Explainable artificial intelligence in anesthesiology prediction models: bridging the gap from black-box algorithms to clinical decision support

Machine learning prediction models consistently outperform conventional clinical scoring tools across perioperative outcomes, yet remain largely unadopted at the bedside. This narrative review examines explainable artificial intelligence (XAI) as a bridge between algorithmic performance and clinical translation, drawing on a targeted search of PubMed and Embase (January 2018 to March 2026). Five prediction domains are synthesized: perioperative hypotension, postoperative complications, difficult airway management, depth-of-anesthesia monitoring, and regional anesthesia and pain management. SHAP (SHapley Additive exPlanations) has emerged as the dominant XAI method, identifying modifiable intraoperative targets. Attention mechanisms and GradCAM provide complementary interpretability for time-series and imaging-based models. Critical gaps persist: most delirium prediction models include no XAI analysis; no standardized evaluation framework exists, and no prospective trial has demonstrated that XAI-enhanced recommendations improve patient outcomes. Addressing these gaps is a research priority if XAI is to fulfill its potential as a facilitator of perioperative AI adoption.