Combining magnetic resonance imaging and evoked potentials enhances machine learning prediction of multiple sclerosis disability worsening.

Introduction: Predicting long-term disability progression in multiple sclerosis (MS) remains a significant challenge. Existing prognostic models often rely on single-modality data or conventional measures, such as lesion count on magnetic resonance imaging (MRI) or latency values from evoked potentials (EPs), overlooking subclinical disease progression. This study aimed to develop a multimodal machine learning (ML) pipeline integrating clinical, high-dimensional MRI, and motor EP time-series (EPTS) features to predict disability worsening in MS. Methods: A retrospective cohort of 127 people with MS (PwMS; 424 episodes) from a tertiary MS center in Belgium was used, including clinical data, T2-weighted fluid-attenuated inversion recovery MRI, and motor EPs. Disability worsening was defined as a change in the expanded disability status scale (EDSS) over two years, stratified by baseline EDSS. MRI features included 42 anatomical and lesion volumes and 100 radiomic descriptors from lesions and the normal-appearing white matter (NAWM). EPTS features included latency, peak-to-peak amplitude (PPA), and high-dimensional descriptors selected using highly comparative time-series analysis (HCTSA) and Boruta. ML models (Light Gradient Boosting Machine (LGBM), random forest, logistic regression) were trained using 20×repeated stratified 3-fold cross-validation. Performance was evaluated using the area under the receiver operating characteristic curve (AUROC), average precision (AP), and Brier score. SHapley Additive exPlanations (SHAP) were used for interpretability. Results: Across 96 model configurations, models combining MRI and EPTS features, with or without clinical data, consistently outperformed single-modality models across AUROC, AP, and Brier score, regardless of algorithm or feature representation. The best-performing model (Brier score = 0.062) was an LGBM using combined MRI and EPTS data. MRI radiomics dominated feature importance, especially shape- and texture-based features from NAWM and lesion regions. EPTS features, particularly waveform dynamics (e.g., Sliding Window) and PPA, provided complementary value and improved sensitivity. EPTS-only models showed the highest AUROC, but combined models achieved the best overall balance across all performance metrics. Conclusion: This is the first study to integrate clinical, MRI radiomics, and motor EPTS features in an ML pipeline for MS prognosis. Combining structural and functional subclinical markers improves the prediction of disability worsening and supports multimodal monitoring for personalized care.