Predicting postoperative length of stay: a feature selection approach to predictive modeling in lumbar fusion surgery

Predictive modeling has the potential to improve preoperative planning and resource allocation in lumbar fusion surgery. This study aimed to identify the 20 most important variables for predicting prolonged postoperative length of stay (pLOS) using machine learning (ML).

The ACS-NSQIP database was queried for lumbar fusion procedures performed between 2012 and 2022, including ALIF, PlatIF, PLIF, and combined PLIF+PlatIF. Variable selection was performed using MUVR and Boruta, followed by hierarchical clustering and 5-fold cross-validation to ensure feature robustness. The 20 selected features were used to train multiple ML models, including tree-based classifiers (Random Forest, XGBoost, CatBoost, LightGBM), support vector classifiers, neural networks, ensemble methods, and logistic regression.

A total of 114,892 patients were included. Eleven patient-specific and nine procedural variables were identified as most predictive of prolonged pLOS. Among patient factors, dialysis, congestive heart failure, and bleeding disorders were strongest predictors. Among procedural factors, osteotomy, billing of additional fusion codes, and longer operation time had the greatest impact. The neural network achieved the highest accuracy (71.2%), recall (79.4%), and F1-score (73.8%), though all models performed similarly, with minimal variation in classification metrics.

These findings underscore that model choice plays a limited role once optimal features are selected - feature selection was the most critical determinant of predictive performance.