Early detection of type 2 diabetes mellitus using machine learning-based prediction models

Area of Science:

• Diabetes Mellitus Research
• Machine Learning in Healthcare
• Predictive Modeling

Background:

• Current type 2 diabetes mellitus (T2DM) screening relies on simplified regression models.
• Growing electronic health data enables development of complex, updateable machine learning prediction models.

Purpose of the Study:

• Compare machine learning (Glmnet, RF, XGBoost, LightGBM) vs. regression models for undiagnosed T2DM prediction.
• Evaluate model performance and stability using fasting plasma glucose prediction over time.

Main Methods:

• Utilized Glmnet, Random Forest (RF), XGBoost, and LightGBM machine learning algorithms.
• Assessed prediction accuracy (RMSE) and variable selection stability with 6-month data batches over 100 bootstrap iterations.

Main Results:

• Simple regression models initially showed the lowest RMSE (0.838).
• Glmnet demonstrated the highest improvement rate (+3.4%) with additional data.
• LightGBM exhibited the greatest variable selection stability over time.

Conclusions:

• No clinically significant performance gain was observed with more sophisticated machine learning models.
• Model interpretability and calibration are crucial, alongside variable stability, for clinical utility.