Towards Trustworthy Explainable AI: A Quantitative Framework for Evaluating Model Explanations in Lung Cancer Diagnosis
Year: 2026
Type: article
Abstract: Reliable evaluation of explanation quality is essential before explainable artificial intelligence (XAI) can be safely adopted in clinical decision support. Existing trust scores typically combine multiple criteria into a single composite measure, yet few studies examine whether these scores can distinguish genuine model-dependent explanations from trivial alternatives. In this work, we propose a trust evaluation framework that integrates four established criteria—faithfulness, stability, chance-corrected clinical plausibility, and responsiveness—using a non-compensatory aggregation strategy. To validate the framework itself, we introduce two control explainers: a random explainer and a constant explainer that always returns the same clinically plausible feature set without accessing the predictive model. The framework is evaluated on four black-box classifiers (Deep MLP, XGBoost, LightGBM, and CatBoost) using two independent clinical lung-cancer data sets. Experimental results show that conventional equal-weight aggregation can incorrectly assign higher trust to the constant explainer than to SHAP, whereas the proposed framework consistently assigns zero trust to both control explainers while preserving positive scores for genuine explainers. Additional weight-sensitivity and component-ablation analyses demonstrate that the limitation lies in the compensatory aggregation itself rather than in the selected weights and identify responsiveness as the criterion that prevents model-independent explanations from being considered trustworthy. The framework also abstains from evaluating explanations of models that are statistically indistinguishable from the majority-class baseline, as demonstrated on the second dataset. These findings provide a practical validation strategy for assessing the reliability of trust metrics used to evaluate XAI methods in clinical applications.
Authors: Nouran Elmasry, Asmaa H Ali, Faten Khalifa
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7203663309