Explainable Prediction of Long-Term Glycated Hemoglobin Response Change in Finnish Patients with Type 2 Diabetes Following Drug Initiation Using Evidence-Based Machine Learning Approaches

Gunjan Chandra,1,* Piia Lavikainen,2,* Pekka Siirtola,1 Satu Tamminen,1 Anusha Ihalapathirana,1 Tiina Laatikainen,3– 5 Janne Martikainen,2 Juha Röning1 1Biomimetics and Intelligent Systems Group, Faculty of Information Technology and Electrical Engineering, University of Oulu, Oulu, Finland; 2School of Pharmacy, University of Eastern Finland, Kuopio, Finland; 3North Karelia Wellbeing Services County (Siun Sote), Joensuu, Finland; 4Department of Public Health and Social Welfare, Finnish Institute for Health and Welfare, Helsinki, Finland; 5Institute of Public Health and Clinical Nutrition, University of Eastern Finland, Kuopio, Finland*These authors contributed equally to this workCorrespondence: Gunjan Chandra, Biomimetics and Intelligent Systems Group, Faculty of ITEE, University of Oulu, Oulu, Finland, Email [email protected]: This study applied machine learning (ML) and explainable artificial intelligence (XAI) to predict changes in HbA1c levels, a critical biomarker for monitoring glycemic control, within 12 months of initiating a new antidiabetic drug in patients diagnosed with type 2 diabetes. It also aimed to identify the predictors associated with these changes.Patients and Methods: Electronic health records (EHR) from 10,139 type 2 diabetes patients in North Karelia, Finland, were used to train models integrating randomized controlled trial (RCT)-derived HbA1c change values as predictors, creating offset models that integrate RCT insights with real-world data. Various ML models—including linear regression (LR), multi-layer perceptron (MLP), ridge regression (RR), random forest (RF), and XGBoost (XGB)—were evaluated using R² and RMSE metrics. Baseline models used data at or before drug initiation, while follow-up models included the first post-drug HbA1c measurement, improving performance by incorporating dynamic patient data. Model performance was also compared to expected HbA1c changes from clinical trials.Results: Results showed that ML models outperform RCT model, while LR, MLP, and RR models had comparable performance, RF and XGB models exhibited overfitting. The follow-up MLP model outperformed the baseline MLP model, with higher R² scores (0.74, 0.65) and lower RMSE values (6.94, 7.62), compared to the baseline model (R²: 0.52, 0.54; RMSE: 9.27, 9.50). Key predictors of HbA1c change included baseline and post-drug initiation HbA1c values, fasting plasma glucose, and HDL cholesterol.Conclusion: Using EHR and ML models allows for the development of more realistic and individualized predictions of HbA1c changes, accounting for more diverse patient populations and their heterogeneous nature, offering more tailored and effective treatment strategies for managing T2D. The use of XAI provided insights into the influence of specific predictors, enhancing model interpretability and clinical relevance. Future research will explore treatment selection models.Keywords: type 2 diabetes, HbA1c, treatment effect estimation, machine learning, SHAP