MRI radiomics and clinical feature model for predicting postoperative functional outcome of lumbar disc herniation: a multicenter retrospective study

Background: Lumbar disc herniation (LDH) is a common degenerative spinal disorder. Surgery is effective in relieving nerve compression and improving function. However, some patients fail to achieve a minimal clinically important difference (MCID) in functional improvement after surgery. Existing prediction models mostly rely on subjective scores and lack the comprehensive integration of preoperative quantitative imaging metrics and magnetic resonance imaging (MRI) radiomics. This study aims to identify independent risk factors among preoperative MRI radiomics features and clinical parameters associated with failure to achieve the MCID 6 months postoperatively in patients with LDH. Furthermore, a combined model integrating preoperative MRI radiomics and clinical parameters was developed. Methods: Preoperative MRI images were preprocessed, and the herniated disc region of interest was manually segmented to extract radiomics features. Key radiomics features were selected. Multivariate logistic regression (LR) analysis incorporated clinical variables identified as independent predictors. The outcome variable was defined as whether the patient achieved an MCID in the Patient-Reported Outcomes Measurement Information System Physical Function(PROMIS PF) scores at 6 months postoperatively. A clinical model (CM), a radiomics model (RM), and a combined clinical-radiomics model (CRM) were constructed using support vector machine classifiers. Model performance was evaluated using receiver operating characteristic curves, calibration curves, and decision curve analysis. Results: A total of 455 patients were included. Multivariate LR analysis showed that the preoperative and 3-month postoperative PROMIS PF scores, symptom duration time, psoas muscle cross-sectional area (CSA), the ratio of functional CSA (FCSA) to vertebral body area (VBA) (FCSA/VBA), and multifidus muscle fat infiltration rate were independent risk factors for the study’s primary outcome – failure to achieve the MCID in PROMIS PF scores at 6 months postoperatively. After eliminating redundant features and performing regularization, 10 core radiomics features were selected. The CRM exhibited the best predictive performance across all the datasets. Conclusions: The CRM integrating MRI radiomics features with clinical indicators demonstrated excellent performance in predicting postoperative functional outcomes and strong generalizability across centers.