Enhancing Trust in Ensemble Machine Learning Formulations for Diabetes Prognosis: A Comparative Analysis of SHAP and LIME Architectures for Black-Box Model Interpretability

• Hazel Maruq

  Author

Diabetes mellitus remains a critical global health challenge, affecting approximately 537 million adults worldwide, with projections indicating a rise to 783 million by 2045. While ensemble machine learning models have demonstrated superior predictive accuracy for diabetes prognosis, their inherent "black-box" nature presents significant barriers to clinical adoption, where interpretability is essential for building trust among healthcare professionals. This study addresses the research gap by conducting a comparative analysis of SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) architectures within an ensemble learning framework for diabetes prediction. Using the PIMA Indian Diabetes Dataset, we developed a stacking ensemble model combining Random Forest, XGBoost, and Logistic Regression as a meta-model. The ensemble achieved a predictive accuracy of 86.0%, precision of 0.82, recall of 0.80, and an ROC-AUC of 0.91, representing a significant improvement over individual base classifiers. Through systematic comparison of SHAP and LIME explanations, we identified distinct interpretability trade-offs: SHAP provided consistent global feature importance rankings, identifying glucose, BMI, and age as the most influential predictors, while LIME offered more intuitive local explanations for individual patient predictions. The findings demonstrate that SHAP is superior for clinical audits and population-level risk assessment, whereas LIME excels in patient-specific consultations. This study contributes a replicable framework for integrating XAI techniques into ensemble learning pipelines, enabling healthcare practitioners to leverage high-accuracy predictive models with transparent, actionable insights. The implications extend to clinical decision-support systems, policy development for AI adoption in healthcare, and future research on hybrid interpretability approaches tailored to medical diagnostics.

• PDF

Copyright (c) 2026 Hazel Maruq (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.