566-P: Development and Evaluation of Pharmacotherapy Decision Support Technology for Type 2 Diabetes Mellitus

Background and Objectives: The established treatment target for most patients with type 2 diabetes mellitus (T2DM) is controlling HbA1c levels to less than 7%; however, about half of patients fail to achieve this target. In order to support optimal prescription selection, we have been developing a prescription selection support system that predicts expected outcomes by treatment options. In this presentation, we report a novel analytical method named Treatment Pathway Based Estimation (TPGE) , which predicts the probability of achieving a treatment goal, such as HbA1c less than 7% in 3 months, using a treatment pathway graph built from the data. Methods: The TPGE model was built and evaluated using EHR data from 27,9 patients with T2DM. TPGE was compared with a conventional machine learning approach, gradient boosting decision tree (GBT) , across two dimensions. First, the prediction error of the probability of achieving treatment was evaluated by the Brier Score. Second, the characteristics of the prescriptions recommended with the highest probability were investigated, referring to the prescriptions actually prescribed by clinicians. Results: TPGE achieved a prediction error of 0.158, which is lower than the prediction error of 0.162 of GBT. In addition, the calibration error was 0.002 (84.6% reduction) , much lower than that of GBT (0.013) . A prescription to add two or more drugs was selected by clinicians for 0.7% of cases. GBT recommended that prescription for 42.2 % of cases, while TPGE recommended for 1.4%, which was closer to the clinicians’ decision. In summary, the superiority of TPGE over GBT was confirmed. Conclusion: We have developed a pharmacotherapy decision support technology for T2DM to help clinicians and patients select the optimal prescription.