CardioPatternFormer: Transformer-Based Interpretable ECG Classification for Multi‐Pathology Detection

Abstract

Background

Interpretation of electrocardiogram (ECG) needs to be critically important to detect cardiovascular diseases early but the traditional machine learning models usually lack long-range temporal dependencies and provide low interpretability. Recently, transformer architectures have gained momentum as one of the powerful options to analyze sequential data, although there is limited literature regarding their use in the area of multi-pathology ECG classification.

Objective

To create CardioPatternFormer a transformer based interpretable ECG classification model that is able to effectively identify various cardiac malady and give clinically meaningful results with the help of attention-based explanations.

Methods

The suggested model uses a multi-head self-attention, duration features encoding, and hierarchical pattern researchers to process raw 1D-ECG gyongy signals. The results were identified in multi-label classification setup to identify arrhythmias, ischemic changes, conduction abnormalities and structural heart conditions. The visualization of attention-map and relevance scoring made it interpretable. CNN, LSTM and hybrid architectures were compared with this model, which was trained and evaluated on big annotated ECG datasets with stratified cross-validation.

Results

CardioPatternFormer gave higher yields with a range of AUC=0.94-0.98 in the key pathologies compared to baseline deep learning models by 4-8%. Heatmap attention showed a positive correlation with ECG pathological segments as identified by the cardiologist and demonstrated interpretability. It was also demonstrated in the model that it was more robust to noisy and irregular ECG signals.

Conclusion

The CardioPatternFormer is a very precise and interpretable ECG classification using a transformer-based system of cardio-patterns to identify multi-pathology. The fact that it can bring physiologically significant waveform details aids the credibility and openness of AI-aided cardiology. The model has good prospects of real time diagnostic assistance and connectivity with the wearable or remote monitoring systems.