Synergistic Approaches for Accurate Arrhythmia Prediction: A Hybrid AI Model Integrating Higuchi Dimensional Fractal, RR-intervals and Attention-based Convolutional Neural Network in ECG Signal Analysis

In recent years, numerous methods for detecting arrhythmias using a 12-lead ECG have emerged, with deep learning approaches notably demonstrating effectiveness and gaining widespread adoption. However, the classification of inter-patient ECG data for arrhythmia detection remains a significant challenge. Despite the increased utilization of deep learning methodologies, a noticeable gap persists in achieving optimal performance in inter-patient ECG classification. In this paper, we introduce a new method based on a 1D deep learning model that incorporates an attention mechanism into convolutional neural networks for arrhythmia detection. 1D-CNN layers automatically extract morphological characteristics from ECG data, providing an accurate technique for spatial feature extraction. Simultaneously, the attention mechanism enables the model to focus on crucial segments of a signal. To enhance temporal context, four RR-interval features are included, and the potential of the Higuchi Dimensional Fractal is explored as a method for extracting additional features from ECG signals. Consequently, the classification layers benefit from the combination of both temporal and deep features, contributing to the final arrhythmia classification. We validated the proposed method using the MIT-BIH arrhythmia dataset, employing an inter-patient paradigm for model training and validation. Additionally, to assess its generalization ability, we tested it on the INCART dataset. The proposed method attained an average accuracy of 98.75% for three classes and 97.96% for four classes on the MIT-BIH arrhythmia dataset. On the INCART dataset, it achieves an average accuracy of 98.12% for three classes. The experimental results indicate the superiority of this method in comparison to existing methods for recognizing arrhythmias. Thus, our method demonstrates enhanced generalization and potential effectiveness in identifying arrhythmias in real-world datasets characterized by class imbalances, showcasing its practical applicability.