Spinal Cord Radiomics-Driven Machine Learning Predicts Meaningful Clinical Improvement After Surgery for Degenerative...

A prospective observational cohort study. To determine whether machine learning models using radiomic features derived from preoperative MRI, clinical variables, or their combination can predict achievement of the minimum clinically important difference (MCID) in function and quality of life after surgery for degenerative cervical myelopathy (DCM). Predicting surgical outcomes in DCM remains challenging, as conventional MRI and clinical scores incompletely reflect spinal cord pathology. Radiomics quantifies voxel-level intensity and texture patterns from routine MRI, providing quantitative measures of tissue heterogeneity that may serve as imaging biomarkers of recovery potential. Forty-six patients with DCM underwent preoperative 3D T2-weighted MRI and surgical decompression. Spinal cord radiomic features (Shape3D, First-Order, GLCM, and GLSZM) were extracted using PyRadiomics. Baseline clinical variables included age, sex, duration of symptoms, T2 hyperintensity, and functional scores assessed with the baseline mJOA and SF-36 PCS scores. Three-month MCID achievement was defined using established thresholds. Predictive models were developed using radiomic features, clinical variables, or their combination. For mJOA MCID, the combined radiomics-clinical model achieved the best performance (AUC = 0.88 ± 0.13). For SF-36 PCS MCID, the combined model achieved an AUC = 0.78 ± 0.17 and an AUCPR of 0.82 ± 0.14. SHapley Additive exPlanations identified texture-based radiomic features and age as dominant predictors for mJOA MCID, whereas first-order radiomic features and baseline SF-36 PCS were most influential for SF-36 PCS MCID. MRI-based spinal cord radiomics improves prediction of meaningful postoperative recovery beyond clinical data, supporting their potential as imaging biomarkers for individualized prognostication in DCM.