An ethnic-sensitive hybrid framework for T2D prediction with explainable AI and weighted ensembles

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Public Health

Background:

• Type 2 diabetes (T2D) poses a significant global health challenge, with over 537 million affected in 2021.
• Early T2D prediction is hindered in low- and middle-income countries by data scarcity, class imbalance, and unique population risk factors.

Purpose of the Study:

• To develop and evaluate a novel predictive framework, FW-CAGIN-WCAE, designed to overcome limitations in early T2D detection.
• To enhance prediction accuracy and interpretability, particularly in resource-constrained environments.

Main Methods:

• A four-stage framework: Zero-Threshold Feature Removal (ZTFR), Feature-Weighted Class-Adaptive Generative Imputation Network (FW-CAGIN), Weighted Classifier Aggregation Ensemble (WCAE), and SHAP analysis.
• Utilized nested five-fold cross-validation on PIDD, FHGDD, and BDD datasets and their combinations.

Main Results:

• Achieved a peak Area Under the Curve (AUC) of 0.936 ± 0.018 on the PIDD-BDD dataset combination.
• Significantly reduced imputation Mean Absolute Error (MAE) from 0.8028 to 0.0033.
• Decreased AUC variability by 36.3% and improved the Diagnostic Odds Ratio (DOR) to 68.4 ± 20.5.

Conclusions:

• The FW-CAGIN-WCAE framework provides an accurate, interpretable, and population-sensitive approach for early T2D detection.
• The method is particularly beneficial for resource-limited healthcare settings, offering a robust solution to data challenges.