Abstract
Explainable Artificial Intelligence (XAI) is increasingly important in healthcare, where transparent and trustworthy predictive models are essential for supporting clinical decision-making and safe adoption of data-driven systems. Local Interpretable Model-agnostic Explanations (LIME) is a widely used post-hoc, model-agnostic explanation technique to generate instance-level explanations. However, LIME often exhibits limited stability and fidelity, particularly in data-constrained healthcare settings, where small or low-quality training datasets can reduce the reliability of its explanations, undermining clinical trust and real-world interpretability. To address these challenges, we propose a Fast Transferable Local Explanation framework, termed FT-Local Explainer, that enhances the stability and fidelity of local explanations in limited-data target domains by effectively leveraging transferable explanation knowledge from related, data-rich source domains with distributional shifts. To enable controlled and privacy-aware cross-domain information sharing, FT-LocalExplainer accesses only a small set of representative source-domain prototypes during transfer. In addition, a fast leave-one-out cross-validation strategy is introduced to adaptively deter mine the extent of explanation transfer between domains. This strategy provides an approximately unbiased estimation of local fidelity error while enabling efficient learning of the transfer parameter, thereby mitigating the risk of inappropriate transfer. FT-LocalExplainer also incorporates a cost-sensitive mechanism to address class imbalance, helping to maintain explanation quality when the locally sampled neighbourhood is imbalanced. Experiments on real-world healthcare tabular datasets demonstrate that FT LocalExplainer consistently improves explanation quality relative to baseline methods and provides faithful local explanations across diverse healthcare domain-shift scenarios, thereby improving the interpretability of black-box prediction models.
