Bayesian Methods for Multi-Objective Optimization of Hybrid Numerical Filters in ECG Signal Processing for Accurate Arrhythmia Classification

Abstract

This study introduces an innovative method for ECG signal processing that combines advanced filtering techniques, multi-objective Bayesian optimization, and a sophisticated deep learning architecture for classification. The methodology starts with Enhanced Empirical Mode Decomposition (EEMD) to break down the ECG signal into Intrinsic Mode Functions (IMFs). These IMFs undergo filtration through a series of Chebyshev Type II, Butterworth, Daubechies Wavelet, and Savitzky-Golay filters. To achieve optimal performance, a Bayesian multi-objective optimization strategy, augmented by reinforcement learning for dynamic weight adjustment and Gaussian process minimization, is utilized to fine-tune filter parameters. This process ensures maximum noise reduction while maintaining signal integrity. The optimized signals are then processed by an advanced deep learning architecture that includes parallel and residual connections, bidirectional GRU layers, and dense classification layers, enabling precise classification of cardiac conditions. The model's performance was rigorously tested across 12 different ECG leads, showing remarkable improvements in classification accuracy (ACC), sensitivity (SNS), and F1 score. Post-optimization results achieved impressive values of 99.24\% for ACC, 99.04\% for SNS, and 99.05\% for F1 score, demonstrating the significant enhancement in ECG signal analysis and diagnostic reliability provided by the proposed approach.