Explainable Prediction of Long-Term Glycated Hemoglobin Response Change in Finnish Patients with Type 2 Diabetes...

Area of Science:

• Diabetes management
• Artificial intelligence in healthcare
• Biomarker analysis

Background:

• Glycemic control in type 2 diabetes (T2D) relies on monitoring HbA1c.
• Predicting HbA1c changes after initiating new antidiabetic drugs is crucial for effective treatment.
• Existing prediction models often lack real-world data integration and interpretability.

Purpose of the Study:

• To apply machine learning (ML) and explainable artificial intelligence (XAI) to predict HbA1c level changes within 12 months of starting new antidiabetic drugs in T2D patients.
• To identify key predictors associated with these HbA1c changes.
• To compare ML model performance against randomized controlled trial (RCT) data.

Main Methods:

• Utilized electronic health records (EHR) from 10,139 T2D patients.
• Developed offset models integrating RCT-derived HbA1c changes with real-world EHR data.
• Evaluated various ML models (LR, MLP, RR, RF, XGB) using R² and RMSE, comparing baseline and follow-up data.

Main Results:

• ML models demonstrated superior performance compared to RCT models.
• The follow-up Multi-layer Perceptron (MLP) model achieved higher R² (0.74) and lower RMSE (6.94) than the baseline model (R²: 0.52, RMSE: 9.27).
• Key predictors included baseline HbA1c, post-drug initiation HbA1c, fasting plasma glucose, and HDL cholesterol.

Conclusions:

• ML and XAI enable more realistic and individualized HbA1c change predictions for T2D management.
• XAI enhances model interpretability, providing clinical relevance for tailored treatment strategies.
• Future research will focus on developing treatment selection models leveraging these predictive capabilities.