TOWARD TRANSPARENT AI IN EPIDEMIOLOGY: IMPROVING LIME FOR INFECTIOUS DISEASE APPLICATIONS

Abstract

Artificial intelligence (AI) models are increasingly used in epidemiology for infectious disease prediction, risk stratification, and outbreak surveillance; however, their limited interpretability remains a major barrier to adoption in public-health decision-making. Local Interpretable Model-Agnostic Explanations (LIME) is a widely used technique for explaining black-box predictions, yet its sensitivity to sampling strategies, feature perturbations, and instability across runs restricts its reliability in high-stakes medical contexts. This study proposes an enhanced LIME-based framework tailored for infectious disease modelling that improves explanation stability, feature attribution consistency, and domain relevance. The proposed approach integrates adaptive neighborhood sampling, epidemiologically informed feature grouping, and regularized local surrogate models to produce more coherent and clinically meaningful explanations. Experiments are conducted on multiple infectious disease datasets involving case-incidence prediction and patient-level risk classification, using tree-based and deep-learning models as base predictors. Comparative evaluation against standard LIME demonstrates notable improvements in explanation fidelity, robustness, and alignment with domain expert expectations, while maintaining computational efficiency. Quantitative metrics such as stability indices, faithfulness scores, and perturbation sensitivity are combined with qualitative assessment by public-health professionals to validate interpretability gains. Results indicate that the improved framework provides clearer insight into drivers such as demographic factors, mobility patterns, environmental variables, and clinical indicators, thereby supporting transparent and trustworthy deployment of AI in epidemiological practice. The study contributes to the growing body of explainable AI research by adapting local explanation techniques to the unique requirements of infectious disease modelling and offers a practical tool for enhancing confidence in AI-assisted public-health interventions.