Using (Automated) Machine Learning and Drug Prescription Records to Predict Mortality and Polypharmacy in Older Type 2 Diabetes Mellitus Patients

We analyse a large drug prescription dataset and test the hypothesis that drug prescription data can be used to predict further complications in older patients newly diagnosed with type 2 diabetes mellitus. More specifically, we focus on mortality and polypharmacy prediction. We also examine the balance between interpretability and predictive performance for both prediction tasks, and compare performance of interpretable models with models generated with automated methods. Our results show good predictive performance in the polypharmacy prediction task with AUC of 0.859 (95% CI: 0.857–0.861). On the other hand, we were only able to achieve the average predictive performance for mortality prediction task with AUC of 0.754 (0.747–0.761). It was also shown that adding additional drug related features increased the performance only in the polypharmacy prediction task, while additional information on prescribed drugs did not influence the performance in the mortality prediction. Despite the limited success in mortality prediction, this study demonstrates the added value of the systematic collection and use of Electronic Health Record (EHR) data in solving the problem of polypharmacy related complications in older age.