Machine learning algorithms mimicking specialists decision making on initial treatment for people with type 2 diabetes mellitus in Japan diabetes data management study (JDDM76)

Objective

To evaluate whether typical machine learning models that mimic specialists’ care can successfully reproduce information, not only on whether to prescribe medications but also which hypoglycemic agents to prescribe as initial treatment for type 2 diabetes.

Research design and methods

A medical records database containing prescriptions for medications for 16,005 patients who visited a diabetologist's office for the first time was utilized to train five typical machine learning models as well-as a model used for logistic analysis. Prescribed were no medications (diet and exercise therapy), insulin, biguanides (BG), sulfonylureas (SU), dipeptidyl peptidase-4 inhibitors (DPP-4I), alpha-glucosidase inhibitors (α-GI) or glinides. Models were compared based on the F1 score and ROC/AUC scores.

Results

XGBoost, which splits decision-making into three sections, was the top performing model (42 % accuracy) among five models and conventional logistic regression (35 % accuracy). The second highest scoring model was Support Vector Machines, which had an accuracy of 40 %. When using XGBoost to compare decisions on no medication needed vs. needing medication the AUC was 0.96. Insulin vs. oral medications had an AUC of 0.78. With all remaining oral medications removed, the AUC was 0.76.

Conclusions

Among the five models investigated, XGBoost outperformed the other machine learning models examined as well as the traditional logistic model, suggesting that its accuracy had the potential to assist non-specialists in decision-making regarding treatment of patients with type 2 diabetes in the future.