Machine learning algorithms for prediction of cerebrospinal fluid leakage after posterior surgery for thoracic ossification of the ligamentum flavum.

To develop and validate a machine-learning (ML) model that pre-operatively predicts cerebrospinal-fluid leakage (CSFL) after posterior decompression for thoracic ossification of the ligamentum flavum (TOLF), and to elucidate the key risk factors driving model decisions. Electronic medical-record and imaging data of 318 consecutive TOLF patients who underwent laminectomy between January 2009 and June 2023 were retrospectively analysed (CSFL = 101, 31.8%). The cohort was randomly split 4:1 into training (n = 254) and testing (n = 64) sets. Class imbalance was addressed with two synthetic oversampling techniques, SMOTE and ADASYN. A baseline logistic-regression (LR) model and four ML algorithms-XGBoost, Random Forest, LightGBM and Support Vector Machine (SVM)-were tuned via Bayesian optimisation. Primary endpoints were F1-score and recall; secondary metrics included AUC, accuracy, calibration curves and Brier scores. Probabilities were recalibrated with Platt Scaling and Isotonic Regression, and model interpretability was assessed with SHAP and LIME. Under SMOTE, SVM achieved the best overall performance (F1 = 0.889, recall = 0.881); its Brier score improved to 0.103 after Isotonic Regression. Feature-attribution analyses consistently identified multi-segment involvement, residual spinal-canal area (RrSCA) and related diametric ratios (RrPD, RrDCM), operative time, and intra-operative blood loss as the strongest predictors of postoperative CSFL. A SMOTE-enhanced, isotonic-calibrated SVM provides accurate and reliable CSFL risk estimation in TOLF patients and is freely available as an online tool ( https://github.com/DebtVC2022/CSFL_predict ). The model supports preoperative risk stratification, patient counselling, and peri-operative management, yet requires prospective, multicentre validation to establish broad clinical utility.