Predicting Time to Diabetes Diagnosis Using Random Survival Forests

Type 2 Diabetes Mellitus (T2DM) is a chronic metabolic disorder with increasing population incidence. However, T2DM takes years to develop, allowing onset prediction and prevention to be a clinically effective treatment strategy. In this study we propose and assess a novel approach to diabetes prediction which integrates a specialized extension of the random forest algorithm known as random survival forest (RSF). Rather than predicting a binary outcome, this machine learning model incorporates survival analysis methodology to predict the time until a patient will receive a diabetes diagnosis if their current lifestyle is maintained. We trained a baseline model on 7,704 electronic medical records from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN) with 14 biomarker and comorbidity features across different measurement dates. Although tuning parameters were purposefully chosen for quick training rather than for predictive performance, our model exceeded expectations with a concordance index of 0.84. Thus, RSF models have been shown to produce accurate timelines of diabetes onset trajectory, providing patients with quantifiable and relatable risks that are easy to understand. The results of our study have substantial implications for advancing machine learning in clinical decision support and patient outcome predictions, emphasizing the role of innovative models in improving predictive accuracy.