Novel WT-MoE-SSM-EchoFormer-Based Hybrid Framework for Intelligent ECG Analysis and Cardiac Disease Prediction

Resumen

Cardiovascular Disease (CVD) continues to be a major global health concern, posing significant risks to human life. Early and accurate detection remains difficult due to noise in Electrocardiograms (ECGs) signals, class imbalance, and the limited generalisation ability of existing Machine Learning (ML) models. Therefore, this work proposes a new hybrid optimized Deep Learning (DL) techniques for the intelligent analysis of ECG signals for the prediction of cardiac disease. The study uses raw ECG Signal from the MIT-BIH Arrhythmia Database as a case study, and applies Savitzky-Golay (SG) filters for denoising to ensure proper noise elimination while maintaining the necessary morphological structures. Exploratory Data Analysis (EDA) is utilized for statistical descriptors such as mean, variance, skewness and kurtosis of RR intervals and QRS widths, complemented by correlation matrix and QRS morphology overlays to uncover temporal patterns. Pan–Tompkins’s algorithm applied for extracting morphological features, such as RR intervals, QRS duration, Heart Rate, and Rhythm Variability enabling precise identification of P, QRS and T complexes. The new proposed framework, Wavelet-Token Swin-1D Mixture-of-Experts State-Space EchoFormer (WT-MoE-SSM-EchoFormer) uses wavelet domain tokenization for feature encoding, Swin 1D hierarchical attention and expert driven selection learning for better represent cardiac signals through state space temporal modelling. The Golden Tortoise Beetle Optimizer (GTBO) algorithm employed for dynamically tuning hyperparameters, thus improving convergence and training stability. The complete framework is implemented in Python and yielded an accuracy of 99.98%, a precision of 99.79%, a recall of 94.84%, an F1-score of 97.03%, a specificity of 99.99%, and a ROC-AUC of 1.0000, which validates the proposed framework's effectiveness and reliability for ECG-based cardiac disease diagnosis. The proposed framework achieves superior classification accuracy, lower misclassification, and greater potential for real-time cardiac disease diagnosis.