Performance-Interpretability Trade-off in Machine Learning Models for Diabetes Prediction: A Comparative Study Using Explainable AI

Diabetes is a rapidly growing public health concern that require early detection to prevent serious complications. In recent years machine learning models have shown strong potential in predicting diabetes using clinical data. However, a key challenge is the trade-off between predicting performance and interpretability. While complex models such as random forest and XGBoost achieve high accuracy but they often lack of transparency. Simpler models like logistic regression, decision tree provides better interpretability but may compromise performance. In this study a comparative analysis of multiple machine learning models for diabetes prediction by evaluating both performance and interpretability. The Pima India diabetes dataset is used and models are evaluated using metrics such as accuracy, precision, recall, F1 score and ROC-AUC. In addition of this interpretability is analyzed using both intrinsic (ante-hoc) methods and post-hoc explainable AI techniques including Shap and Lime. The results show that ensemble learning models outperform interpretable models in term of predictive performance whereas interpretable models provide greater transparency. Explainable AI techniques help in filling this gap by providing meaningful insights into complex model predictions. The study highlights the importance of balancing accuracy and interpretability in healthcare applications and support the use of XAI to enhance trust and usability of machine learning systems.