Abstract
Background: Identifying Arrhythmia for healthcare professionals is critical, considering time and effort and existing struggle with complex spatial and temporal artifacts. The current Machine Learning focuses on accurate classification instead of having a deeper look at the signal origination and cause of disease.
Method: Addressing these issues, this paper presents Bidirectional Arrhythmia Sequence extractor, a unique deep learning model for Electrocardiogram -based arrhythmia classification. The three main parts are: 1) a Squeeze-and-Excitation Temporal Attention Module to model long-range temporal dependencies; 2) a Multi-Receptive Convolutional Module to extract spatial patterns at multiple scales; and 3) an Adaptive Class-Balanced Loss to minimize class imbalance.
Result: The combination of using Multi-Receptive Convolutional Module and Squeeze-and-Excitation Temporal Attention Module helps the classification and identification of these electrocardiogram signals considering both the spatial and temporal factors and also use of Adaptive Class-Balanced Loss to dynamically adjusts class weights during training to emphasize underrepresented arrhythmia types.
Conclusion: The proposed Bidirectional Arrhythmia Sequence extractor architecture advances electrocardiogram arrhythmia classification by learning discriminative spatio-temporal representations while handling data challenges. Bidirectional Arrhythmia Sequence extractor can improve clinical decision support and heart disease diagnosis.