left(\widehat{x_i}\right)$ is provided by Equation (2):

$$\widehat{{\mathit{x}}_i}={\displaystyle \frac{{\mathit{x}}_i-{\mathsf{\mu }}_B}{\sqrt{{\mathsf{\sigma }}_B^2+\mathsf{ϵ}}}}$$

(2)

The mini-batch mean and variance are represented by ${\mathsf{\mu }}_B$ and ${\mathsf{\sigma }}_B^2,$ respectively, with $\mathsf{ϵ}$ being an insignificant constant for numerical stability. The mathematical model’s ability to acquire complex patterns is supported through the introduction of non-linearity via the ReLU activation function. Pooling layers, such as average pooling and maximum pooling, lower the dimensionality of feature maps, thus accelerating processing as well as facilitating feature selection [57]. The max-pooling and average pooling layers are mathematically defined by Equations (3) and (4), respectively.

$${\mathit{Y}}_{i,j,k}=\underset{\left(m,n\right)\in P}{\max }{\mathit{X}}_{i+m,j+n,k}$$

(3)

$${\mathit{Y}}_{i,j,k}={\displaystyle \frac{1}{\left|P\right|}}\sum _{\left(m,n\right)\in P}{\mathit{X}}_{i+m,j+n,k}$$

(4)

where $P$ is the pooling region. A fully connected layer at the end of the network categorizes input data into distinct output classes, leveraging on high-level features extracted by the previous layers [58], defined by Equation (5).

$$\mathit{z}={\mathit{W}}^T\mathit{x}+\mathit{b}$$

(5)

where $\mathit{W}$ is the weight matrix, x is the input vector, and $\mathit{b}$ is the bias vector.

2.3. LSTM Model

LSTM is adept at processing time series and sequential data like ECG signals. LSTM effectively addresses short-term memory problems in deep learning models, which occur when earlier layers in the network struggle to learn and retain information from extended data sequences due to vanishing gradients [36,59,60,61]. The key components of an LSTM cell are defined mathematically in Equations (6)–(11) as follows:

$$\mathrm{Forget} \mathrm{Gate} \left(f_t\right)=\mathsf{\sigma }\left({\mathit{W}}_f\cdot \left[{\mathit{h}}_{t-1},{\mathit{x}}_{\mathit{t}}\right]+{\mathit{b}}_f\right)$$

(6)

$$\mathrm{Input} \mathrm{Gate} \left(i_t\right)=\mathsf{\sigma }\left({\mathit{W}}_i\cdot \left[{\mathit{h}}_{t-1},{\mathit{x}}_t\right]+{\mathit{b}}_i\right)$$

(7)

$$\mathrm{Candidate} \mathrm{Cell} \mathrm{State} \left(\widetilde{C_t}\right)=\tanh \left({\mathit{W}}_C\cdot \left[{\mathit{h}}_{t-1},{\mathit{x}}_t\right]+{\mathit{b}}_C\right)$$

(8)

$$\mathrm{Cell} \mathrm{State} \left(C_t\right)=f_t\odot C_{t-1}+i_t\odot \widetilde{C_t}$$

(9)

$$\mathrm{Hidden} \mathrm{Sate} \left(h_t\right)=o_t\odot \tanh \left(C_t\right)$$

(10)

$$\mathrm{Output} \mathrm{Gate} \left(o_t\right)=\sigma \left({\mathit{W}}_o\cdot \left[{\mathit{h}}_{t-1},{\mathit{x}}_t\right]+{\mathit{b}}_o\right)$$

(11)

where $\mathsf{\sigma }$ is the sigmoid activation function; ${\mathit{W}}_{\mathit{f}},{\mathit{W}}_{\mathit{i}},{\mathit{W}}_{\mathit{C}},$ and ${\mathit{W}}_o$ are weight matrices; ${\mathit{h}}_{t-1}$ is the hidden state from the previous time step; ${\mathit{x}}_t$ is the input at the current time step; ${\mathit{b}}_{\mathit{f}},{\mathit{b}}_{\mathit{i}},{\mathit{b}}_{\mathit{C}},\mathrm{a}\mathrm{n}\mathrm{d} {\mathit{b}}_{\mathit{o}}$ are the bias; $(\cdot )$ represents the inner product (dot product or matrix multiplication); and $\odot$ represents the Hadamard product (element-wise multiplication).

2.4. Bidirectional LSTM

The Bi-LSTM network is an expansion of the LSTM architecture that integrates intelligence derived from time steps both in antecedent and subsequent temporal domains, which enhances its ability to capture temporal dependencies in sequential data [62,63]. The basic structure of the Bi-LSTM used in this work is illustrated in Figure 2a. For a sequence of inputs $\left(\right\{{\mathit{x}}_1,{\mathit{x}}_2,\dots ,{\mathit{x}}_T\left\}\right)$, the forward LSTM processes the sequence from $(t=1 to T)$ as given by Equation (12) [64,65].

$$\overrightarrow{{\mathit{h}}_t}=\overrightarrow{LSTM}\left({\mathit{x}}_t,{\overrightarrow{\mathit{h}}}_{t-1}\right)\mathrm{for}t=1 \mathrm{to} T$$

(12)

The backward LSTM processes the sequence from $(t=\mathrm{T} \mathrm{t}\mathrm{o} 1)$, as given by Equation (13).

$$\overleftarrow{{\mathit{h}}_t}=\overleftarrow{LSTM}\left({\mathit{x}}_t,{\overleftarrow{\mathit{h}}}_{t+1}\right)\mathrm{for}t=T \mathrm{to} 1$$

(13)

The final output at each time step $t$ is given by Equation (14).

$${\mathit{h}}_t=\left[\overrightarrow{{\mathit{h}}_t};\overleftarrow{{\mathit{h}}_t}\right]$$

(14)

Figure 2. (a) Schema of Bi-LSTM architecture (left) and (b) schema of attention model architecture (right).

Figure 2. (a) Schema of Bi-LSTM architecture (left) and (b) schema of attention model architecture (right).

2.5. Attention Model

The attention model is a crucial component in sequence-related tasks, enabling the model to concentrate on specific segments of an input sequence during each iteration, making it highly effective for understanding context in long sequences; for example, text summarization and machine translation provide pertinent instances [66,67,68,69]. In the context of sequence generation tasks, an attention mechanism is employed for various calculations. The compatibility score $\left(\widetilde{{\mathit{h}}_t}\right)$ quantifies the alignment between source hidden states $\left({\mathit{h}}_s\right)$ and the target hidden state $\left({\mathit{h}}_t\right)$. Alignment scores $\left(a_s\right)$ are derived from these compatibility scores using a softmax function, making them interpretable as probabilities [70]. These scores determine the attention weights assigned to source hidden states. The context vector $\left({\mathit{c}}_t\right)$ is computed as a weighted sum of source hidden states $\left({\mathit{h}}_s\right)$ based on alignment scores $\left(a_s\right)$, representing the relevant part of the source sequence for generating the target sequence at time step $\left(t\right)$ (Figure 2b). Finally, the decoder output $\left({\mathit{y}}_t\right)$ is predicted based on the context vector, the previous hidden state $\left({\mathit{h}}_{t-1}\right)$, and the previously generated token $\left({\mathit{y}}_{t-1}\right)$, if available, using the decoder’s LSTM cell [71] (see Algorithm 1). The equations for the compatibility scores, source hidden states, alignment score, target hidden state, context vector, and decoder output are provided in Equations (15)–(20), respectively [72].

$$\widetilde{{\mathit{h}}_t}={\mathit{v}}_a^{\mathrm{\top }}\tanh \left({\mathit{W}}_a\left[{\mathit{h}}_s;{\mathit{h}}_t\right]\right)$$

(15)

$${\mathit{h}}_s=\mathrm{LSTM}\left({\mathit{X}}_s,{\mathit{h}}_{s-1}^{\left(s\right)},{\mathit{c}}_{s-1}^{\left(s\right)}\right)$$

(16)

$${\mathit{h}}_t=\mathrm{LSTM}\left({\mathit{X}}_t,{\mathit{h}}_{t-1}^{\left(t\right)},{\mathit{c}}_{t-1}^{\left(t\right)}\right)$$

(17)

$$a_s={\displaystyle \frac{\exp \left(\widetilde{{\mathit{h}}_t}\right)}{{\sum }_{s^{\prime }}\exp \left(\widetilde{{\mathit{h}}_{t^{\prime }}}\right)}}$$

(18)

$${\mathit{c}}_t={\sum }_s\left(a_s\cdot {\mathit{h}}_s\right)$$

(19)

$${\mathit{y}}_t=\mathrm{DecoderLSTM}\left({\mathit{c}}_t,{\mathit{h}}_{t-1},{\mathit{y}}_{t-1}\right)$$

(20)

where ${\mathit{X}}_s$ is the input at source time step $\left(s\right),\left({\mathit{h}}_{s-1}^{\left(s\right)}\right)$ is the previous hidden state of the LSTM at source time step $\left(s\right), \mathrm{a}\mathrm{n}\mathrm{d} \left({\mathit{c}}_{s-1}^{\left(s\right)}\right)$ is the previous cell state of the LSTM at source time step $s$.

Algorithm 1: Attention mechanism for sequence generation.

Initialization: $\{{\mathit{h}}_{\mathit{s}}\text{:}$ Source hidden states, ${\mathit{h}}_{\mathit{t}}\text{:}$ Target hidden state, ${\mathit{y}}_{\mathit{t}-\mathbf{1}}\text{:}$ Previously generated token, ${\mathit{c}}_{\mathit{t}}\text{:}$ Context vector, $\widetilde{{\mathit{h}}_{\mathit{t}}}\text{:}$ Compatibility score, ${\mathit{a}}_{\mathit{s}}\text{:}$ Alignment score, ${\mathit{y}}_{\mathit{t}}\text{:}$ Decoder output, ${\mathit{N}}_{\mathit{t}}\text{:}$ Length of target sequence} Input: $\{{\mathit{h}}_{\mathit{s}},{\mathit{h}}_{\mathit{t}},{\mathit{y}}_{\mathit{t}-\mathbf{1}}$} Output: {${\mathit{y}}_{\mathit{t}}$} For each target time step $t$ from 1 to $N_t$: Do process: $\widetilde{{\mathit{h}}_t}=\mathrm{score}\left({\mathit{h}}_s,{\mathit{h}}_{\mathit{t}}\right)$ Do process: $a_s={\displaystyle \frac{\exp \left(\widetilde{{\mathit{h}}_t}\right)}{{\sum }_{s^{\prime }}\exp \left(\widetilde{{\mathit{h}}_{t^{\prime }}}\right)}}$ Do process: ${\mathit{c}}_t={\sum }_s\left(a_s\cdot {\mathit{h}}_s\right)$ Do process: ${\mathit{y}}_t=\mathrm{DecoderLSTM}\left({\mathit{c}}_t,{\mathit{h}}_{t-1},{\mathit{y}}_{t-1}\right)$ Update: ${\mathit{h}}_{t-1}\leftarrow {\mathit{h}}_t,{\mathit{y}}_{t-1}\leftarrow {\mathit{y}}_t$ End For

2.6. Performance Evaluation Metrics

In our analysis, we utilized standard evaluation metrics, as described in Equations (21)–(30).

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(21)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(22)

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(23)

$$\mathrm{F}1\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\displaystyle \frac{2\ast \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\ast \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}}$$

(24)

$$\mathrm{W}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}=\sum _{i=1}^n{\displaystyle \frac{{\left(\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}\right)}_i}{n}}$$

(25)

$$\mathrm{W}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}=\sum _{i=1}^n{\displaystyle \frac{{\mathrm{C}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}}_i\mathrm{W}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}.{\mathrm{M}\mathrm{e}\mathrm{t}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{s}}_i}{Total Samples}}$$

(26)

$$\mathrm{M}\mathrm{a}\mathrm{c}\mathrm{r}\mathrm{o} \mathrm{a}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}=\sum _{i=1}^n{\displaystyle \frac{{\mathrm{M}\mathrm{e}\mathrm{t}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{s}\_\mathrm{C}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}}_i}{n}}$$

(27)

$$\mathrm{S}\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{c}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l} \mathrm{s}\mathrm{i}\mathrm{m}\mathrm{i}\mathrm{l}\mathrm{a}\mathrm{r}\mathrm{i}\mathrm{t}\mathrm{y} \mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x} (\mathrm{S}\mathrm{S}\mathrm{I}\mathrm{M}\left(\mathrm{G},\mathrm{R}\right))={\displaystyle \frac{\left(2{\mathsf{\mu }}_{\mathrm{G}}{\mathsf{\mu }}_{\mathrm{R}}+\mathrm{C}1\right)\times \left(2{\mathsf{\sigma }}_{\mathrm{G}\mathrm{R}}+\mathrm{C}2\right)}{\left({\mathsf{\mu }}_{\mathrm{G}}^2+{\mathsf{\mu }}_{\mathrm{R}}^2+\mathrm{C}1\right)\times \left({\mathsf{\sigma }}_{\mathrm{G}}^2+{\mathsf{\sigma }}_{\mathrm{R}}^2+\mathrm{C}2\right)}}$$

(28)

$$\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n} \mathrm{s}\mathrm{q}\mathrm{u}\mathrm{a}\mathrm{r}\mathrm{e} \mathrm{e}\mathrm{r}\mathrm{r}\mathrm{o}\mathrm{r} (\mathrm{M}\mathrm{S}\mathrm{E}(\mathrm{G},\mathrm{R}))={\displaystyle \frac{\sum {({\mathrm{G}}_{\mathrm{i}}-{\mathrm{R}}_{\mathrm{i}})}^2}{\mathrm{N}}}$$

(29)

$$\mathrm{P}\mathrm{e}\mathrm{a}\mathrm{r}\mathrm{s}\mathrm{o}\mathrm{n} \mathrm{c}\mathrm{o}\mathrm{r}\mathrm{r}\mathrm{e}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n} \mathrm{c}\mathrm{o}\mathrm{e}\mathrm{f}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{e}\mathrm{n}\mathrm{t} (\mathrm{r}(\mathrm{G},\mathrm{R}))={\displaystyle \frac{\sum \left(\right({\mathrm{G}}_{\mathrm{i}}-{\mathsf{\mu }}_{\mathrm{G}})\ast ({\mathrm{R}}_{\mathrm{i}}-{\mathsf{\mu }}_{\mathrm{R}}\left)\right)}{({\mathsf{\sigma }}_{\mathrm{G}}\ast {\mathsf{\sigma }}_{\mathrm{R}})}}$$

(30)

Here, TP, FP, TN, and FN signify the quantities of true positives, false positives, true negatives, and false negatives, respectively, and n is the number of samples [73]. ${\mathsf{\mu }}_{\mathrm{G}}$ and ${\mathsf{\mu }}_{\mathrm{R}}$ represent means of G and R, respectively; ${\mathsf{\sigma }}_{\mathrm{G}}$ and ${\mathsf{\sigma }}_{\mathrm{R}}$ are standard deviations of G and R, respectively; ${\mathsf{\sigma }}_{\mathrm{G}\mathrm{R}}$ is the covariance of G and R; and C1 and C2 are small constants to stabilize the division [74]. ${\mathrm{G}}_i$ and ${\mathrm{R}}_i$ represent corresponding data points in G and R, respectively, and N is the total data samples [75].

3. Proposed Methodology

Our model is illustrated in Figure 3. To address data imbalance, we proposed to use GAN to generate synthesized ECGs and extend the representation of minority arrhythmia categories in the MIT-BIH-derived study dataset. The generated synthetic data were then partitioned into training and testing sets, ensuring that the model would be tested on unseen data separate from the training set, thus preempting bias. For end-to-end model training and classification, we employed three deep learning architectures: (1) a convolutional neural network with skip connections (SkipCNN), the foundational unit of our proposed model, which, by incorporating skip connections, enables the network to skip one or more layers to facilitate the more efficient capture of both local and global features; (2) SkipCNN+LSTM, a hybrid model that combines the capabilities SkipCNN and LSTM to process spatial and temporal information, respectively; (3) SkipCNN+LSTM+Attention, which integrates an attention mechanism to the hybrid model to boost model performance through selective emphasis on and capture of pertinent information in the input data; and (4) an ensemble model, which leverages the complementary strengths of different model architectures to improve model accuracy for arrhythmia detection. The proposed models, SkipCNN, SkipCNN+LSTM, SkipCNN+LSTM+Attention, and Ensemble, are mathematically defined by Equations (31), (32), (33), and (34), respectively.

$${\mathit{y}}_{\mathrm{S}\mathrm{k}\mathrm{i}\mathrm{p}}=\mathcal{F}\left(\mathit{x},\left\{{\mathit{W}}_i\right\}\right)+\mathit{x}+\mathit{b}$$

(31)

$$\begin{array}{cc}\hfill {\mathit{h}}_t,{\mathit{c}}_t=& \left[\sigma ,\sigma ,\sigma ,tanh\right]\ast \mathit{W}\ast \left[\mathit{F}\left({\mathit{x}}_t,{\mathit{W}}_i\right)+\mathit{x}+\mathit{b};{\mathit{h}}_{\left(t-1\right)}\right]+\left[{\mathit{b}}_i,{\mathit{b}}_f,{\mathit{b}}_o,{\mathit{b}}_c\right]\hfill \\ & \hspace{1em}\hspace{1em}\hspace{1em} \oplus \left(f_t\odot {\mathit{c}}_{\left(t-1\right)}+i_t\odot tanh\left({\mathit{W}}_c\ast \left[\mathit{F}\left({\mathit{x}}_t,{\mathit{W}}_i\right)+\mathit{x}+\mathit{b};{\mathit{h}}_{\left(t-1\right)}\right]+{\mathit{b}}_c\right)\right)\hfill \\ & \hspace{1em}\hspace{1em}\hspace{1em} \odot tanh\left({\mathit{c}}_t\right)\hfill \end{array}$$

(32)

$${\mathit{h}}_t=softmax\left(f_A\left({\mathit{h}}_t,{\mathit{W}}_a\ast \left[\mathit{F}\left({\mathit{x}}_t,{\mathit{W}}_i\right)+\mathit{x}+\mathit{b};{\mathit{h}}_{\left(t-1\right)}\right]\right)\right)\ast \left[\mathit{F}\left({\mathit{x}}_t,{\mathit{W}}_i\right)+\mathit{x}+\mathit{b};{\mathit{h}}_{\left(t-1\right)}\right]$$

(33)

$${\mathit{y}}_{\mathrm{ensemble}}=w_{\mathrm{SkipCNN}}\cdot {\mathit{y}}_{\mathrm{SkipCNN}}+w_{\mathrm{LSTM}}\cdot {\mathit{h}}_{t_{\mathrm{Attention}}}+w_{\mathrm{Attention}}\cdot f\left({\mathit{h}}_{t_{\mathrm{LSTM}}},{\mathit{c}}_{t_{\mathrm{LSTM}}}\right)$$

(34)

where $\oplus$ is element-wise addition, $\odot$ is element-wise multiplication, $\ast$ is a convolution operation, and $\cdot$ is the dot product or matrix multiplication. The weights denoted as $w_{\mathrm{SkipCNN}},w_{\mathrm{LSTM}},$ and $w_{\mathrm{Attention}}$ are determined via a training process. In order to reduce the ensemble model’s total loss function, these weights are often tuned using backpropagation. Through the use of each component model’s strengths, the optimization process modifies these weights in accordance with how well it performs on the validation set, enabling the ensemble to provide predictions that are more accurate overall.

3.1. Proposed GAN Architecture

We designed a novel GAN architecture based on Bi-LSTM for generating minority ECG samples, which comprised two primary components: a generator and a discriminator, both utilizing Bi-LSTM networks (Figure 4).

3.1.1. Generator

This employs a combination of layers, including the core Bi-LSTM layer, to transform an input ECG sequence into a desired ECG output format. A vector of random noise $\mathit{z}$ of size $\{batch\_size\}\times \mathrm{Noise}\_\dim$ is given to the generator. A sequence of fully connected (FC) layers is succeeded by the Bi-LSTM layer [76]. The first FC layer applies a linear transformation to the input noise, which increases the representational capacity of the data, as shown in Equation (35).

$${\mathit{h}}_1=\max \left(0,{\mathit{W}}_1\mathit{z}+{\mathit{b}}_1\right)$$

(35)

A second fully connected layer performs a similar operation, further enhancing the capacity, as shown in Equation (36).

$${\mathit{h}}_2=\max \left(0,{\mathit{W}}_2{\mathit{h}}_1+{\mathit{b}}_2\right)$$

(36)

To capture temporal dependencies, the modified input is processed by the core Bi-LSTM layer (shown in Equation (37)).

$$\left({\mathit{h}}_t,{\mathit{c}}_t\right)=Bi-LSTM\left({\mathit{h}}_{t-1},{\mathit{x}}_t\right)=\left\{\left({\overrightarrow{\mathit{h}}}_t,{\overleftarrow{\mathit{h}}}_t\right),\left(\overrightarrow{{\mathit{c}}_t},\overleftarrow{{\mathit{c}}_t}\right)\right\}$$

(37)

With the aim of averting overfitting and upholding the model’s propensity for generalization, dropout regularization is next applied: 20% of input units are randomly set to zero during training, as shown in Equation (38).

$${\mathit{h}}_{\mathrm{dropout}}={\mathit{h}}_t\cdot \mathrm{Bernoulli}\left(1-{\displaystyle \frac{1}{1+{\left(\frac{\left|{\mathit{h}}_{ti}\right|}{\alpha }\right)}^{\alpha }}}\right)$$

(38)

where the binary mask represented by $\mathrm{B}\mathrm{e}\mathrm{r}\mathrm{n}\mathrm{o}\mathrm{u}\mathrm{l}\mathrm{l}\mathrm{i}\left(1-p_i\right)$ keeps elements with probability $1-p_i,$ and $h_t$ is the input tensor. The final fully connected layer generates the output sequence, which corresponds to ECG data of a specific uniform length (187), shown in Equation (39).

$$\mathit{y}={\mathit{W}}_{\mathrm{out}}{\mathit{h}}_{\mathrm{dropout}}+{\mathit{b}}_{\mathrm{out}}$$

(39)

The linear transformation carried out by this layer maps the model’s internal representation to the desired output format. To ensure that the output adheres to the desired shape, a reshaping operation is applied. This ensures that the output sequence retains its batch size while taking the desired shape of (−1, 1, 187), where “−1” denotes that the batch size is preserved.

3.1.2. Discriminator

The architecture incorporates Bi-LSTM and is designed to differentiate between real ECG beats sourced from the MIT-BIH arrhythmia dataset and synthetic ECG beats generated by the generator. To capture the temporal features from the input data sequence (ECG), the Bi-LSTM layers are used as defined in Equation (40).

$$\left({\mathit{h}}_t,{\mathit{c}}_t\right)=\mathrm{Bi}-\mathrm{LSTM}\left({\mathit{W}}_{\mathrm{out}}{\mathit{h}}_{\mathrm{dropout}}+{\mathit{b}}_{\mathrm{out}}\right)$$

(40)

The first fully connected layer applies a linear transformation to the Bi-LSTM output, which increases the representational capacity, as shown in Equation (41).

$${\mathit{h}}_1=\left\{\begin{array}{cc}\hfill {\mathbf{W}}_1{\mathbf{h}}_{\mathrm{t}}+{\mathbf{b}}_1,& \mathrm{i}\mathrm{f} {\mathbf{W}}_1{\mathbf{h}}_{\mathrm{t}}+{\mathbf{b}}_1>0\hfill \\ \hfill 0,& \mathrm{O}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\hfill \end{array}\right.$$

(41)

The second FC layer is applied to further process the data, priming it for the discrimination task, as given in Equation (42).

$${\mathit{h}}_2=\left\{\begin{array}{cc}\hfill {\mathbf{W}}_2{\mathbf{h}}_1+{\mathbf{b}}_1,& \mathrm{i}\mathrm{f} {\mathbf{W}}_2{\mathbf{h}}_1+{\mathbf{b}}_1>0\hfill \\ \hfill 0,& \mathrm{O}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\hfill \end{array}\right.$$

(42)

To prevent overfitting, dropout regularization is next applied: random deactivation is applied to around 20% of the neurons at each forward pass during training. This regularization technique boosts the model’s effectiveness in generalizing efficiently to unseen data. The final fully connected layer serves as a binary classifier using a sigmoid activation function, and the layer output falls within the range of [0, 1], as given in Equation (43).

$$p={\displaystyle \frac{1}{1+\exp \left(-\left({\mathit{W}}_{\mathrm{out}}{\mathit{h}}_{\mathrm{dropout}}+{\mathit{b}}_{\mathrm{out}}\right)\right)}}$$

(43)

This output embodies the discriminator’s assessment of the input data: values approaching 1 indicate that the input ECG data closely resemble real data from the MIT-BIH arrhythmia database, whereas values approaching 0 suggest that the input data are more likely synthetic, i.e., generated by the generator. The synthetic ECG signal generated using the proposed GAN is presented by Algorithm 2.

To address the data imbalance issue, we proposed a unique GAN model that generates a synthetic ECG dataset. There are many reasons why it is vital to distinguish between ECGs obtained from the original dataset and those generated by the generator. First, this distinction enables us to evaluate the model’s endurance and its ability to generalize well. By assessing how well the model distinguishes between actual and synthetic ECGs, we can ensure that it performs consistently across both types of data. Second, it is required to validate the quality of the synthetic data generated. The synthetic ECGs should closely resemble the properties of real ECGs, and differentiating between the two aids in ensuring that the created data appropriately reflect the variability and patterns observed in the actual dataset. In addition, this differential is critical for the discriminator in our GAN model, since it must properly differentiate between real and synthetic ECGs in order to train both itself and the generator. Finally, examining how the model differentiates between actual and synthetic ECGs over time offers useful insights into its performance and serves as a roadmap for future improvements.

Algorithm 2: ECG synthetic data generation using GAN with Bi-LSTM.

$\mathrm{Initialization}\text{:} \{\mathit{G}\text{:}$ Generator, $\mathit{D}\text{:}$ Discriminator, $\mathit{z}\text{:}$ Random noise vector, $\mathit{x}\text{:}$ Synthetic ECG data, ${\mathit{x}}_{\mathbf{real}}\text{:}$ Real ECG samples, ${\mathit{x}}_{\mathbf{fake}}\text{:}$ Generated ECG samples, ${\mathit{a}}_{\mathit{s}}\text{:}$ Alignment score, ${\mathit{y}}_{\mathit{t}}\text{:}$ Decoder output, ${\mathit{N}}_{\mathit{t}}\text{:}$ Length of target sequence} $\mathrm{Input}\text{:} \{\mathbf{z},{\mathit{x}}_{\mathit{r}\mathit{e}\mathit{a}\mathit{l}},{\mathit{x}}_{\mathbf{fake}}$} $\mathrm{Output}\text{:} \{\mathit{x},P_{\mathrm{real}},P_{\mathrm{fake}}$} $\mathbf{For} \mathrm{each} \mathrm{batch} \mathrm{of} \mathrm{random} \mathrm{noise} \mathrm{vector} \mathit{z}$:                  $\mathit{x}=G\left(\mathit{z}\right)$ If real ECG samples $\left({\mathit{x}}_{\mathrm{real}}\right)$ are provided:     $P_{\mathrm{real}}=D\left({\mathit{x}}_{\mathrm{real}}\right)$   Else:                  $P_{\mathrm{real}}=0$   If generated ECG samples $\left({\mathit{x}}_{\mathrm{fake}}\right)$ are provided:     $P_{\mathrm{fake}}=D\left({\mathit{x}}_{\mathrm{fake}}\right)$ Else:        $P_{\mathrm{fake}}=0$ 5. End For

3.2. Proposed Model Architectures

This work centers around the detection of arrhythmias in the MIT-BIH ECG dataset as well as GAN-generated synthetic ECG signals. To classify the ECG rhythms into distinct AAMI categories, we introduced three deep learning neural networks: SkipCNN, SkipCNN+LSTM, and SkipCNN+LSTM+Attention (Figure 5).

The SkipCNN model (Neural Net No.1) comprises three ConvNormPool blocks, which perform ECG feature extraction and pattern recognition processes. Every ConvNormPool block comprises three 1D convolutional layers (Conv1d), each with a kernel size of 5, that capture local patterns and features, followed by batch normalization layers (BatchNorm1d) to standardize the inputs, ensuring stable and efficient training. Non-linear characteristics are introduced by integrating Swish activation functions, thus providing the model with the capability to acquire intricate data relationships. Furthermore, each ConvNormPool block contains a max-pooling layer, and the spatial dimensions of feature maps are reduced, while the most crucial information is retained in the process, rendering the data more manageable for subsequent layers. Of note, the inclusion of skip connections facilitates the flow of gradients during training, thereby mitigating vanishing gradients, which helps the model learn more effectively. A final FC layer with a kernel size of 5 and 128 hidden neurons aggregates features from preceding layers to make final predictions using its softmax function. Similar to the convolutional layers, the fully connected layer benefits from a skip connection. This connection ensures that the model maintains access to critical information from earlier stages, even as it nears the final classification decision.

The SkipCNN+LSTM model (Neural Net No.2) combines SkipCNN and LSTM. The SkipCNN component comprises two ConvNormPool blocks. Each ConvNormPool block contains three 1D convolutional layers (Conv1d)—each of kernel size 5—that are augmented with batch normalization layers (BatchNorm1d) and Swish activation functions to facilitate feature extraction and maintain training stability. Max-pooling layers, each of kernel size 2, are added. Notably, a skip connection is incorporated between the first and second convolutional layer blocks to ensure smooth gradient flow and preserve crucial information. The LSTM component comprises a Bi-LSTM of hidden size 64, which allows the model to consider both past and future contexts during learning, enhancing the ability to recognize intricate temporal patterns and relationships within the ECG signals. The overall model architecture is completed with a fully connected layer (fc)—with kernel size 5 and hidden size 64—which aggregates features to make final predictions. It benefits from both bidirectional processing and skip connection; the former allows the model to capture temporal information from both directions in the sequential data, and the latter facilitates the retention of essential information from earlier stages of feature extraction. The SkipCNN+LSTM+Attention model (Neural Net No.3) adds an attention mechanism to the above. Similar to the aforementioned models, the SkipCNN component comprises two ConvNormPool blocks, within each are three 1D convolutional layers (Conv1d) with kernel size 5, batch normalization layers (BatchNorm1d), and Swish activation functions. Max-pooling layers, each with kernel size 2, are included to downsample the feature maps. A skip connection between the first and third convolutional layers enhances the model’s gradient flow while preserving vital information. After the third block’s max-pooling layer, an attention layer (attn) is integrated. Attention mechanisms have proven effective in focusing on specific portions of the input sequence that are most relevant for making predictions. In this case, a linear layer is used to implement the attention mechanism, enabling the model to dynamically weigh the importance of different elements within the ECG signal. An LSTM layer (which is not bidirectional), configured with a hidden size of 64, is incorporated, allowing the model to capture temporal patterns in a unidirectional manner. The architecture is completed with a fully connected layer (fc), with kernel size 5 and hidden size 64.

4. Experimental Results

The study dataset comprised 109,446 ECG beats from the MIT-BIH arrhythmia dataset, which had been mapped into five significantly imbalanced classes (Table 2). To address this imbalance, we first focused on generating a synthetic ECG dataset for the two minority classes, S and F (Table 2), using a novel GAN that incorporated a Bi-LSTM architecture in both its generator and discriminator functions (Section 3.1). By generating synthetic ECG beats that closely resembled real ECG beats from the minority classes, we aimed to balance the dataset and ensure that the arrhythmia detection model had sufficient representation of the underrepresented classes, thereby optimizing model training to detect arrhythmias across all classes. Both the real and synthesized ECGs were then input to deep learning networks (Section 3.1), as well as an ensemble of these networks, for downstream classification.

We ran our experiments on a laptop computer with the following hardware specifications: Intel Xeon CPU comprising two virtual CPUs; 12GB of DDR4 RAM; and NVIDIA Tesla K80 GPU with 12GB dedicated graphics memory, which was necessary to speed up deep learning tasks and complex data processing. The model was implemented using Python programming within the PyTorch environment, a popular open-source deep learning framework, which allowed for flexible and efficient experimentation. We experimented and obtained the following optimized hyperparameters for training the GAN model: learning rate (lr), 0.0002; batch size, 96; number of epochs, 3000 iterations; optimizer (for both the generator and discriminator); and Adam optimizer with a learning rate of 0.0002 (

Table 3. Hyperparameters for the proposed deep learning models.

Hyperparameter	Values
Model Architecture	SkipCNN	SkipCNN+LSTM	SkipCNN+LSTM+Attention
Input size a	187	187	187
Hidden size b	128	64	64
Kernel size c	5	5	5
Number of classes	5	5	5
Learning rate d	$1\times {10}^{-3}$	$1\times {10}^{-3}$	$1\times {10}^{-3}$
Optimizer e	Adam	Adam	Adam
Batch size	96	96	96
Number of epochs	100	100	100

^a Every input signal has a fixed length of size 187, which is achieved with zero-padding pre-processing. ^b The hidden size, which determines the number of hidden neurons within certain layers of the models, is set to 128 for the SkipCNN model and 64 for both the SkipCNN+LSTM and SkipCNN+LSTM+Attention models. ^c The kernel size, which determines the size of the convolutional kernel used within the convolutional layers of the models, is set to 5 across all architectures. ^d The learning rate, which dictates the step size during optimization, is set at $1\times {10}^{-3}$ (0.001) for gradual parameter updates during training. ^e All three models share the Adam optimizer, an established and effective function for optimizing neural network parameters, which simplifies the training process and also allows for fair model comparisons.

).

4.1. GAN for ECG Beat Generation

We input 2085 and 460 ECG signals of the S and F AAMI classes, respectively, to the GAN model to generate synthetic ECG signals of these two minority classes. The training was conducted over a total of 3000 epochs to refine the GAN model. For the synthesis of ECG beats of the S class, with an increasing number of training epochs, the generated ECG signals converged (Figure 6) and became more similar to the original ECG signals (Figure 7) with demonstrable progressive decremental discriminator training error (

Table 4. Log of the training process for ECG beat generation of the S class using the GAN model. Cumulative training times at fixed intervals of 600 epochs and training errors, expressed as discriminator loss (Loss_D) and generator loss (Loss_G), are shown.

Epoch	Loss_D	Loss_G	Time
0	0.81992757	1.3727783	00:15:16
600	0.32295054	2.08986354	00:23:41
1200	0.34246421	2.03715849	00:32:05
1800	0.41835696	2.19828391	00:40:26
2400	0.31349421	2.24707913	00:48:47

and Figure 8). This signifies the GAN model’s ability to capture the underlying patterns and characteristics of the target S ECG class. Of note, the training time expended for each set of 300 epochs remained relatively constant throughout the training process (Table 4), suggesting that the training and signal generation processes were able to maintain a consistent uniform pace throughout, underscoring the model’s reliability.

In a like manner, for the synthesis of ECG beats of the F class, with an increasing number of training epochs, the generated ECG signals converged (Figure 9) and became more similar to the original ECG signals (Figure 10), with demonstrable progressive decremental discriminator training error (

Table 5. Log of the training process for ECG beat generation of the F class using the GAN model. Cumulative training times at fixed intervals of 600 epochs and training errors, expressed as discriminator loss (Loss_D) and generator loss (Loss_G), are shown.

Epoch	Loss_D	Loss_G	Time
0	1.02077174	1.19444096	01:46:38
600	0.86434889	1.63123429	01:49:03
1200	0.98127455	1.7189554	01:51:26
1800	0.90554142	1.35714304	01:53:51
2400	0.75086892	1.52905619	01:56:18

and Figure 11). This signifies the GAN model’s ability to capture the underlying patterns and characteristics of the target F ECG class. As noted for the S class, the training time expended for each set of 300 epochs remained relatively constant throughout the training process (Table 5), suggesting that the training and signal generation processes were able to maintain a consistent uniform pace throughout, underscoring the model’s reliability for data augmentation across different minority ECG classes.

In the context of generative adversarial networks (GANs), the generator loss and discriminator loss are critical components for training the GAN model. These losses are used to optimize the generator (G) and discriminator (D) networks in a two-player minimax game. Equations (44) and (45) are mathematical representations of these losses:

$$Generator Loss\left(Loss\_G\right)=\mathrm{m}\mathrm{i}\mathrm{n}\left[\log \left(\mathrm{D}\left(\mathrm{x}\right)\right)+\log \left(1-\mathrm{D}\left(\mathrm{G}\left(\mathrm{x}\right)\right)\right)\right]$$

(44)

$$Discriminator Loss\left(Loss\_D\right)=\mathrm{m}\mathrm{a}\mathrm{x}\left[\log \left(\mathrm{D}\left(\mathrm{x}\right)\right)+\log \left(1-\mathrm{D}\left(\mathrm{G}\left(\mathrm{x}\right)\right)\right)\right]$$

(45)

where D(x) signifies the likelihood that x originated from the actual data. G is a differentiable function, depicted as a multilayer perceptron. D and G engage in the ensuing two-player minimax game, involving the value function.

4.2. Similarity Matching of Synthetic ECG Signals

We analyzed the similarity between the synthesized and real ECGs using the MSE, SSIM, and Pearson correlation coefficient, r. For S-class ECG beats, (1) the MSE ranged from 0.0004 to 0.0057, indicating relatively low levels of error in the synthetic signals; (2) the SSIM ranged from 0.9314 to 0.9675, indicating high similarity; and (3) r ranged from 0.8991 to 0.9782, indicating high linear correlation (Figure 12). Similarly, for F-class ECG beats, (1) the MSE ranged from 0.0020 to 0.0136, indicating relatively low levels of error in the synthetic signals; (2) the SSIM ranged from 0.9680 to 0.9742, indicating high similarity; and (3) r ranged from 0.8167 to 0.9727, indicating high linear correlation (Figure 13). These results confirm the proposed GAN model’s ability to generate synthetic signals in minority ECG classes versus real ECGs.

4.3. Dataset Distribution

The GAN was able to synthesize 2400 and 1536 realistic ECGs with high quantifiable similarity matching from 2085 S-class and 460 F-class training ECGs, respectively, yielding an augmented dataset with substantially improved representation of the minority ECG classes for more effective training and an equitable train–test split across all classes (

Table 6. Dataset ECG class distribution before and after GAN augmentation with the train–test split.

Dataset	Distribution	N	S	V	F	Q	Total
Before GAN	Beats, n	90,589	2779	7236	803	8039	109,446
	Beats, %	82.77	2.54	6.61	0.73	7.35	100
	Training, n	76,900	2085	6155	460	6838	92,438
	Testing, n	13,689	694	1081	343	1201	17,008
	Training, %	85	75	85	57	85	84
	Testing, %	15	25	15	43	15	16
After GAN	Beats, n	90,589	5179	7236	2339	8039	113,382
	Generated beats, n	0	2400	0	1536	0	3936
	Beats, %	79.9	4.6	6.4	2.1	7.1	100%
	Training, n	76,900	4485	6155	1996	6838	96,374
	Testing, n	13,689	694	1081	343	1201	17,008
	Training, %	85	87	85	85	85	85
	Testing, %	15	13	15	15	15	15

694 and 343 ECG beats from the S and F classes, respectively, were reserved for testing, while 2085 and 460 beats from the S and F classes, respectively, were used as input to the GAN model to synthesize 2400 S-class and 1536 F-class ECG beats. The synthetic signals were added to the original training set, resulting in an expanded training dataset comprising 96,374 ECG signals. The train–test data split was performed on the total ECG dataset including all the classes, and then the GAN data augmentation was applied to the two minority classes.

). Of note, the number of ECG beats of all classes in the test set (approximately 16% of the total ECG beats in the MIT-BIH database) remained unchanged, which ensured unbiased model testing using only original and unseen ECG data.

4.4. Arrhythmia Detection Using Proposed Models

The expanded 113,382-sample combined original and synthetic dataset was divided into training and test sets comprising 96,374 (85%) and 17,008 (15%) ECG bests, respectively. We conducted prior experiments to optimize the train–test split ratio and found an 85:15 train–test split to yield the most favorable outcomes for arrhythmia detection. To validate our model, we subdivided the 85% training set into a training subset of 90,705 (80%) ECG beats and a validation subset of 5669 ECG beats (5%).

During training spanning 100 epochs with the SkipCNN model, there was a progressive diminution of training and validation losses, which quantify how well the model fits the training data and how well the model generalizes to unseen data, respectively, and a progressive increase in training and validation accuracies and F1 scores (F1 score is the geometric mean of precision and recall) that approximated 100% (Figure 14). Next, the SkipCNN model was tested on the unseen test set, which demonstrated excellent performance for intra-class ECG classification, as well as across all five AAMI ECG classes (Figure 15).

Similarly, during training spanning 100 epochs with the SkipCNN+LSTM model, there was also a progressive diminution of training and validation losses and a progressive increase in training and validation accuracies and F1 scores that approximated 100% (Figure 16). Next, the SkipCNN-LSTM model was tested on the unseen test set, which also demonstrated excellent performance for intra-class ECG classification, as well as across all five AAMI ECG classes (Figure 17). Overall, the performance of SkipCNN-LSTM was similar to that of SkipCNN.

Finally, during training spanning 100 epochs with the SkipCNN+LSTM-Attention model, there was also a progressive diminution of training and validation losses and a progressive increase in training and validation accuracies and F1 scores that approximated 100% (Figure 18).

Compared with the SkipCNN and CkipCNN+LSTM models, loss function curves demonstrated greater convergence, while the training and validation F1 scores were marginally better, suggesting a positive impact of the attention mechanism on training performance. Next, the SkipCNN-LSTM+Attention model was tested on the unseen test set, which also demonstrated excellent performance for intra-class ECG classification, as well as across all five AAMI ECG classes (Figure 19). Overall, the performance of SkipCNN-LSTM+Attention was similar to that of the other models.

Upon the completion of all three experiments, we proceeded to validate the test dataset using an ensemble model. Constructed as a weighted average of the optimized outcomes of all three individual models, the ensemble model yielded excellent classification performance for individual ECG classes as well as across all classes, with average precision, recall, and F1 scores of 99.60% (

Table 7. Confusion matrix with calculated performance metrics of arrhythmia detection using the ensemble model.

Class	TP	FP	FN	TN	Precision	Recall	F1
N	13,495	29	94	390	0.9998	0.9925	0.9961
S	620	73	1	16,314	0.8919	0.9986	0.9422
V	1060	21	18	15,909	0.9806	0.9843	0.9825
F	339	12	4	16,653	0.9658	0.9883	0.9769
Q	1197	8	3	15,700	0.9934	0.9975	0.9954
Average	-	-	-	-	0.9963	0.9963	0.9963
Macro avg.	-	-	-	-	0.9663	0.9922	0.9786
Weighted avg.	-	-	-	-	0.993	0.9925	0.9926

and Figure 20).

5. Discussion

The performance of our proposed models was compared against that of recently published models for arrhythmia detection (

Table 8. Performance analysis of the proposed methodology with state-of-art models for arrhythmia detection, with and without mitigation strategies for data imbalance.

Study	Model	Types of Arrhythmias	Data Balancing	Performance (%)
Rajesh and Dhuli, 2018 [35]	AdaBoost	5	RO, SMOTE, DBB	Acc 99.10 Sen 97.90 Spe 99.40
Oh et al., 2018 [59]	CNN, LSTM	5	---	Acc 98.10 Sen 97.50 Spe 98.70
Gao et al., 2019 [36]	LSTM	8	Focal Loss	Acc 99.26 Rec 99.26 Spe 99.14 Pre 99.30 F1 99.27
Pandey and Janghel, 2019 [30]	CNN	5	SMOTE	Acc 98.30 Pre 86.06 Rec 95.51 F1 89.87
Shaker et al., 2020 [37]	Two-stage deep CNN	15	GANs	Acc 98.00 Pre 93.95
Zhou et al., 2021 [39]	GAN with auxiliary classifier for ECG	5	GAN	Acc 97.00
Fan et al., 2022 [41]	Majority voting	4	MBLS	Acc 99.12
Ma et al., 2022 [42]	Convolutional block attention modules ResNet	5	WGAN-GP	Acc 99.23 Pre 99.13 Sen 97.50 Spe 99.81 F1 98.29
Qin et al., 2022 [43]	Squeeze-and-excitation ResNet1D	5	---	Pre 95.80 Rec 96.75 F1 96.27
Qin et al., 2023 [34]	ECG anomaly detection using GAN	2	GAN	Acc 95.50 Pre 96.90 Rec 91.80 F1 94.30 AUC 95.90
Asadi et al., 2023 [11]	CNN	2	GAN	Acc 99.00
Din et al. 2024 [79]	CNN+CNN–LSTM+Transformer	2	---	Acc 99.56 Pre 99.82 Rec 98.87 F1 99.34
Han et al. 2024 [78]	CNN+LSTM	2	---	Acc 99.60 F1 99.81
Proposed work	CNN	5	GAN	Pre 99.12 Rec 99.12 F1 99.12
	CNN+LSTM			Pre 99.30 Rec 99.30 F1 99.30
	CNN+LSTM+ Attention			Pre 99.29 Rec 99.29 F1 99.29
	Ensemble			Pre 99.60 Rec 99.60 F1 99.60

Acc, accuracy; AUC, area under the curve; CNN, convolutional neural network; ECG, electrocardiogram; F1, F1 score; GAN, generative adversarial network; LSTM, long short-term memory; MBLS, modified broad learning system; MIT-BIH, Massachusetts Institute of Technology–Beth Israel Hospital; Pre, precision; Rec, recall; ResNet, residual network; Sen, sensitivity; SMOTE, synthetic minority oversampling technique; Spe, specificity, WGAN-GP, Wasserstein generative adversarial network with gradient penalty.

). We considered articles that utilized the MIT-BIH database for arrhythmia detection from ECG signals, published between 2018 and 2024. Our models outperformed most of the comparator models, with two notable exceptions. In Ullah et al. [77], the DenseNet-based model attained 99.80% four-class arrhythmia classification accuracy, separately assessed on the MIT-BIH and St. Petersburg Institute of Cardiological Technics ECG datasets. However, the authors did not address data imbalance in their work. In Han et al. [78], a hybrid CNN+LSTM technique was utilized for the detection of cardiac arrhythmias, achieving 99.60% accuracy in classification and a 99.81% F1 score, without using any data balancing techniques.

GAN has been used to augment minority-class ECG data for both binary [34] and multi-class [37,39] arrhythmia detection, as well as in other disease classification models. In addition to GAN for balancing data, other methods have also been utilized, including RO, SMOTE, DBB [35], Focal Loss [36], SMOTE [30], and MBLS [41]. Some studies such as refs. [43,59,78,79], did not utilize any data balancing techniques and still achieved over 95% accuracy. The number of classes for arrhythmia classifications varies across studies, ranging from a maximum of 15 classes [37] to a minimum of 2 classes, with our study utilizing 5 classes according to the AAMI classification.

The ability to synthesize realistic ECGs is a distinct advantage of our ECG generation model that combined GAN with Bi-LSTM. The latter confers the model with the ability to process preceding and subsequent data points efficiently, which enables the GAN to synthesize ECGs with a high degree of matched similarity to real ECGs of the minority classes. By integrating the synthetic ECGs with the ECG training, the data imbalance was ameliorated, and a consistent train–test split was obtained across all ECG classes, which enhanced the model training. As a result, our model yielded superlative classification performance for the individual arrhythmia classes, as well as across all classes, which outperformed many extant models. Among our proposed models, the ensemble model, with 99.60% accuracy, emerged as the top performer; more noteworthy is the fact that all our models consistently attained precision, recall, and F1-score values exceeding 99%, underscoring the accuracy and robustness of our GAN plus SkipCNN-based classification approach to arrhythmia detection.

While our proposed hybrid DL models show promising results for detecting cardiac arrhythmias, it is important to acknowledge some limitations. First, the models were trained and tested on a single dataset, restricting their applicability to additional datasets. Second, more testing with various DL architectures and hyperparameters may be required to determine the best configuration for this task. Third, the suggested models’ complexity, particularly the SkipCNN+LSTM+Attention design, may provide computational and interpretability issues. Fourth, the diversity in ECG recordings from different datasets, as well as the presence of noise or artifacts, may have an impact on these models’ generalization capabilities.

6. Conclusions

In this manuscript, we present the GAN+SkipNet model, which addresses the challenge of detecting cardiac arrhythmias by mitigating data imbalances. To achieve this, we leveraged the MIT-BIH arrhythmia ECG dataset, comprising 109,446 ECG beats categorized into five classes according to the AAMI standard. Our approach involved several key steps. First, we proposed a Bi-LSTM-based GAN model to generate signals for minority classes. Specifically, we focused on generating two types of ECG signals: FVN and PAC signals, which represent just 0.73% and 2.54% of the total dataset, respectively. To assess the quality and similarity of the synthetic signals, we calculated the mean squared error (MSE), structural similarity index (SSIM), and Pearson correlation coefficient (r) scores for both signal types. Subsequently, we combined the synthetically generated dataset with the original dataset, resulting in a total of 113,382 ECG beats. We introduced three innovative models, namely SkipCNN, SkipCNN+LSTM, and SkipCNN+LSTM+Attention, for arrhythmia detection using the combined dataset. Additionally, we enhanced detection accuracy by introducing an ensemble model. Model performance was evaluated using various metrics, including precision, recall, and F1-score, as well as their average, macro average, and weighted average values. The best results for arrhythmia detection were achieved with the SkipCNN+LSTM model (although all other proposed models also provided similar outcomes), which achieved 99.3% precision, recall, and F1-score. Further improvements were made using the ensemble model, pushing the metrics to 99.6%. We validated our proposed model against state-of-the-art techniques from recently published peer-reviewed articles.

There are a few limitations to our work. We performed training and validation with an 80:20 data split instead of standard k-fold cross-validation. The latter represents a potential enhancement that might add rigor and generalizability to our model results. Overall, our work demonstrates a significant step forward in the realm of arrhythmia detection, providing valuable insights into addressing data imbalance and achieving exceptional accuracy.

In the future, we aim to enhance our research by incorporating cross-validation techniques for rigorous model validation and real-time data integration to enable real-time clinical trials. Additionally, we plan to leverage generative AI to provide valuable feedback based on error-prone data, improving the accuracy and reliability of our arrhythmia detection system.