Enhancing Interpretability in Diabetics Prediction: A Comparative Study of SHAP, LIME and Permutation Feature Importance

Diabetes is a prevalent condition with rising global impacts on morbidity and mortality. This paper presents an in-depth analysis of machine learning (ML) models for diabetes prediction. To improve interpretability, the study incorporates multiple Explainable AI (XAI) techniques, including SHAP, LIME, and Permutation Feature Importance, which provide both global and local insights into model predictions. Using multiple XAI methods allows for a comprehensive understanding of model behavior from different perspectives—SHAP offers consistent, mathematically sound feature attributions; LIME provides localized, instance-specific explanations; and Permutation Feature Importance highlights overall feature relevance. Consistently across these XAI methods, Glucose emerged as the most influential predictor, followed by BMI and Age, aligning with established clinical risk factors. Features such as Pregnancies and DiabetesPedigreeFunction exhibited moderate impact, while Insulin and Skin Thickness had minimal effect on predictions. By comparing the advantages and limitations of different XAI methods, this research fosters trust in ML-driven diabetes diagnostics, enabling more transparent and informed decision-making. The study offers a framework for ethical AI integration in clinical practice, advancing responsible AI use in diabetes management.