Trustworthy AI in Healthcare: A Modality-Aware Evaluation Framework for XAI Fidelity and Stability in Clinical Decision Support

Trustworthy AI in Healthcare: A Modality-Aware Evaluation Framework for XAI Fidelity and Stability in Clinical Decision Support

seyedehSamira Rabinataj1

1) master student of language technologies and digital humanities, university of Turin

IOI Code : ACCA-ACAZFF
Publication : 4th International Congress on Science, engineering & New Technologies(secongress.com)
Abstract :
    Abstract The clinical translation of artificial intelligence (AI) into decision support workflows remains contingent not only on predictive accuracy but on the reliability of the explanations that accompany model outputs. Methods such as SHAP, LIME, and Integrated Gradients (IG) are widely deployed in healthcare AI, yet their fidelity and stability are typically reported on a single dataset within a single clinical task, leaving open the question of whether ranking relationships transfer across the heterogeneous data structures encountered in clinical decision support. This paper proposes the Modality-Aware Evaluation Framework for XAI (MAEF-XAI), a methodological pipeline that jointly evaluates three normalized properties—fidelity, stability, and clinical plausibility—and aggregates them via a documented Trustworthiness Index (TI). The framework is demonstrated on a fully reproducible micro-experiment that contrasts a tabular electronic-health-record-style classification task with a low-resolution medical-image-style classification task. The experiment shows that ranking relationships between explainers do not transfer between task types: on the tabular task, SHAP achieves the highest stability and ground truth plausibility, whereas on the imaging task IG outperforms both SHAP and LIME on stability and plausibility while LIME achieves the highest composite TI. We also show that a small (±0.1) perturbation of the TI weights can change method rankings in the imaging condition, underscoring that aggregation choices must be reported transparently. The paper is positioned as a methodological contribution: MAEF-XAI provides an audit-ready protocol that frames XAI evaluation as a per-task design problem rather than a universal property of the explainer. Keywords: explainable AI, trustworthy AI, clinical decision support, fidelity, stability, SHAP, LIME, Integrated Gradients, evaluation methodology

Keywords : explainable ai trustworthy ai clinical decision support fidelity stability shap lime integrated gradients evaluation methodology