A single-source preoperative prediction model for recurrent lumbar disc herniation after primary discectomy:...

Recurrent lumbar disc herniation (rLDH) follows primary discectomy in 5% to 25% of patients. Existing models often embed postoperative variables, merge data from unlinked repositories, or omit calibration and clinical-utility assessment. We aimed to develop and internally validate a strictly preoperative model for rLDH using a single multi-center registry, with calibration-first evaluation, decision curve analysis, and a SHAP-aligned clinical score (SHapley Additive exPlanations).

This retrospective development and split-sample internal validation study used the Spine Patient Outcomes Research Trial (SPORT) lumbar disc herniation cohort; 738 of 1,244 enrolled patients met inclusion criteria. The outcome was symptomatic rLDH at the index level within 2 years, confirmed radiologically or by reoperation. Thirteen preoperative predictors and 4 algorithms (elastic-net logistic regression, random forest, gradient boosting, support vector machine) were evaluated, trained on 70% with stratified 5-fold cross-validation and 500-resample optimism correction, then tested on a 30% holdout. Reporting followed TRIPOD+AI.

Recurrence occurred in 110 patients (14.9%). Logistic regression performed best: AUC 0.738 (95% CI 0.694-0.782; optimism-corrected 0.724), calibration slope 0.94, Brier score 0.112. SHAP identified annular defect size, smoking, and younger age as the most influential predictors. A 10-item score stratified patients into low- (5.8%), moderate- (15.2%), and high-risk (32.6%) groups. Decision curve analysis showed net benefit across threshold probabilities of 8% to 30%.

A timing-consistent, single-registry preoperative model achieved acceptable discrimination and calibration for rLDH. The SHAP-aligned score offers transparent preoperative risk stratification; external validation is required before clinical use.