Using Machine Learning Techniques to Identify Key Risk Factors for Diabetes and Undiagnosed Diabetes

Abstract: This paper reviews a wide selection of machine learning models built to predict both the presence of diabetes and the presence of undiagnosed diabetes using eight years of National Health and Nutrition Examination Survey (NHANES) data. Models are tuned and compared via their Brier Scores. The most relevant variables of the best performing models are then compared. A Support Vector Machine with a linear kernel performed best for predicting diabetes, returning a Brier score of 0.0654 and an AUROC of 0.9235 on the test set. An elastic net regression performed best for predicting undiagnosed diabetes with a Brier score of 0.0294 and an AUROC of 0.9439 on the test set. Similar features appear prominently in the models for both sets of models. Blood osmolality, family history, the prevalance of various compounds, and hypertension are key indicators for all diabetes risk. For undiagnosed diabetes in particular, there are ethnicity or genetic components which arise as strong correlates as well.