CARE-X: Capability-Aware Reliable and Explainable AI for Healthcare Operations

Abstract

Healthcare artificial intelligence increasingly functions as part of an augmented system in which information, computational resources, external interfaces, and operating rules can alter what a fixed predictive component is able to accomplish. Yet evaluation commonly focuses on final performance, leaving these dependencies largely unexamined. We introduce CARE-X, a non-diagnostic framework for auditing capability in assistive healthcare workflows across the complete 2^4 factorial lattice of resources, information, actions/interfaces, and rules. CARE-X combines exact Boolean-lattice attribution with minimal enabling sets, minimal failure cuts, survival and concentration measures, and workflow-level Capability Passports. We show formally that two augmented systems can exhibit identical technical completion under every intervention while differing in safe capability. Controlled experiments include a 2,048-case simulator and a frozen Random Forest trained on 14,000 synthetic cases. We further examine CARE-X using the open MIMIC-IV-ED Demo, comprising 222 emergency-department stays from 64 patients and 666 non-diagnostic workflow cases. Full augmentation achieves a safe-handling rate of 0.9384. Removing either information or actions/interfaces reduces this rate by 0.6051, with the tied dependency pattern reproduced in a patient-disjoint split. Exact Möbius reconstruction error is zero. In contrast, removal of the rules channel produces a loss of only 0.0015, showing that dependencies observed under controlled conditions need not transfer uniformly to real records. CARE-X thus provides a way to distinguish predictive competence from the operational dependencies that enable safe system capability, while treating the real-record analysis as feasibility evidence rather than clinical validation.