OBJECTIVES: To identify predictors of disease modifying therapy (DMT) initiation among treatment-naïve multiple sclerosis (MS) patients using machine learning and structured data from a large, geographically diverse electronic health records (EHR) database. METHODS:  Optum-Humedica de-identified EHR dataset was used to select MS patients, ≥18 years with no prior DMT experience, from integrated delivery networks (1/1/2007-12/31/2013). First observed MS diagnosis was the index date, and patients had evidence of continuous clinical activity 12-months pre- and post-index. We used a proprietary machine learning platform, Reverse Engineering and Forward Simulation (REFS™), to build an ensemble of models to examine the association of patients’ baseline characteristics and DMT initiation post-index. The area under the curve (AUC) statistic assessed accuracy of prediction models, and we validated prediction models in an independent dataset. RESULTS:  Sample selection yielded 12,516 MS patients (DMT initiation=25%; mean age=49.9 years; females=76%). Predictors identified in every model of the REFS™ ensemble included year of MS diagnosis, geographic location of the patient, and prescriptions for oil-soluble vitamins. Patients diagnosed in 2012 (versus earlier years) had the largest median odds ratio (OR) in the ensemble for DMT initiation (OR, interquartile range [IQR]: 3.10, 3.09-3.12) followed by patients living in the Northeast and West (respectively, 2.59, 2.57-2.60; 2.55, 2.50-2.58). Additional predictors of DMT initiation with selection frequency >90% included eye disorders (1.44, 1.43-1.45), stimulants (1.72, 1.69-1.73), and income (1.73, 1.70, 1.74). When validated in an independent dataset AUC was 0.71. CONCLUSIONS:  Using REFS™ to analyze structured EHR data, we identified demographic and clinical predictors of DMT initiation with moderate to strong predictive accuracy. Further analyses should be completed using additional granular data from unstructured fields in this data source and machine learning to refine the accuracy of our models and the predictors of DMT initiation in the MS population.