A study on the predictive value of preoperative MRI-based radiomics for postoperative neurological function in patients with ossification of the posterior longitudinal ligament

Abstract

This study proposes to develop an integrative predictive model that synergistically combines quantitative radiomic features, established clinical risk parameters, and advanced machine learning methodologies to forecast postoperative neurological functional outcomes in patients diagnosed with cervical ossification of the posterior longitudinal ligament (OPLL). This retrospective cohort study recruited 113 consecutive patients with cervical OPLL who underwent surgical decompression at the Second Affiliated Hospital of Naval Medical University.Postoperative neurological outcomes were assessed over a 12-month follow-up period, with patients stratified into two groups based on the modified Japanese Orthopaedic Association (mJOA) score recovery rate. Radiomic features were extracted from preoperative T2-weighted MRI scans covering the C2–C7 spinal segments. After feature selection using least absolute shrinkage and selection operator (LASSO) logistic regression, eleven machine learning algorithms were employed to build radiomic signature models. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis, with the Hosmer-Lemeshow test assessing calibration. Clinically significant risk factors were identified through univariate and multivariate logistic regression analyses and subsequently incorporated into clinical prediction models. An integrated predictive framework was then developed by combining radiomic signatures with clinical risk factors, and its discriminative ability was validated using ROC analysis. SHAP (SHapley Additive exPlanations) analysis was performed post-hoc to enhance the interpretability of the MLP-based combined model. Finally, decision curve analysis (DCA) was performed to evaluate the clinical utility of the radiomics-clinical nomogram in both training and validation cohorts. A total of 1,688 radiomic features were extracted from the preoperative MRI datasets. After reproducibility filtering (ICC > 0.75) and redundancy elimination (Pearson correlation > 0.9), 283 features entered LASSO regression, which ultimately retained 9 optimal features (λ = 0.045). Among the evaluated machine learning models, Multi-Layer Perceptron algorithm demonstrated superior diagnostic performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.870 (95%CI:0.789–0.951) in the training cohort and 0.796 (95%CI: 0.643–0.949) in the internal validation cohort. The integrated radiomics-clinical combined model further enhanced predictive accuracy, attaining an AUC of 0.878 (95% CI:0.802–0.953) in the training cohort and 0.830 (95% CI:0.685–0.976) in the internal validation cohort. The integrative analysis of clinical parameters and radiomic signatures demonstrates robust predictive efficacy for postoperative neurological recovery in cervical OPLL patients, thereby offering a data-driven framework to inform personalized therapeutic decision-making.