Machine learning approaches for predicting prolonged hospital length of stay after lumbar fusion surgery in patients aged 75 years and older: a retrospective cohort study based on comprehensive geriatric assessment.

Objective: Postoperative recovery following lumbar fusion surgery in patients aged 75 years and older often requires a prolonged length of stay (PLOS) in the hospital. Accurately predicting the risk of PLOS and assessing its risk factors for preoperative optimization are crucial to guide clinical decision-making. The aim of this study was to select the risk factors for PLOS and develop a machine learning (ML) model to estimate the likelihood of PLOS based on comprehensive geriatric assessment (CGA) domains in older patients undergoing lumbar fusion surgery.

Methods: An observational cohort of 242 patients aged ≥ 75 years (median age 80 years) undergoing lumbar fusion surgery at a single center from March 2019 to December 2021 was retrospectively reviewed. Predictor variables consisted of clinical characteristics, CGA variables, and intraoperative variables. The primary outcome was PLOS, defined as a hospital LOS above the 75th percentile in the overall study population. Patients were randomly divided into two groups (7:3) for model training and validation. Ensemble ML algorithms were used to select the significant variables associated with PLOS, and 9 ML models were used to develop predictive models. The Shapley Additive Explanations (SHAP) method was used for model interpretation and feature importance ranking.

Results: Three ensemble ML algorithms selected 9 CGA and clinical variables as influential factors of PLOS. The random forest (RF) model had the best predictive performance among the models evaluated, with an area under the receiver operating characteristic curve of 0.822 (95% CI 0.727-0.917) and F1-score of 0.571. SHAP values indicated that the duration of surgery, the number of fusion levels, and age were the most important predictors, while the Fried frailty phenotype was the most important CGA variable for PLOS. The RF model and its SHAP interpretations were deployed online for clinical utility.

Conclusions: This ML model could facilitate individual risk prediction and risk factor identification for PLOS in older patients undergoing lumbar fusion surgery, with the potential to improve preoperative optimization.