Predicting 10-Year Diabetes Risk Through Physiological Acceleration: A Longitudinal Deep Learning Ensemble Approach.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Diabetes Research

Background:

• Type 2 diabetes (T2D) development is a gradual process with a prolonged preclinical phase.
• Traditional static risk scores often fail to capture dynamic metabolic changes.
• Predicting T2D onset requires modeling longitudinal physiological trajectories.

Purpose of the Study:

• To develop a longitudinal deep learning framework for predicting 10-year T2D risk.
• To model physiological acceleration of routine clinical biomarkers for enhanced T2D prediction.
• To improve upon static risk scores by incorporating dynamic metabolic data.

Main Methods:

• Utilized an 18-year longitudinal dataset (Korean Genome and Epidemiology Study - KoGES).
• Constructed a 3D tensor of 21 clinical variables over a 6-year window.
• Developed a stacking ensemble of LSTM and GRU architectures with a logistic regression meta-learner.

Main Results:

• The dynamic framework achieved 0.90 accuracy and 0.94 AUROC on an independent test set.
• The model demonstrated high performance with a Positive Predictive Value (PPV) of 0.97, sensitivity of 0.80, and specificity of 0.98.
• Significantly outperformed a static XGBoost baseline model.

Conclusions:

• The proposed framework offers a highly accurate and resource-efficient tool for T2D screening.
• This dynamic approach can reduce unnecessary clinical alerts and enhance screening efficiency.
• Longitudinal modeling of biomarkers provides superior T2D risk prediction compared to static methods.