Multi-label Neural Model for Prediction of Myocardial Infarction Complications with Resampling and Explainability

Abstract

With myocardial infarctions accounting for the largest percent of cardiovascular-related deaths, the need for machine learning tools in prediction and prevention has never been clearer. Specifically, in the case of in-hospital complications following acute myocardial infarction (AMI), even with decreased in-hospital mortality rate due to improved hospital care, patients who survive the acute phase of MI remain at risk for MI-associated complications or recurrent AMI such as bundle branch blocks and angina. In this paper, we propose a multi-label framework to predict the occurrence of 5 complications following admission of 1,700 patients after suffering an AMI episode. We evaluate the models using several multi-label prediction metrics as a test of robustness of our method beating numerous other alternatives and comment on the balance of cost-effectiveness of a compact deep learning model versus shallow machine learning in the multi-label context. Our neural network outperformed 13 other algorithms across all metrics, except Hamming loss. We also implement Shapley value analysis to this multi-label problem and observe interesting behaviour such as the duration of arterial hypertension and time elapsed from the beginning of the attack to the hospital being key predictive features of lethal outcome. This framework presents a novel approach in using multi-label learning, and especially compact cost-effective deep learning, simultaneous for prediction of several AMI complications which has not been explored previously.