Preliminary Development of a Multimodal Deep Learning Model Integrating Clinical Data and Dual-Plane MRI to Predict 1-Year Patient-Reported Lower-Limb Numbness Improvement After Tubular Microdiscectomy

Abstract Objective To develop and preliminarily evaluate a multimodal model integrating clinical data and dual-plane magnetic resonance imaging (MRI) features to predict 1-year patient-reported lower-limb numbness improvement after tubular microdiscectomy (TMD). Materials and Methods This retrospective two-center study included 214 adults: 111 in the training cohort, 48 in the internal validation cohort, and 55 in the external validation cohort. Mean ages were 53.2 ± 15.0, 54.1 ± 14.1, and 51.4 ± 14.7 years, respectively; male participants comprised 58 (52.3%), 28 (58.3%), and 28 (50.9%), and female participants 53 (47.7%), 20 (41.7%), and 27 (49.1%), respectively. Deep features were extracted from sagittal and transverse T2-weighted MRI using residual networks and combined with clinical variables in a support vector machine. Outcome-informed feature selection and tuning were confined to the training cohort using 10-fold cross-validation. Candidate models and the decision threshold were selected using Hospital 1 internal validation data. ComBat parameters were estimated from Hospital 1 training data and applied to Hospital 2 without re-estimation; Hospital 2 did not inform model development. Results The final combined model incorporated 25 dual-plane MRI-derived features and four clinical variables. It yielded an area under the receiver operating characteristic curve of 0.868 in the Hospital 1 internal validation cohort used for model selection and 0.833 (95% CI, 0.720–0.947) in the Hospital 2 external validation cohort. Comparisons with the clinical-only and dual-plane imaging-only models were descriptive and should be interpreted as exploratory. Conclusion This study provides preliminary model-development evidence that clinical and dual-plane MRI features may aid prediction of 1-year patient-reported numbness improvement after TMD. Larger prospective, multicenter, multivendor studies using validated sensory outcomes are needed before individualized clinical application.