Artificial intelligence in perioperative pressure injury management: a scoping review of evidence, gaps, and implications for population health and digital public health policy

Background Perioperative pressure injury (POPI) is a significant global public health challenge, contributing to substantial iatrogenic morbidity and heavy economic burden worldwide. Traditional prevention is limited by insufficient assessment tools, the “black box” nature of intraoperative monitoring, and delayed intervention. Artificial intelligence (AI) offers a novel approach by integrating multidimensional data for real-time risk prediction; however, no scoping review has systematically mapped AI applications across the full perioperative continuum from a population health perspective. Methods Systematic searches were conducted in PubMed, Embase, CINAHL, Cochrane Library, IEEE Xplore, ACM Digital Library, Web of Science, and Scopus from inception to December 11, 2025. Studies applying machine learning, conventional regression-based benchmarking tools, or intelligent sensor-based monitoring for POPI prediction, monitoring, or prevention during the perioperative period were included. Two reviewers independently performed screening and data extraction, followed by descriptive synthesis and evidence mapping. Results Six studies (2018–2025), all from East Asia, were included. This geographic concentration highlights an unassessed generalizability and algorithmic fairness risk, rather than an established inequity finding. Evidence concentrated on post-hoc risk stratification (five prediction studies), predominantly using machine learning models (e.g., Random Forest, XGBoost) with AUCs of the best-performing models ranging from 0.806 to 0.836; one study reported a prediction accuracy of 0.9733, which is not directly comparable to AUC. Models identified specialty-specific risk factors (e.g., cardiopulmonary bypass time, forced positioning). Only one proof-of-concept study addressed intraoperative real-time monitoring, visualizing pressure increases after 2 hours of surgery. Evidence for postoperative early warning systems is absent. Conclusion Current studies demonstrate the feasibility of specialty-specific post-hoc risk stratification using intraoperative covariates, though truly preoperative models remain absent. Intraoperative monitoring shows early promise, but postoperative early warning is a critical gap, and evidence lacks global diversity. Future research should develop closed-loop intelligent systems integrating preoperative prediction, intraoperative multimodal monitoring, and postoperative warning. Nursing protocols should consider incorporating AI-identified risk factors as hypothesis-generating inputs pending prospective validation, while policymakers should establish frameworks for continuous validation and cross-population calibration to ensure equitable impact.