ROBUSTNESS ANALYSIS OF EXPLAINABLE ARTIFICIAL INTELLIGENCE METHODS FOR MALARIA PREDICTION UNDER DATA PERTURBATION

Abstract

Explainable artificial intelligence (XAI) methods such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model agnostic Explanations (LIME) are increasingly deployed in health surveillance systems to provide transparency in machine learning predictions. However, concerns persist regarding the reliability of these explanations under imperfect data conditions commonly encountered in resource limited settings. This study investigates the robustness of SHAP and LIME explanations for malaria test positivity rate prediction under systematic data perturbations. Three gradient boosting models (XGBoost, LightGBM, CatBoost) were trained on malaria surveillance data from Bayelsa State, Nigeria comprising 2,100 records across eight local government areas. Model explanations were evaluated under controlled perturbations including Gaussian noise (5 to 100 percent), missing value injection (5 to 50 percent), and feature corruption (5 to 50 percent). Stability was quantified using Spearman rank correlation and top k feature overlap metrics. Results demonstrate exceptional robustness of SHAP explanations, with mean Spearman correlation coefficients of 0.976 for XGBoost, 0.981 for LightGBM, and 0.982 for CatBoost. SHAP consistently outperformed LIME across all conditions. The top five most important features remained consistent across most perturbation scenarios with 100 percent overlap for XGBoost and CatBoost SHAP. These findings provide confidence for deploying XAI based decision support systems in malaria surveillance programs where data quality may be suboptimal.