Knowledge Discovery with Electrocardiography Using Interpretable Deep Neural Networks

Abstract

AbstractDespite the potentials of artificial intelligence (AI) in healthcare, very little work focuses on the extraction of clinical information or knowledge discovery from clinical measurements. Here we propose a novel deep learning model to extract characteristics in electrocardiogram (ECG) and explore its usage in knowledge discovery. Utilising a 12-lead ECG dataset (nECGs= 2,322,513) collected from unique subjects (nSubjects= 1,558,772) in primary care, we performed three independent medical tasks with the proposed model: (i) cardiac abnormality diagnosis, (ii) gender identification, and (iii) hypertension screening. We achieved an area under the curve (AUC) score of 0.998 (95% confidence interval (CI), 0.995-0.999), 0.964 (95% CI, 0.963-0.965), and 0.839 (95% CI, 0.837-0.841) for each task, respectively; We provide interpretation of salient morphologies and further identified key ECG leads that achieve similar performance for the three tasks: (i) AVR and V1 leads (AUC=0.990 (95% CI, 0.982-0.995); (ii) V5 lead (AUC=0.900 (95% CI, 0.899-0.902)); and (iii) V1 lead (AUC=0.816 (95% CI, 0.814-0.818)). Using ECGs, our model not only has demonstrated cardiologist-level accuracy in heart diagnosis with interpretability, but also shows its potentials in facilitating clinical knowledge discovery for gender and hypertension detection which are not readily available.