An interpretable multi-scale hybrid attention network (MSHAN) for explainable ECG arrhythmia classification.

Artificial Intelligence (AI) holds the promise of transforming patient care in healthcare, but the general adoption of AI in the healthcare sector is impeded by issues that stem from a lack of transparency and trust in AI models. We noted a lack of faith in artificial intelligence (AI) systems working together with cardiologists, which can be considered "black-boxes. This observation led to the idea of incorporating interpretability into the architecture itself, which is achieved by the key innovation of multi-scale feature abstraction and hierarchical attention mechanisms. A novel architecture, namely MSHAN (Multi-Scale Hybrid Attention Network), is proposed to address the critical challenge with an overall accuracy of 96.25% for ECG arrhythmia detection. The MIT-BIH Arrhythmia Database was used for experiments. The proposed model exhibited good discrimination and classification performance for the inter-patient DS2 test set with macro-averaged sensitivity more than 95.34%, macro-averaged precision more than 95.01%, and specificity more than 98.36% for all the classes of arrhythmia, and the discriminative capability was also observed by probability-based ROC analysis, with an average AUC of 0.980. The results of the comparative analysis showed the superiority of the MSHAN method in comparison with traditional methods, including the ResNet-1D (92.14%) and 1D CNN (88.73%), and the explanation of the predictions made by the MSHAN for the inter-patient evaluation mode. We performed an informal clinical observation that the visualizations of the model's decision process, based on SHAP, helped the cardiologists in better understanding and checking the correctness of the model's predictions; formal clinical check and validation are topics that will be explored in future. A wide range of experiments with various architectures show that channel and temporal attention modules combined with parallel convolutional pathways yielded the best performance. Beyond its role in learning to interpret ECGs, the findings in this research provide a blueprint for designing interpretable systems in other time-series applications that involve AI.