Artificial Intelligence-Assisted Monitoring in Anaesthesia: A Systematic Review and Meta-analysis of Diagnostic Accuracy and Clinical Impact AI in Anaesthesia: Monitoring and Outcomes Review

Abstract

Antonio Andrea Camastra MD, Matheus Requena Escobar MD, Daniel Macedo Oliveira MD, Laiz G. C. Novaes MD, Andre Busatto de Donato MD, Lucas Teixeira Baldo MD, Joao Evangelista Ponte Conrado MD, Cecília Schettini Gueiros MS, Raphael Matheus de Souza Makiyama Lopes MD and Thomas Rolf Erdmann MD, MsC, PhD

Background: Artificial intelligence (AI) is transforming medicine by enabling real-time data analysis and improved decision-making. In anaesthesiology, AI tools are increasingly used for perioperative risk assessment and intraoperative monitoring, but evidence on their real-world performance and safety remains limited.

Methods: We conducted a systematic review and meta-analysis following PRISMA guidelines, including studies from 2010 to May 2025 that evaluated AI applications—machine learning (ML), deep learning, neural networks, and fuzzy logic— in adult patients undergoing general or regional anaesthesia. Primary outcomes were perioperative complications (e.g., hypotension, hypoxia, bradycardia, delirium, vomiting, cardiac arrest, mortality, acute kidney injury [AKI]); secondary outcomes included haemodynamic stability, ICU admission, and length of stay. Risk of bias was assessed using RoB 2 and ROBINS-I, and random-effects models were applied.

Results: Eighteen studies with diverse surgical settings and sample sizes (60 to >450,000 patients) were included. ML models consistently outperformed conventional statistical methods. Ensemble algorithms, such as XGBoost and random forests, achieved AUROC values of 0.942 and 0.96, respectively. Deep learning models, including Max-Pooling Convolutional Neural Networks, predicted mortality with AUROC 0.867. Hypotension Prediction Index (HPI) trials showed 88% sensitivity, 87% specificity, and a 77% reduction in hypotension burden. Hybrid models integrating waveform and electronic health record data reported AUROCs of 0.807 for mortality and 0.766 for AKI.

Conclusions: AI-based monitoring, especially ML and biomarker-guided strategies, offers substantial improvements in perioperative risk stratification and haemodynamic management. Wider clinical adoption requires external validation, explainable AI frameworks, and rigorously designed randomized controlled trials demonstrating meaningful patient outcome benefits.

PDF