Development and validation of a machine learning model utilizing the ACS-NSQIP database for predicting early postoperative Re-operation in lumbar microdiscectomy

Introduction: Early re-operation secondary to postoperative surgical site infection (SSI) in lumbar microdiscectomy, although rare, poses significant morbidity and healthcare burden. Research question: Given the variable performance of traditional statistical and machine learning (ML) models in predicting this outcome, this study aimed to develop and validate a ML model to accurately predict 30-day re-operation following SSI in patients undergoing lumbar microdiscectomy. Material and methods: De-identified patient data were obtained from American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP), focusing on patients with microdiscectomy. After applying inclusion criteria, 79,870 eligible cases were analyzed, of which 462 (0.6%) experienced re-operation following SSI. A nested cross-validation pipeline with Bayesian optimization was implemented using extreme gradient boosting (XGBoost) combined with Synthetic Minority Over-sampling Technique (SMOTE), and multiple classification thresholds were evaluated. Model interpretability was assessed through SHapley Additive Explanations (SHAP) to identify the most influential predictors. Results: The final model achieved an area under the receiver operating characteristic curve of 0.996, with accuracy consistently at 0.993 and sensitivity of 0.924-0.985 across thresholds. Specificity remained at 0.993, while the positive predictive value fluctuated modestly (0.436-0.437), and the negative predictive value was 1.000. SHAP analysis highlighted any SSI, race, smoking status, diabetic status, and revised Risk Analysis Index scores as top predictors influencing re-operation risk. Discussion and conclusion: An optimized ML approach incorporating synthetic data augmentation yielded high predictive performance for an infrequent yet critical complication of lumbar microdiscectomy. These findings underscore the potential of data-driven models to enhance perioperative risk stratification and support targeted preventive strategies.