XAI-Powered Comparative Diabetes Prediction with XGBoost, SVM and Artificial Neural Networks

In this study, diabetes prediction was conducted using XGBoost, Support Vector Machines (SVM), and a Multilayer Artificial Neural Network (ANN). Attributes that cannot be biologically ‘0’ (e.g. glucose, BMI) were first marked as NaN. After filling the missing values with the median, the data were transformed with StandardScaler. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied, and feature selection was performed using the VarianceThreshold method. Hyperparameter optimization for all models was conducted using the RandomizedSearchCV approach.Model performance was evaluated based on ROC-AUC (Receiver Operating Characteristic – Area Under Curve), accuracy, and F1 score. The best results were achieved by the ANN model (ROC-AUC: 0.829, accuracy: 0.76, F1 score: 0.678), followed by XGBoost (ROC-AUC: 0.820, accuracy: 0.734, F1 score: 0.667), and SVM (ROC-AUC: 0.809, accuracy: 0.734, F1 score: 0.661).As part of the Explainable Artificial Intelligence (XAI) framework, global SHAP (SHapley Additive exPlanations) analysis was applied to the XGBoost model, while local explanations were generated using LIME (Local Interpretable Model-Agnostic Explanations) for the SVM and ANN models. SHAP analysis identified glucose, body mass index (BMI), and age as the most influential features. LIME provided case-specific interpretations by highlighting the contribution of individual features for each patient. These approaches contributed to enhanced transparency in clinical decision-making processes.