Explainable Deep Learning Models for High-Stakes Decision Systems

Abstract

Deep learning models have achieved state-of-the-art predictive performance across a broad range of high-stakes decision domains -- including clinical diagnosis, credit scoring, recidivism prediction, and autonomous vehicle control -- yet their intrinsic opacity remains a fundamental barrier to deployment in regulated and accountability-sensitive contexts. Explainability, broadly defined as the capacity of a model to produce human-interpretable accounts of its predictions, is now a legal requirement under the EU Artificial Intelligence Act (2024) and the General Data Protection Regulation for AI systems making decisions with significant effects on individuals. This paper presents a systematic comparative evaluation of six explainability methods applied to deep learning models across three high-stakes domains: clinical image classification (chest X-ray pathology detection), tabular credit risk scoring, and natural language processing (NLP)-based legal document classification. A total of 18 deep learning model-explainer combinations (six explainers applied to three domain models) are evaluated using a four-dimensional explainability quality framework comprising fidelity, comprehensibility, stability, and actionability. Quantitative results demonstrate substantial variation in explainability quality across methods and domains: gradient-based methods (Integrated Gradients, GradCAM) achieve the highest fidelity scores in image domains (mean fidelity = 0.84, SD = 0.06) but poor stability in NLP contexts (mean stability = 0.51, SD = 0.09). Perturbation-based methods (SHAP, LIME) achieve higher comprehensibility and stability across domains (mean = 0.79, SD = 0.07) but lower fidelity in complex image tasks (mean = 0.67, SD = 0.10). The study contributes a domain-stratified explainability benchmark, a practitioner selection guide mapping explainer methods to domain requirements, and empirical guidance for EU AI Act Article 13 transparency compliance in deep learning systems.