OBJECTIVES: Identifying potential non-responders to interferon-beta/glatiramer acetate (BRACE) treatment among Multiple-Sclerosis (MS) patients is critical to individualise and optimise treatments. This study assessed the applicability of advanced predictive models based on commercial claims data to predict non-response and to identify key predictors. METHODS: MS patients in the PharMetrics PlusTM US claims database were included who had initiated or switched to new BRACE treatments (the index date) and who had experienced ≥1 relapse in the year prior to index date. Non-response was defined as ≥1 relapse in the two years following index date. Predictions of non-response were estimated using logistic regressions with and without a Lasso penalty, Support Vector Machines and Random Forests. Model performance was assessed using the Area Under the Curve (AUC) computed on left-out folds based on tenfold cross-validation. Model discrimination was also assessed by comparing actual non-response rates between patients in the highest and lowest risk group based on quintiles of predicted scores. Sensitivity analysis was conducted using various outcome definitions. RESULTS: The study included 1767 responders and 1902 non-responders. AUCs for the models ranged between 62%-64.5% Compared with standard logistic regression, controlling for overfitting and exploited non-linearities led to modest improvements in model performance. Based on results from the logistic regression with Lasso, 73.5% patients in the highest risk group did not respond to BRACE treatments compared with 32.8% of patients in the lowest risk group. Non-responders were more likely to be younger, have high pre-index medication use and to have experienced ≥2 pre-index relapses.  Prediction accuracy was substantially improved by modifying the outcome definition to patients with ‘many’ relapses post-index. CONCLUSIONS: Medical claims data can provide insight into treatment non-response, with the non-response rate twice as high in the highest compared with the lowest risk group. The overall performance of the models was moderate.