Domain-principled Inference with ResNet-Transformer Model for 12-lead ECG Classification

Automated detection of cardiovascular diseases (CVDs) from Electrocardiogram (ECG) recordings is a problem of immense practical interest and it is associated with considerable research challenges. In this paper, we develop an ECG classification model that is capable of satisfying a practical clinical requirement of automated CVD screening solution, where the medical domain principle is to minimize the false negative rates of decisive diagnosis or to improve sensitivity of the critical CVD classes. In this work, we attempt to solve this unique domain principle challenge and propose novel hybrid deep neural model architecture called domain-principled ResNet-Transformer, which consists of two components-ResNet-Transformer network and domain-principled inference. Residual Network or ResNet acts as the feature extractor from the ECG signal to construct lower dimensional feature embeddings, which are fed to the transformer network to capture the patterns in ECG signals and learn the internal information between the embeddings for classification. While the proposed ResNet-Transformer model demonstrates reliable classification performance, the domain-principled inference algorithm ensures that the model is capable of higher sensitivity measures of the critical diagnosis classes, where affinity propagation-based unsupervised learning approach is proposed to maximize the reward of successfully predicting the critical CVDs. We have experimented with Physionet 2020 ECG datasets, one of the largest publicly available 12-lead ECG datasets of its kind and empirical results demonstrate that the classification performance of our ResNet-Transformer model significantly outperforms the current state-of-the-art algorithms. Moreover, the ablation study establishes the unique proposition of our reward-centric inference method to maximize sensitivity towards critical CVD classes.