Explainable AI Frameworks for High-Stakes Applications

Abstract

As machine learning models are increasingly deployed in high-stakes domains -- clinical diagnosis, criminal sentencing, credit scoring, autonomous driving -- the demand for explanations of model behaviour has moved from academic interest to regulatory requirement. This study presents a controlled empirical evaluation of seven explanation methods -- SHAP (KernelSHAP and TreeSHAP), LIME, Integrated Gradients, Grad-CAM, counterfactual explanations, and concept-based explanations (TCAV) -- across four high-stakes application domains: medical image classification (chest X-ray pathology detection), recidivism risk prediction (COMPAS-equivalent tabular data), loan default prediction (Home Credit dataset), and autonomous driving decision explanation (nuScenes). Each method was evaluated on five dimensions: faithfulness (does the explanation accurately reflect model reasoning?), stability (do similar inputs produce similar explanations?), comprehensibility (can domain experts understand and act on the explanation?), computational cost, and regulatory compliance alignment with the EU AI Act transparency requirements. A total of 840 evaluation trials including 168 structured interviews with domain experts (42 clinicians, 42 legal professionals, 42 loan officers, 42 safety engineers) were conducted. SHAP achieved the highest faithfulness scores across all domains (mean faithfulness = 0.82 +- 0.06) but the lowest comprehensibility ratings from non-technical experts (mean = 2.8/5). Counterfactual explanations achieved the highest comprehensibility (4.4/5) and strongest regulatory alignment but the weakest faithfulness (0.64 +- 0.11). No single method satisfied all five dimensions simultaneously, confirming that explanation method selection must be tailored to the specific regulatory, domain, and stakeholder context.