Abstract / Summary
Abstract Background/Purpose Preoperative prediction of neurological recovery remains challenging for patients with cervical ossification of the posterior longitudinal ligament (OPLL) undergoing anterior controllable antedisplacement and fusion (ACAF). This study aimed to develop and rigorously internally validate an interpretable CT-based radiomics nomogram integrating multi-regional features for predicting 12-month neurological recovery after ACAF in a large single-center cohort. Methods This retrospective study enrolled 280 consecutive patients with cervical OPLL treated with ACAF at a single tertiary center between 2017 and 2024, randomly stratified into a training cohort (n = 200) and an internal validation cohort (n = 80) at a 7:3 ratio. Radiomics features were extracted from four regions of interest (ROIs): whole ossification mass, mature core/active edge sub-regions, target vertebral body, and a 3-mm ossification–vertebra interface zone generated by morphological dilation. A multi-method feature selection pipeline (LASSO, mRMR, SVM-RFE) with intersection strategy was applied to construct the radiomics score (Rad-score). Nested 10-fold cross-validation was used to avoid information leakage, and 1000-repeat random resampling validation was performed to assess model stability. SHAP analysis was performed for model interpretability. The nomogram was validated by discrimination, calibration, clinical utility, subgroup robustness and sensitivity analysis. Results Eight radiomics features were selected as the final signature, including three texture features from the active edge, two first-order density features from the interface zone, two texture features from the vertebral body, and one shape feature from the whole ossification mass. All final features showed inter-observer ICC ≥ 0.78. The combined nomogram yielded AUCs of 0.832 (95% CI 0.775–0.879) in the training cohort and 0.805 (95% CI 0.706–0.880) in the internal validation cohort. The 1000-repeat resampling validation yielded a mean AUC of 0.798 ± 0.021, indicating stable performance. The model showed good calibration (Brier score = 0.128) and superior net benefit over the clinical-only model, with a net reclassification improvement of 0.312 (95% CI 0.154–0.470). Subgroup analyses confirmed robust performance across ossification types, K-line status and surgical extents. Conclusions The multi-regional CT radiomics nomogram enables accurate and stable preoperative prediction of neurological recovery after ACAF in a large single-center cohort, with rigorous internal validation and interpretable biological features. Large-scale multicenter prospective studies are needed to confirm generalizability.