Develop and validate a machine learning model to predict the risk of persistent pain after percutaneous transforaminal endoscopic discectomy.

Background: Persistent pain is a common complication following percutaneous transforaminal endoscopic discectomy (PTED) for lumbar disc herniation. Identifying associated risk factors and developing a predictive model are crucial for guiding clinical decisions. This study aims to utilize machine learning models to predict persistent pain, identify key influencing factors, and construct a risk model to assess the likelihood of persistent pain.

Methods: We first compared baseline characteristics and pathological indicators between patients who developed persistent pain and those who did not after PTED. Significant factors were used as input features in four machine learning models: Logistic Regression (LR), Support Vector Machine (SVM), XGBoost, and Multilayer Perceptron (MLP). Each model was optimized through grid search and 10-fold cross-validation. Performance was evaluated using ROC curves, F1 score, accuracy, recall, and precision. Models with AUC values exceeding 0.9, specifically XGBoost and MLP, were selected for SHAP visualization and risk prediction model construction.

Results: Among the four machine learning models, XGBoost and MLP achieved the best performance, with AUC values of 0.907 and 0.916, respectively. SHAP analysis identified a history of lumbar spine trauma and herniation calcification as key features positively influencing persistent pain risk. Elevated inflammatory markers (e.g., CRP, ESR, and WBC) and older age also significantly impacted predictions. Using the most important features from XGBoost and MLP, a risk prediction model was constructed and externally validated, achieving an AUC of 0.798, indicating good predictive accuracy.

Conclusion: History of lumbar spine trauma, herniation calcification, and inflammatory markers are important predictors of persistent pain after PTED. The risk prediction model based on XGBoost and MLP shows high predictive accuracy and can serve as a valuable tool for clinical decision-making.