Rectified Adam Optimizer and LSTM with Attention Mechanism for ECG-Based Multi-class Classification of Cardiac Arrhythmia

Abstract

Cardiac Arrhythmia (CA) is one of the most prevalent cardiac conditions and prime reasons for sudden death. The current CA detection methods face challenges in noise removal, R-peak detection, and low-level feature selection, which can impact diagnostic accuracy and signal stability. The research aims to develop an effective framework for detecting and classifying CA using advanced signal processing, feature extraction, feature selection, and classification for reliable medical diagnosis. The input electrocardiogram (ECG) signals are processed using hybrid noise reduction techniques such as cascaded variable step size normalized least mean square and sparse low-rank filter. The complex and high-level features are extracted using higher-order spectral energy distributed image, wavelet transform, and R-wave peak to R-wave peak interval to enhance the representation of cardiac data. Recursive feature elimination is applied to select the most relevant diagnostic features and the Rectified Adam optimizer is used to fine-tune parameters to achieve better training stability. The model integrates long-term memory with an attention mechanism to enhance the classification performance of arrhythmia detection. Simulation results demonstrate that the proposed model achieves 99.40% accuracy, outperforming existing models and showing its efficiency in classifying CA for better diagnosis and early treatments.