An irregular heartbeat, known as arrhythmia, occurs when the heart's rhythm is disrupted, causing it to beat erratically, excessively quickly, or unusually slowly, distracting its ability to pump blood efficiently. It can be analysed with an Electrocardiogram (ECG), which records the heart's electrical activity without invasive procedures. Traditionally, diagnosing arrhythmia relied on human observation, which was time-consuming. However, automatic detection and classification now save time and improve accuracy. Over the past few decades, cardiovascular disease diagnosis has increasingly relied on machine learning (ML) and deep learning (DL) techniques. Outmoded ML requires extensive feature extraction from ECG signals, while DL models simplify this process by automatically identifying relevant features for prediction and classification 1. Yanfang Dong et al.2 proposed CNN-DVIT, the new architecture for deep learning classification of multi-label arrhythmias using 12-lead ECG signals containing different-length recordings. This model combines CNNs with depth wise separable convolutions and incorporates a ViT with a deformable attention mechanism to extract spatial and temporal characteristics from ECG data efficiently. Sattar et al.3 developed a deep learning-based approach for ECG classification by digitising ECG images to time-series data and applying models such as CNN, LSTM, and SSL-based autoencoders. The CNN model achieves the highest accuracy for classifying cardiac arrhythmias. The results have proven that the digitised ECG signals allow such accurate real-time monitoring for cardiologists, offering an essential instrument for the early and efficient diagnosis of cardiac diseases. Answer: Ansari et al.4 review deep learning (DL) architectures for ECG arrhythmia detection, comparing models like CNNs, MLPs, Transformers, and RNNs used from 2017 to 2023. It offers a roadmap for researchers entering the field, providing insights into current trends and practical models for detecting ECG anomalies. The survey also highlights areas for future research, aiming to inspire further advancements in ECG arrhythmia detection and classification.

ECG is the most widely employed and pertinent method for evaluating a patient's cardiac activity. Each cardiac phase embraces successive atrial and ventricular depolarisation 5, emanating from the atrial sinoatrial node and spreading across the heart. This process produces electrical currents on the body's surface, inducing skin surface electrical potential variations. Surface electrodes capture these signals, which are visually depicted in the ECG. A standard ECG cycle's key attributes include amplitudes, morphologies, and durations of waves such as P, QRS, and T waves 3, as illustrated in Figure 1. A simple ECG wave starts with a P wave. The QRS wave, notably, is the most significant waveform in the ECG signal and represents the depolarisation of the ventricles. T waves represent ventricular repolarisation.

Baseline wander, Power Line Interference, contact noise, and motion artefacts represent diverse categories of noise found in ECG signals 6. M Wu et al.7, 8 proposed a resilient and effective deep 1D CNN with 12 layers to categorise five arrhythmia classes. X Hua et al.9 presented a comprehensive ECG signal classification technique that leverages a unique 1D CNN segmentation strategy. MMR Khan et al.10 presented a 1-D CNN based on DL for the automated categorisation of ECG heartbeats, aiming to classify five distinct types of abnormal rhythms. BM Maweu et al.11 presented the CNN Explanations Proposed model for ECG Signals (CEFEs), designed to provide interpretable explanations. Y Zhao et al.12 introduced an automated technique for ECG signal classification exploiting a deep CNN, complemented by wavelet transform for data filtering. Savalia S and Emamian V 13 introduced a computerised classification system employing the Multi-Layer Perceptron (MLP) network and the CNN.

 Despite their effectiveness, DL-based methods face specific challenges 14. They often require intricate convolution and recursive structures, producing a series of concealed states with limited parallelisation due to dependencies on previous states. While CNNs have become famous for pattern classification, they may not capture pathological variations in ECG signals, such as irregularities in shape, duration, or timing 15. Transformer networks present a promising alternative, offering advantages in handling sequential data and extracting relevant features to classify cardiac abnormalities better.

Key contributions of the paper include:

(a) In pre-processing, denoising, signal segmentation, and data resampling are meticulously executed.

(b) The proposed model employs transformer network (TransNet) based CNN for feature mining.

(c) A comparative study is performed against the standard CNN architecture to evaluate the model's effectiveness.

ECG monitoring generates a continuous flow of data, yielding large amounts of complicated waveforms that cannot be analysed manually, especially in real-time. The volume becomes even larger in large-scale clinical environments where numerous patients are monitored simultaneously. Arrhythmia can take such diverse forms and severities. Some are subtle and not easily detectable and require specialists' expert analyses for pattern recognition. Manual interpretation is prone to error, particularly regarding infrequent or unusual arrhythmias. This analysis takes more time and requires great attention to detail, which would be impractical for a clinical setting that needs prompt diagnoses to intervene effectively. It is hard to find many professionals in clinical settings whose qualifications warrant extensive training and experience for proper ECG interpretation. In high-volume settings, the absence of available experts might reduce the capability to keep pace with demand. Automated techniques, like Transformer-based Convolutional Neural Networks, help address these challenges through efficient, scalable, and accurate tools for classifying arrhythmias. Specifically, Transformer-based CNNs apply deep learning techniques to capture temporal and spatial features in ECG data, allowing for improved accuracy and speed up, which is critical in vast, real-time applications.

The proposed methodology integrates TransNet with the CNN network to address its limitations, particularly its suboptimal performance in handling temporal features. The paper's structure is outlined as follows: The proposed model is detailed in Section 2; preliminary findings are presented in Section 3; and a concise overview and concluding remarks are provided in Section 4.

The proposed model is represented in Figure 2. In Figure 2 ECG signals from MIT-ArrhyDB are given as input to the model. Due to undesired random instabilities, the signal is exposed to eradicate noise as part of the pre-processing. After the signal has been de-noised, R-peak information is accessible from the database, which separates the individual heartbeats. Next, resampling has been used to enrich the data, and lastly, the classifier architecture receives the processed and expanded data. The following subsections include a full description of the complete procedure.

In Figure 2, the raw data is pre-processed to eliminate unwanted noise or artefacts of the ECG signal, improving the signal quality. Then, we extract all the relevant segments that might correspond to individual heartbeats or specific intervals, after which we bring the data into a uniform sampling rate for consistency, which is valuable for further processing. Filter count and kernel dimension with values of 32, 64, and 128 and dimensions of 1.5, 1.7, and 1.9, respectively, are used within each layer for identifying the different patterns appearing in ECG signals. After extracting the features by CNN, these are fed into a Transformer Network, which, using attention mechanisms, helps to understand complex relationships between data points. This approach also accounts for sequential relationships among heartbeats in ECGs, making it suitable for analysing ECG data, which is inherently time-dependent. This output from the Transformer Network is given a softmax function to provide different classes' probabilities. The pipeline implements CNN for feature extraction and Transformer networks for sequential learning of ECG data to classify it into different heartbeat categories. The approach efficiently suits the appropriately complex temporal patterns of ECG signals and will potentially contribute to automated arrhythmia detection systems.

This architecture uses 109,446 ECG signal segments from the MIT-ArrhyDB to evaluate model performance. The dataset is divided 80:20 for learning and evaluation, with 87,554 beats for training and 21,892 for assessment. The model classifies five categories of CVDs—N, S, V, F, and Q achieving an overall accuracy of 98.95% and an F1-score of 98.52%. The training was conducted for 50 epochs. A gradual increase in training accuracy was observed up to the 40th epoch, after which it reached a saturation point, maintaining stability from the 40th to the 50th epoch. Therefore, training was halted at 50 epochs. The Adam optimiser with a learning rate 0.001 was utilised to avoid overfitting. Despite some misclassifications in the normal class, these are minor relative to the total number of beats tested. Although S and F beats are less frequent, the model achieves 83% and 86% accuracy for these classes, respectively. In contrast, a reference model [20] using a transformer for ECG classification reported an accuracy of 90.52%. The commonly employed metrics, namely Accuracy, Precision, Recall, and F1-score (represented in , , , , respectively), have been utilised to evaluate the proposed classification model quantitatively. True positives and true negatives are denoted as TP and TN, while false positives and false negatives are represented as FP and FN.

Each category is evaluated based on accuracy, precision, recall, and F1-score. The model achieves high accuracy across all categories, with N and Q at 99.00%, V at 95.00%, S at 83.00%, and F at 86.00%. Precision and recall values closely match accuracy, indicating consistent performance in correctly identifying each category. F1-scores, which balance precision and recall, range from 83.00% to 99.00%, highlighting the model's effectiveness in classifying ECG signals across diverse cardiac conditions.

To analyse the performance of the proposed model, it is also trained using a convolutional neural network, which gives an overall accuracy of 96.82%. When the transformer block is incorporated with CNN, the overall accuracy is enhanced to 98.95%. Table 1 shows the comparison of the transformer-based CNN model with previous architectures. Figure 4, Figure 5 represent training and validation accuracy and training and validation loss, respectively. Figure 6, Figure 7 illustrate the confusion matrices for the model employing CNN-Transformer and CNN, respectively. The application of CNN alone results in an F1-score of 96.31% in the proposed model. Incorporating the transformer into the CNN leads to a notable improvement, with the F1-score improving by 2.21%.

Figure 4 compares the training and validation accuracy of two models over 50 epochs. Figure 4 (a) shows the performance of a CNN, where training accuracy stabilises close to 1.0, but validation accuracy fluctuates and stabilises around 0.975, showing potential overfitting. Figure 4 (b) depicts the performance of a combined CNN and Transformer model, with both training and validation accuracies increasing rapidly and stabilising close to 1.0, representing better generalisation and reliable performance. The enclosure of transformers augments the model's capability to capture complex patterns, leading to superior overall performance compared to CNN alone. Figure 5 (a) shows that the training loss decreases rapidly and stabilises near zero, while the validation loss also decreases but remains higher and fluctuates more, suggesting potential overfitting. Figure 5 (a) shows that the training and validation losses decrease rapidly and closely follow each other, stabilising near zero, indicating that the TransNet-based CNN model achieves better generalisation and more stable performance on both training and validation data.

As shown in Figure 6, Figure 7, the CNN with TransNet performs better than the basic CNN. The proposed model demonstrates an accuracy of 86% for class S and 83% for class F, despite these classes having a smaller quantity of data.

Table 1 showcases the superior performance of the proposed TransNet-based CNN model, which achieves an impressive accuracy of 98.95% and an F1 score of 98.52%. The table demonstrates that the proposed architecture significantly outperforms other compared models, particularly excelling in accurately classifying low-quantity signals such as Supraventricular and Fusion beats.

Implementing an automated Transformer-based CNN system for classifying arrhythmia in a real-world clinical environment presents several potential challenges. Hospitals and clinics use many legacy systems that cannot support advanced deep-learning models; thus, much modification and integration may be necessary for seamless data transfer and processing. Integrations with electronic health records and other medical databases are technically complicated and resource-demanding processes. Healthcare data is most sensitive, and implementing automated systems requires strict adherence to privacy regulations in the healthcare sector, such as HIPAA in the U.S. or GDPR in Europe. Secure data handling, encryption, and anonymisation protocols will protect privacy but add some layers of complexity to the deployment of such a system. Clinical environments would particularly demand near real-time if not real-time, processing of ECG data for expedited patient diagnosis and treatment. Transformer-based models are powerful but could also be computationally expensive, posing significant challenges in achieving adequate speed and efficiency without high-performance hardware. ECG data differ considerably from patient to patient due to age, physical condition, and sensor quality. Noises and artefacts from patient movement, electrode placement, or other environmental reasons can affect signal quality and make accurate classification difficult. The system built should ensure robustness against such variability to limit misdiagnosis. Advanced deep learning models integrated and maintained within healthcare require special hardware, maintenance, and technical support for proper functionality. These costs can be highly prohibitive for smaller clinics or facilities with limited budgets, slowing adoption across diverse clinical environments.

This research proposed an automated categorisation model that combines a CNN and TransNet to categorise ECG signals. Before being fed into the CNN, the ECG signal undergoes pre-processing procedures, including denoising, segmentation, and resampling. The parameter information extracted by the CNN retains temporal characteristics. Through the synergistic combination of the CNN and an enhanced transformer, the proposed model achieves a categorisation accuracy of 98.95%. The suggested model may be used because of its performance and the fact that transformers were included to address the complexity issues with sequential models. This model can be used with wearable technologies in future research to help save more lives in emergencies by providing ongoing monitoring. Future scopes in this direction are promisingly developing Transformer-based CNNs for arrhythmia classification in clinical settings, which extend to all constituencies of healthcare domains. Further work could incorporate these models with wearable devices for health monitoring, such as smartwatches and mobile ECG monitors. This would enable seamless, in situ, immediate arrhythmia detection outside clinics, allowing for remote monitoring and early interventions, especially among high-risk or elderly patients. Automated systems can permit clinicians to monitor and manage patients with arrhythmias who reside in areas inaccessible by specialists. To combine it with telemedicine, future systems can be designed to integrate with existing platforms for reviewing and diagnosing patients by remote specialists, which is highly valuable in managing chronic cardiac conditions.