Explainable artificial intelligence for cross domain evaluation of predictive models in multi-disease diagnosis

Abstract Accurate and interpretable disease prediction is one of the major challenges faced in healthcare, especially for breast, heart, and lung cancers. This study proposes a highly structured, leakage-safe benchmarking framework for comparing conventional tabular Machine Learning (ML) models for multi-disease prediction, which is not a new ML model. Six conventional ML models, namely support vector machine (SVM), logistic regression (LR), random forest (RF), XGBoost, decision tree (DT), and k-nearest neighbors (KNN), were evaluated using nested cross-validation for proper model selection and performance on three benchmarking datasets. To improve the interpretability of the models, the authors incorporated Explainable AI (XAI) techniques, namely local interpretable model-agnostic explanations (LIME) for better instance-level interpretability and permutation feature importance (PFI) for better global interpretability. The results indicate high discriminative ability of the models, with random forest and XGBoost models achieving the best classification accuracy. SVM and logistic regression models also achieved the best results for ROC-AUC metric under outer-fold validation. The novelty of the paper is not the architecture of the ML models but the fact that the authors propose a highly structured, leakage-safe preprocessing pipeline, nested validation, statistically sound multi-model comparisons, and robust local–global interpretability aggregation, all of which are incorporated into a single benchmarking template.