A machine learning model integrating radiomics and clinical factors to predict recurrence risk after rehabilitation in lumbar disc herniation.

This study aimed to construct and internally validate a machine learning model integrating radiomics features and clinical factors to predict post-rehabilitation recurrence risk in patients with lumbar disc herniation (LDH). In this single-center retrospective cohort study, 260 adults aged 18 to 70 years with LDH who completed standardized rehabilitation were analyzed (78 with recurrence and 182 without recurrence). Radiomics features were extracted separately from sagittal T1-weighted, sagittal T2-weighted, and axial T2-weighted magnetic resonance imaging, standardized, concatenated across sequences, and reduced using least absolute shrinkage and selection operator regression. Three logistic regression models were developed: radiomics-only, clinical plus rehabilitation, and combined models. A stratified 70/30 split provided internal validation; receiver operating characteristic curves, calibration analysis, decision curve analysis, and risk stratification were used to evaluate performance. A separate multivariable logistic regression analysis was used to identify clinically interpretable predictors. Age, body mass index, disc height index, improvement in the Oswestry Disability Index, diabetes, Modic changes, and poor rehabilitation adherence differed significantly between the recurrence and non-recurrence groups (all P < .05). The combined model demonstrated the best predictive performance, with an area under the curve (AUC) of 0.85 (95% confidence interval: 0.78-0.91), outperforming the radiomics model (AUC = 0.78) and the clinical model (AUC = 0.76). The calibration curve showed strong agreement between predicted and observed recurrence, and decision curve analysis indicated superior net clinical benefit across a wide range of thresholds. Based on predicted probabilities, patients were stratified into low- (<20%), intermediate- (20%-40%), and high-risk (>40%) groups, with recurrence rates of 15.0%, 30.0%, and 66.0%, respectively (P < .001). Multivariable logistic regression identified disc height index (odds ratio [OR] = 1.80), Oswestry Disability Index improvement (OR = 0.65), poor rehabilitation adherence (OR = 2.50), diabetes (OR = 2.10), and Modic changes (OR = 1.90) as independent predictors of recurrence. Integrating multisequence weighted magnetic resonance imaging radiomics with clinical and rehabilitation variables improved internal prediction of post-rehabilitation recurrence in LDH. The combined model may support risk-adapted rehabilitation intensity and follow-up, but it requires prospective, multicenter external validation before routine clinical use.