Clinical Event Prediction with Machine Learning for Clinical Decision Support―A Case of Pediatric Cardiology―

Abstract

　Predicting adverse clinical events leads to preventing patients from severe condition. In the prediction of clinical events by machine learning, not only accuracy but also usability and explainability in a clinical decision support (CDS) system are crucial because they help to understand and trust the model. This research aims to develop a machine learning model to predict clinical events, to implement a CDS system that provides the prediction results and the explanations, and to evaluate the usefulness of the system. An acute heart failure predictive model was developed with records of 475 patients who were hospitalized for congenital heart diseases from 2015 to 2017. We calculated 65 features with sliding window approach from numeric time-series data extracted from electronic health records (EHR) and constructed a random forest model. The acute heart failure events were predicted at AUC=0.88. We developed a CDS system that connects to EHR and PACS, and used the predictive model to prospectively detect the sign of acute heart failure. The prediction accuracy of the prospective evaluation was AUC=0.76. By evaluating the accuracy, usability and explainability, the usefulness of CDS with a machine learning model was shown.