Operationalizing a human-centered explainable AI framework for smart health systems: a case study of perioperative acute kidney injury

• Proposes a Human-Centered Explainable AI framework for smart health system deployment. • Demonstrates the framework through perioperative Acute Kidney Injury, a high-risk and high-cost use case. • Achieves strong predictive performance (F2 score of 0. 92) with estimated cost savings exceeding 19, 000 per patient. • Integrates clinician involvement to support interpretability, trust, and effective human–AI collaboration. • Shows that successful AI-enabled smart systems require usability, workflow integration, and organizational readiness, beyond predictive accuracy alone. Despite rapid advances in artificial intelligence, the adoption of predictive analytics in healthcare remains constrained by persistent human-centric and operational challenges, including limited interpretability, poor usability, and weak integration into clinical workflows. These challenges impede the development of effective smart health systems and limit the real-world impact of AI-driven decision-support tools. To address these gaps, this study proposes a Human-Centered Explainable AI (HCXAI) framework that incorporates explainability, workflow alignment, and user engagement as core design principles for intelligent clinical systems. The framework is demonstrated through a case study on perioperative Acute Kidney Injury (AKI), a high-risk complication with significant implications for patient safety and hospital costs. Using real-world data from a tertiary hospital in Abu Dhabi, the proposed system integrates cost-sensitive learning, SHAP-based explainability, and a clinician-oriented interface to support human–AI collaboration in perioperative care. Results show that class weighting provides a strong baseline, while combining resampling with class weighting yields modest improvements, with the best-performing models achieving an F2 score of up to 0. 922. From an economic perspective, the proposed approach reduces the estimated cost relative to a no-model baseline by approximately 40%–42%, primarily through improved identification of high-risk cases. Beyond predictive performance, qualitative feedback from clinicians indicates that trust and adoption depend critically on system usability, transparency, and alignment with existing workflows. While explainability features improve confidence, interface design, electronic health record compatibility, and structured onboarding are identified as essential enablers for deployment. Overall, the findings highlight that effective smart health systems require a balanced integration of technical intelligence and human-centered design, demonstrating the value of HCXAI approaches in bridging the gap between prediction and practice.