Comparison of Machine Learning Algorithms for Predicting Spine Surgery Duration

Le résumé fourni par la source

Background and Objectives: Accurate prediction of surgical duration is essential for efficient operating room management. Spine surgery frequently shows discrepancies between estimated and actual surgical duration, which can disrupt surgical scheduling and resource allocation. The aim of this study was to develop and compare machine learning (ML) algorithms for predicting spine surgery duration and identify the most effective approach. Materials and Methods: Electronic medical records of 3376 patients who underwent spine surgery were retrospectively analyzed at a single center. The dataset was divided into training (80%, n = 2700) and internal test (20%, n = 676) sets using stratified random sampling based on surgical duration quintiles. To match the intended use at the time of operating room scheduling, four models (Random Forest, XGBoost, multilayer perceptron [MLP], and weighted least squares [WLS] regression) were developed using only predictors available at scheduling and evaluated on the independent internal test set; a full-information model that additionally included intraoperatively recorded variables was examined for comparison. Results: XGBoost demonstrated the best predictive performance, achieving a mean squared error (MSE) of 3014.6 min2 (equivalent to a root mean squared error [RMSE] of 54.9 min; 95% CI for MSE, 2558.3–3556.0) and an R2 of 0.622 (95% CI, 0.566–0.675). The other ML models showed comparable performance, whereas WLS regression performed less favorably; the full-information model performed equivalently (R2 0.622). Compared with surgeon-estimated duration alone, XGBoost reduced mean absolute error by approximately 20 min (paired ΔMAE −20.1 min; 95% CI, −23.7 to −16.5) and improved prediction accuracy within 60 min by 14.0 percentage points. SHapley Additive exPlanations (SHAP) analysis identified surgeon identity as the most influential predictor across all models (24.7–41.9%), followed by procedure type and surgeon-estimated duration. Conclusions: Machine learning models substantially improved prediction of spine surgery duration compared with conventional approaches, with XGBoost showing the highest predictive accuracy. Surgeon identity emerged as the most important predictor of surgical duration. Implementation of such models may improve operating room scheduling efficiency and resource allocation but requires prospective evaluation of clinical and workflow outcomes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.