Explainable Meta-Learning Ensemble Framework for Predicting Insulin Dose Adjustments in Diabetic Patients: A...

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Clinical Decision Support
• Diabetes Mellitus Management

Background:

• Diabetes mellitus is a global health challenge requiring precise insulin management to prevent complications.
• Optimizing insulin dosage is complex, with risks of hypoglycemia and hyperglycemia from incorrect dosing.
• Machine learning offers potential for clinical decision support but needs high accuracy and interpretability.

Purpose of the Study:

• To develop and evaluate an explainable machine learning framework for predicting insulin dose adjustments.
• To compare ensemble learning approaches for balancing predictive performance and clinical interpretability.
• To utilize SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) for model transparency.

Main Methods:

• Utilized a dataset of 10,000 patient records with 12 clinical/demographic features.
• Implemented and compared nine machine learning models, including gradient boosting variants and ensemble strategies (Voting, Stacking, Blending, Meta-Learning).
• Evaluated models using accuracy, F1-score, AUC-ROC, PR-AUC, sensitivity, specificity, and cross-entropy loss, with SHAP and LIME for interpretability.

Main Results:

• The Meta-Learning Ensemble achieved superior performance (81.35% accuracy, 0.9637 AUC-ROC, 0.9317 PR-AUC).
• Demonstrated high sensitivity (86.61%) and specificity (91.79%), with perfect sensitivity for dose reduction detection.
• SHAP analysis identified insulin sensitivity, medications, sleep, weight, and BMI as key predictors; LightGBM probability estimates were crucial for the ensemble.

Conclusions:

• The explainable Meta-Learning Ensemble framework effectively predicts insulin dose adjustments with robust interpretability.
• SHAP explanations enhance clinician trust and understanding, supporting informed diabetes management decisions.
• This approach advances the clinical application of AI for personalized insulin therapy.