276 - Predicting surgery times with machine learning versus traditional methods

Description:

Purpose: ORs represent one of the most expensive hospital resources, with an average cost of $46 per minute. Inaccurate case duration estimates result in inefficiencies, including idle time, delays, and increased staffing costs. Traditional estimation methods rely on surgeon input or historical averages, which often lack precision and accuracy. This systematic review investigates whether machine learning (ML) models provide more accurate predictions of surgical duration and identifies the key factors that influence model performance.

Literature Review: Emerging evidence suggests that ML algorithms capable of integrating large, multidimensional data sets can enhance prediction accuracy. However, the variability in reported results, methodology, and clinical integration challenges underscores the need for a comprehensive review.

Research Question: We aimed to review the ML methods for surgical duration prediction comprehensively, describe ML models' performance relative to traditional estimation, investigate the factors influencing the performance of ML models, and address the implications on the adoption of ML in real-world practice.

Conceptual Framework: The PICO framework was used to develop the search strategy in this systematic review.

Methodology: A systematic review of peer-reviewed studies published between January 2019 and October 2024 was conducted using MEDLINE, Embase, and CINAHL. Studies were included if they compared ML models to traditional surgical time estimation techniques and reported predictive performance metrics, such as absolute error from actual surgery times.

Data Analysis: Eleven studies met the inclusion criteria. ML model performance was assessed using related, but non-uniform metrics. Risk of bias was evaluated using the PROBAST tool; certainty of evidence was appraised using the GRADE framework.

Results: Most ML models outperformed traditional methods, with an average relative accuracy improvement of 25.7%. Neural networks and tree-based algorithms (e.g., XGBoost) were the most effective. Narrowly trained models, focused on specific procedures or specialties, tended to outperform broad models. However, generalizability was limited, with only one study conducting external validation. Common limitations included poor description of missing data, lack of internal validation, and inconsistent definitions of surgical duration.

Conclusions/

Discussion: ML models show promise in improving surgical duration prediction, particularly when tailored to specific clinical contexts. However, successful implementation requires investment in infrastructure, local model retraining, and integration with existing scheduling workflows. Greater transparency in model development and validation is critical for future research.

Perioperative Nursing Implications: More accurate surgical scheduling improves OR efficiency, reduces staff burnout, minimizes patient delays, and enhances interprofessional collaboration. Perioperative nurses can advocate for and help guide the integration of predictive technologies, ensuring that ML solutions align with frontline clinical workflows and data availability. Nurses' insights into OR processes are vital to selecting relevant predictors and promoting successful model adoption.