Adaptive Attention-Driven ALEXIS Framework for High-Precision ECG Signal Classification

Abstract

Accurate classification of electrocardiogram (ECG) signals is essential for the early identification of cardiac abnormalities and for supporting timely clinical decision-making. However, conventional machine learning and deep learning approaches often assume stationary signal characteristics and therefore struggle to model the complex non-linear, non-stationary, and patient-specific nature of ECG signals. These limitations reduce their ability to capture subtle morphological variations and temporal dependencies associated with abnormal cardiac rhythms. To address these challenges, this paper proposes an Adaptive Attention-Driven ALEXIS Framework, a hybrid deep learning architecture that integrates multiscale signal decomposition, hierarchical feature learning, and adaptive attention-guided fusion for five-class ECG abnormality classification. The proposed framework combines the Stockwell Transform (ST) for time-frequency analysis, Empirical Mode Decomposition (EMD) for adaptive extraction of intrinsic oscillatory modes, and Local Phase Quantization (LPQ) for robust morphological feature representation. The extracted multiscale features are processed through an AlexNet-based spatial learning module and an LSTM-based temporal modeling module, followed by an attention-guided feature fusion mechanism and Support Vector Machine (SVM) classifier for robust decision making. The framework was evaluated using the MIT-BIH Arrhythmia Database comprising 109,494 annotated ECG beats, which were partitioned into 80% training (approximately 87,595 beats) and 20% testing (approximately 21,899 beats) using a stratified beat-level split. Experimental evaluation on the independent test set achieved an overall accuracy of 98.43%, sensitivity of 98.27%, specificity of 98.78%, and F1-score of 98.31%. These results demonstrate that the proposed ALEXIS framework effectively captures complementary spectral, temporal, and morphological ECG characteristics, providing accurate and interpretable ECG abnormality classification while establishing a strong foundation for future validation using inter-patient evaluation protocols and cross-database clinical studies