Comparison of Machine Learning Models for Predicting Recurrent Lumbar Disc Herniation After Percutaneous Endoscopic...

Area of Science:

• Neurosurgery
• Artificial Intelligence
• Medical Informatics

Background:

• Recurrent lumbar disc herniation (rLDH) negatively impacts outcomes after percutaneous endoscopic lumbar discectomy (PELD).
• Accurate prediction of rLDH remains a clinical challenge.
• This study developed and compared machine learning models for rLDH risk stratification.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting rLDH within two years post-PELD.
• To identify key predictors of rLDH.
• To assess the clinical utility and interpretability of developed models.

Main Methods:

• Retrospective analysis of 1483 patients undergoing single-level PELD.
• Development and comparison of six machine learning algorithms, including LightGBM and logistic regression.
• Feature selection using univariate screening and LASSO regression; performance evaluation via AUC, F1-score, Brier score, calibration, and DCA; interpretability using SHAP.

Main Results:

• The overall rLDH rate was 4.25%.
• LightGBM achieved the highest discrimination (AUC = 0.768), while logistic regression had the best F1-score (0.286).
• Increased sagittal range of motion, reduced facet orientation, advanced age, Modic changes, and MSU zone C were significant predictors.

Conclusions:

• An internally validated machine learning framework for rLDH risk stratification was developed.
• LightGBM shows moderate predictive ability but is limited by low recurrence rates for standalone use.
• Clinical utility is highest at low probability thresholds (<10%), suggesting a role in identifying high-risk patients for targeted management.