Clinical interpretation of machine learning models for prediction of diabetic complications using electronic health records

Abstract

AbstractType 2 diabetes is a massive public health issue that continues to grow. The rate of diabetic complication progression varies across individuals. Understanding factors that alter the rate of complication onset may uncover new clinical interventions and help prioritize individuals for more aggressive management. Here, we explore how various machine learning models and types of electronic health records can predict fast versus slow diabetic complication onset using only patient data prior to diabetes diagnosis. We find that optimized random forests generally perform best among the tested models and combining all data sources yields the best predictive performance. A key differentiator of our study is our model interpretation, which identifies specific patient metrics from each dataset that play a unique role in the progression of each complication. Overall, our clinical interpretation of machine learning models can identify patients at risk for poorer outcomes years in advance of their diabetic complication.