Explainable Artificial Intelligence for Critical Healthcare Applications

Abstract

The deployment of artificial intelligence in critical healthcare settings demands not only high predictive accuracy but also transparent, interpretable
decision-support mechanisms that clinicians can audit, trust, and act upon. Existing black-box deep learning models, despite strong performance on
clinical benchmarks, lack the explanatory transparency required for regulatory compliance and clinical adoption. This study proposes a multi-modal
Explainable AI (XAI) framework — termed ClinXAI — that integrates gradient-weighted class activation mapping (Grad-CAM), SHapley Additive
exPlanations (SHAP), and Local Interpretable Model-agnostic Explanations (LIME) within a unified clinical decision pipeline. ClinXAI was validated
across three critical healthcare domains: sepsis onset prediction (MIMIC-III ICU dataset, n = 52,847), diabetic retinopathy grading (EyePACS fundus
image dataset, n = 88,702), and early-stage lung cancer detection (NLST CT scan dataset, n = 26,314). The framework achieved AUC-ROC scores
of 0.943, 0.961, and 0.938 respectively, while clinician trust surveys (n = 84 specialists) rated explanation adequacy at 4.31 out of 5.0 — a 41.2%
improvement over standard model outputs without explanations. Temporal feature attribution identified creatinine and lactate trajectories as the most
influential sepsis predictors, consistent with established clinical guidelines. ClinXAI provides a scalable, domain-agnostic architecture for transparent
AI-assisted diagnosis in high-stakes healthcare environments.
Keywords: ; ; ; ; ;  ;