INTERPRETABLE MACHINE LEARNING-BASED SCORING SYSTEMS FOR EMERGENCY CARE

Abstract

The emergency department (ED) is usually the starting point of the patient flow through a hospital for urgent care, Risk stratification plays a part in the decision-making in prioritizing patients, treatment, and level of monitoring. For this purpose, an inherently interpretable model that allows doctors to easily understand how the model makes predictions is highly preferred. As a representative, scoring systems have been widely used in clinical settings. This thesis advanced interpretable machine learning-based scoring systems for emergency care. First, I established a ten-year large-scale electronic health records (EHR) database of emergency patients and characterized the workflow of the ED. Second, I invented a new methodology, AutoScore, to automatically generate scoring systems using EHR based on interpretable machine learning. On this basis, a parsimonious and point-based scoring tool, the Score for Emergency Risk Prediction (SERP), was developed for triaging patients at the ED. Then, I extended the AutoScore to survival data and derived the Score for Emergency Readmission Prediction (SERAP) for estimating the time to emergency readmission. Finally, I explored deep learning in EHR with temporal patterns through a systematic review. Overall, this work provides evidence to support intervention by scoring systems to improve emergency care.