Beyond Lesion Burden: An Explainable Machine Learning Framework for Characterizing Clinical Resilience in Multiple Sclerosis

Multiple sclerosis (MS) presents a clinico-radiological paradox in which patients with similar magnetic resonance imaging lesion burdens may exhibit different levels of neurological disability. This study investigates whether combining clinical subsystem scores with structural lesion measurements improves the characterization of disability-defined patient groups. Data from 59 patients were categorized as resilient (EDSS ≤ 3.0; n = 42) or vulnerable (EDSS > 3.0; n = 17). Random Forest, Linear Support Vector Machine, and L1-regularized Logistic Regression classifiers were evaluated using five-fold stratified cross-validation. The corresponding ROC-AUC scores were 0.6130, 0.6819, and 0.7271, respectively. SHAP analysis identified cerebellar and pyramidal scores as influential predictors, while lesion volume contributed additional information. Counterfactual explanations generated using DiCE produced 22 valid alternatives across vulnerable patient profiles, with lesion volume modified in all alternatives. These exploratory findings suggest that combining clinical subsystem measurements with structural lesion burden may help characterize disability heterogeneity in MS. Larger independent datasets and further validation are required before clinical application.