Objective: To synthesize and evaluate prediction models for post-anesthesia care unit length of stay (PACU LOS), focusing on prolonged recovery and delayed discharge after surgery.
Methods: A systematic review was conducted following PRISMA 2020 guidelines. Cochrane Library, Embase, PubMed và Google Scholar were searched for studies published between 1998 and 2025. Eligible studies developed or validated prediction models related to PACU recovery time, including prolonged PACU LOS, readiness for discharge, or delayed PACU discharge, and reported at least one performance metric (AUC, C-index, or R²). Risk of bias was assessed using the PROBAST tool.
Results: A total of 747 records were identified, and 11 studies met inclusion criteria. Models were categorized into three groups: traditional logistic regression, logistic-based models (nomograms or risk scores), and machine learning models. Predictive performance ranged from moderate to high (AUC/C-index: 0.66–0.94). Common predictors included surgical duration, type of surgery, age, BMI and ASA classification and comorbidities. Definitions of prolonged PACU LOS varied across studies. Only a minority of studies performed external validation, and reduced performance was observed when models were applied to external populations.
Conclusion: Prediction models for PACU recovery show moderate to good performance; however, heterogeneity in outcome definitions and predictors limits comparability. While machine learning models may achieve higher internal performance, models with appropriate validation appear more suitable for clinical implementation. In Vietnam, evidence remains limited, highlighting the need for local validation and model adaptation.