Machine learning models for predicting treatment outcomes in chronic non-specific back pain patients undergoing lumbar extension traction.

Conservative treatment for chronic non-specific low back pain (CLBP) includes lumbar extension traction (LET) to re-align lumbar lordosis (LL). This study explores the use of machine learning (ML) models to predict post-treatment outcomes in patients with CLBP undergoing LET and how these predictions can support clinical decision-making. We utilized a retrospective database of 431 consecutive patients with uncomplicated CLBP. Post-treatment variables predicted included LL, NRS pain score, and Oswestry Disability Index (ODI). Input model variables included pre-treatment LL, sacral base angle (SBA), ratio of LL/SBA fit type, NRS, ODI, frequency, duration, LET compliance, and demographic variables of age and BMI. Initial variables were analyzed to predict post-treatment outcomes. Three ML models—Random Forest (RF), XGBoost, and Multilayer Perceptron (MLP)—were employed to handle both continuous and categorical variables, and performance was evaluated for predictive accuracy. Factors affecting outcomes were identified using Shapley Additive Explanations. Treatment was a multimodal spine rehabilitation program featuring LET applied 3–6 times per week, varied between 4 and 10 weeks, and follow-up was performed at the end of care. Improvements in LL, NRS, and ODI were − 11.5° to − 23.6°, 7.3/10 to 3.3/10, and 33.2% to 10.4%, respectively. Among the ML models, XGBoost demonstrated the highest predictive accuracy for lumbar lordotic angle (R2 = 0.728) and pain score (R2 = 0.648), while Random Forest slightly outperformed XGBoost for ODI (0.631 vs. 0.616). MLP performed poorly for ODI predictions (R2 = 0.201), indicating difficulty in capturing functional disability patterns. SHAP analysis identified fit type, compliance, traction frequency, pre-treatment lumbar curve, and BMI as the most influential predictors. These predictors offer actionable insight for clinical decision-making by allowing clinicians to stratify patients based on predicted responsiveness, tailor LET frequency and duration, and educate patients on the importance of compliance. This study demonstrates that ML models, particularly XGBoost and Random Forest, can effectively predict LET outcomes, supporting personalized treatment strategies for CLBP patients.