To develop and internally validate a clinically deployable supervised machine-learning (ML) model for identifying advanced disability in multiple sclerosis (MS), defined as Expanded Disability Status Scale (EDSS) ≥ 6, using routinely collected clinical and magnetic resonance imaging (MRI) variables; and to quantify the marginal discriminative contribution of MRI, particularly cervical-cord metrics, beyond clinical features.

Anticipating advanced disability in routine care remains difficult. Prior models often use small cohorts, omit spinal measures, or rely on non-routine inputs. A clinic-ready tool that operates on standard data and reports uncertainty is needed.

We analyzed a harmonized real-world registry (N = 5,018), randomly split 80/20 with stratification. Predictors reflected standard care: demographics; MS type; clinical history and examination (initial EDSS, disease duration, Symbol Digit Modalities Test (SDMT)); and brain/spinal MRI fields, including cervical-cord measures. A gradient-boosting classifier (HistGradientBoosting) was trained within a preprocessing pipeline (imputation, scaling, one-hot encoding). Test-set discrimination was summarized by AUC with bootstrap 95% confidence intervals (CI); statistical significance was evaluated via label permutation. Feature influence was estimated on the test set using permutation ΔAUC.

The model identified EDSS ≥ 6 with AUC 0.961 (95% CI 0.940–0.978); permutation p < 0.001. The most influential clinical features were MS type (ΔAUC 0.0645), initial EDSS (ΔAUC 0.0551), age (ΔAUC 0.0166), disease duration (ΔAUC 0.0109), and SDMT (ΔAUC 0.0080). MRI provided supportive information, with cervical-cord metrics conferring incremental discrimination beyond clinical variables.

A supervised ML approach using routine clinical inputs, complemented by cervical-cord MRI, achieved high discrimination for advanced disability. Clinical variables drove performance, while MRI added value. These findings support immediate test-bed implementation with prospective and external validation, workflow integration, and assessment of clinical impact.