Type 2 Diabetes Mellitus Classification

We consider the problem of predicting whether a given patient has Type 2 Diabetes Mellitus using his or her electronic health records (EHR), which often lack common indicators of diabetes. Effective heuristics such as FINDRISC exist for detecting undiagnosed diabetes, but require that patients be screened to collect all necessary information. We compare FINDRISC, restricted to the information in our data set, with models of varying complexity. We note that an ensemble of random forest and gradient boosting machine models vastly outperforms FINDRISC. Furthermore, the ensemble’s performance on our data is competitive with FINDRISC’s performance in the field. We conclude that diabetes can be effectively detected using EHR alone. Computer diagnosis can, thus, complement more expensive screening.