INTERPRETABLE DEEP LEARNING MODELS FOR PREDICTION OF CLINICAL OUTCOMES FROM ELECTRONIC HEALTH RECORDS

Abstract

The rapid adoption of electronic health records (EHRs) has generated tremendous amounts of valuable clinical data on complex diseases and health trajectories. Yet, achieving successful secondary use of this EHR data for expanding our knowledge about diseases, expediting scientific discoveries in medicine, and facilitating clinical decision-making has remained challenging, owing to the complexity and data quality issues of these EHR data. Artificial intelligence, specifically deep learning, presents a promising approach for analyzing this rich EHR data, represented as a series of timestamped multivariate data packed in irregular intervals. Deep learning-based predictive modeling with longitudinal EHR data offers a great promise for accelerating personalized medicine, enabling disease prevention, better informing clinical decision making, and reducing healthcare costs. However, employing deep learning on EHR data for personalized prediction of clinical outcomes requires coping with numerous issues simultaneously. In this thesis, we focus on addressing three important challenges: data heterogeneity, data irregularity, and model interpretability. We utilize state of the art deep learning techniques and modern machine learning methods to develop accurate and interpretable predictive models using EHR data. Specifically, we demonstrate how temporal clinical data contained in EHRs can be harnessed for providing patient specific predictions and interpretations for several clinical outcomes. We focus on two aspects: 1) code level and visit-level interpretations for predicted outcomes using recurrent neural networks (RNNs), attention mechanism, and contextual decomposition interpretation method, and 2) leveraging the non-stationarity characteristics in EHR data into the predictive models using self-attention mechanism and kernels approximation technique. Our proposed EHR-based deep learning models demonstrate improved performance in terms of predictive accuracy and interpretability on multiple clinical prediction tasks, compared to existing work in this area. These tasks include preterm birth prediction, school-age asthma prediction, and predicting the set of diagnosis codes in the next visit. Such models have a great potential to assist healthcare professionals in making decisions, which are not only dependent on the clinician's clinical knowledge and expertise, but also based on personalized and precise insights about future patient outcomes.