Development of a predictive model for adverse outcomes after unilateral biportal endoscopic spine surgery in the treatment of lumbar disc herniation using six interpretable machine learning models

Abstract

Objective This study aimed to develop and validate an interpretable machine learning model incorporating clinical and radiological features to predict the risk of adverse outcomes in patients with lumbar disc herniation (LDH) following unilateral biportal endoscopic (UBE) surgery. The model was designed to facilitate preoperative risk stratification and assist individualized clinical decision-making. Methods A retrospective cohort study was conducted involving 418 patients with LDH who underwent UBE surgery at our institution between January 2022 and January 2025. Patients were randomly allocated to the training and validation cohorts in a 7:3 ratio. The collected variables comprised demographic characteristics, perioperative parameters, and radiological features. Four complementary feature selection approaches, including the least absolute shrinkage and selection operator (LASSO), Boruta, minimum redundancy maximum relevance (mRMR), and recursive feature elimination (RFE), were applied. Six machine learning algorithms were developed and optimized using five-fold cross-validation within the training cohort. The predictive performance of each model was subsequently evaluated in the validation cohort, and the area under the receiver operating characteristic curve (AUC) was used to identify the optimal model. The best-performing model was further evaluated using multiple performance metrics. SHapley additive explanation (SHAP) analysis was performed to interpret model predictions and enhance transparency in the decision-making process. Finally, an online risk prediction calculator was developed based on the optimal model to facilitate clinical application. Results Modic changes, Pfirrmann grade, APHC, and age were identified as important predictors of adverse outcomes following UBE surgery. Among the six evaluated machine learning models, the ExtraTrees model demonstrated the highest predictive performance. In the training cohort, the model achieved an AUC of 0.933 (95% CI, 0.901–0.960), with an accuracy, precision, sensitivity, and specificity of 0.850, 0.867, 0.703, and 0.934, respectively. In the validation cohort, the model maintained satisfactory performance, achieving an AUC of 0.884 (95% CI, 0.818–0.941), with an accuracy, precision, sensitivity, and specificity of 0.818, 0.811, 0.652, and 0.913, respectively, suggesting good generalizability. Decision curve analysis demonstrated that the ExtraTrees model achieved greater net clinical benefit than the other models across a wide range of threshold probabilities. SHAP-based global importance plots, dependence plots, and individual-level waterfall and force plots demonstrated that higher Modic grades, higher Pfirrmann grades, increased APHC values, and older age were associated with an increased risk of adverse outcomes. The direction and magnitude of these associations were largely consistent with clinical expectations, thereby enhancing model interpretability and clinical acceptance. Furthermore, a web-based calculator (https://mtt123456.shinyapps.io/dynnomapp/) was developed to facilitate the practical application of the prediction model. Conclusion The proposed prediction model demonstrated favorable predictive performance in the internal validation cohort and may serve as a practical tool for preoperative risk assessment of adverse outcomes following UBE surgery for LDH. This model may assist in identifying high-risk patients and provide evidence to support preoperative planning, patient counseling, and individualized treatment optimization. However, because validation was limited to a randomly divided single-center dataset, further external validation involving multicenter cohorts, different temporal periods, and surgeons with varying levels of experience is required to determine the model's robustness and generalizability.