MSFT: A multi-scale feature-based transformer model for arrhythmia classification

Abstract

Electrocardiogram (ECG) stands as a pivotal non-invasive technique utilized for the diagnosis of heart diseases. Deep learning methodologies are progressively being employed to effectively classify cardiovascular diseases based on ECG signals. Nonetheless, developing a precise and efficacious arrhythmia classification system remains a formidable challenge. In response, this paper proposed the Multi-Scale Feature-based Transformer (MSFT) model for the major types of arrhythmia classification. In the model, we employed a multi-branch convolutional layer structure to extract features that encompass the optimal receptive field. To improve the ability of the model in sensing the location of the time series, we introduced both absolute and relative position encoding methods. Particularly, biased relative position encoding (B-RPE) was innovatively embedded into the probsparse multi-head attention (PS-MHA) structure, allowing the model to balance inference capability and inference speed. For empirical validation, ECG records are drawn from the esteemed MIT-BIH arrhythmia database, a publicly accessible and authoritative source. We discretized the ECG records into fixed-length segments using the continuous-time data streaming segmentation approach. Simultaneously, the overlapping window technique was strategically employed to redress the distribution imbalance within sample categories. The experimental outcomes across various evaluation metrics unequivocally validated the remarkable superiority of our methodology. Specifically, a macro-averaged accuracy of 99.40 %(±0.03 %), precision of 99.40 %(±0.03 %), sensitivity of 99.40 %(±0.03 %), specificity of 99.85 %(±0.01 %), and the F1 score of 99.40 %(±0.03 %). Furthermore, the model introduced in this study presents a significant methodological paradigm that can be extended to analogous bio-signal studies.