Abstract
OBJECTIVES: Lumbar spinal stenosis (LSS) is an increasingly important issue related to back pain in elderly patients, resulting in significant socioeconomic burdens. Postoperative complications and socioeconomic effects are evaluated using the clinical parameter of hospital length of stay (LOS). This study aimed to develop a machine learning-based tool that can calculate the risk of prolonged length of stay (PLOS) after surgery and interpret the results. METHODS: Patients were registered from the spine surgery department in our hospital. Hospital stays greater than or equal to the 75th percentile for LOS was considered extended PLOS after spine surgery. We screened the variables using the least absolute shrinkage and selection operator (LASSO) and permutation importance value and selected nine features. We then performed hyperparameter selection via grid search with nested cross-validation. Receiver operating characteristics curve, calibration curve and decision curve analysis was carried out to assess model performance. The result of the final selected model was interpreted using Shapley Additive exPlanations (SHAP), and Local Interpretable Model-agnostic Explanations (LIME) were used for model interpretation. To facilitate model utilization, a web application was deployed. RESULTS: A total of 540 patients were involved, and several features were finally selected. The final optimal random forest (RF) model achieved an area under the curve (ROC) of 0.93 on the training set and 0.83 on the test set. Based on both SHAP and LIME analyses, intraoperative blood loss emerged as the most significant contributor to the outcome. CONCLUSION: Machine learning in association with SHAP and LIME can provide a clear explanation of personalized risk prediction, and spine surgeons can gain a perceptual grasp of the impact of important model components. Utilization and future clinical research of our RF model are made simple and accessible through the web application.