Explainable Meta-Learning Ensemble Framework for Predicting Insulin Dose Adjustments in Diabetic Patients: A Comparative Machine Learning Approach with SHAP-Based Clinical Interpretability.

Background and

Objectives

Diabetes mellitus represents one of the most prevalent chronic metabolic disorders worldwide, necessitating precise insulin dose management to prevent both acute and long-term complications. The optimization of insulin dosing remains a significant clinical challenge, as inappropriate dosing can lead to hypoglycemia or hyperglycemia, each carrying substantial morbidity risks. Machine learning approaches have emerged as promising tools for developing clinical decision support systems; however, their practical implementation requires both high predictive accuracy and model interpretability. This study aimed to develop and evaluate an explainable machine learning framework for predicting insulin dose adjustments in diabetic patients. We sought to compare multiple ensemble learning approaches and identify the optimal model configuration that balances predictive performance with clinical interpretability through comprehensive SHAP and LIME analyses. Materials and

Methods

A comprehensive dataset comprising 10,000 patient records with 12 clinical and demographic features was utilized. We implemented and compared nine machine learning models, including gradient boosting variants (XGBoost, LightGBM, CatBoost, GradientBoosting), AdaBoost, and four ensemble strategies (Voting, Stacking, Blending, and Meta-Learning). Model interpretability was achieved through SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) analyses. Performance was evaluated using accuracy, weighted F1-score, area under the receiver operating characteristic curve (AUC-ROC), precision-recall AUC (PR-AUC), sensitivity, specificity, and cross-entropy loss.

Results

The Meta-Learning Ensemble achieved superior performance across all evaluation metrics, attaining an accuracy of 81.35%, weighted F1-score of 0.8121, macro-averaged AUC-ROC of 0.9637, and PR-AUC of 0.9317. The model demonstrated exceptional sensitivity (86.61%) and specificity (91.79%), with particularly high performance in detecting dose reduction requirements (100% sensitivity for the 'down' class). SHAP analysis revealed insulin sensitivity, previous medications, sleep hours, weight, and body mass index as the most influential predictors across different insulin adjustment categories. The meta-model feature importance analysis indicated that LightGBM probability estimates contributed most significantly to the ensemble predictions.

Conclusions

The proposed explainable Meta-Learning Ensemble framework demonstrates robust predictive capability for insulin dose adjustment recommendations while maintaining clinical interpretability. The integration of SHAP-based explanations facilitates clinician understanding of model predictions, supporting transparent and informed decision-making in diabetes management. This approach represents a significant advancement toward the clinical implementation of artificial intelligence in personalized insulin therapy.