Multi-modal machine learning for prediction of post-anesthesia care unit recovery time

Abstract Background Efficient recovery management in the Post-Anesthesia Care Unit (PACU) is important for patient safety and care coordination, yet estimating recovery duration remains challenging. Traditional regression models and discharge scoring systems have shown limited accuracy and reliability across patient populations. Methods We developed and evaluated machine learning models for PACU recovery prediction using an institution-wide dataset of 203,393 surgical cases collected between January 2012 and July 2025. Structured variables (demographics, surgical metrics, provider identifiers, timing features) were integrated with unstructured operative narratives transformed into biomedical SBERT embeddings, yielding a multi-modal dataset. Model performance was evaluated using robust validation procedures to ensure findings were not dependent on a single patient group or time period. Results Gradient-boosted tree ensembles demonstrated the best performance, with XGBoost achieving a mean absolute error of approximately 42 minutes and explaining over one-third of the variance in recovery time. Nearly 70% of predictions fell within 60 minutes of observed recovery duration. Statistical testing confirmed significant improvements over linear baselines, conventional ensembles, and neural networks. Explainability analyses revealed that operative narratives provided the greatest predictive signal, while provider identifiers and surgical location also contributed, reflecting both patient-level complexity and system-level workflow factors. Conclusions Combining structured perioperative variables with embeddings of operative narratives enables accurate and interpretable PACU recovery prediction at scale. These findings demonstrate the feasibility of multimodal prediction of PACU discharge-ready time and provide a foundation for future evaluation of operational and clinical applications. Future work should include external validation and integration into real-time perioperative systems.