We propose a two-stage prediction model. First, we calculated baseline risk scores (linking the probability of developing the outcome to patient characteristics but not to the treatment) using the least absolute shrinkage and selection operator (LASSO) and ‘pre-specified’ (based on previously identified prognostic factors) models. Second, we estimated the probabilities of post-treatment outcome as a function of baseline risk score using an individual patient data network meta-regression model in a Bayesian framework. We applied this approach to predict relapse at two years in 3,590 patients receiving placebo, natalizumab, dimethyl fumarate, or glatiramer acetate.

Characteristics such as age and disability status at the start of the treatment affected baseline risk of relapse. Both the LASSO and ‘pre-specified’ model estimate a mean baseline risk of 37%, although the ‘pre-specified’ model showed greater estimation variation. For high-risk patients (baseline risk > 50%), the absolute benefit of natalizumab versus dimethyl fumarate was 10% and 15% for LASSO and ‘pre-specified’, respectively. For low risk patients (baseline risk < 30%), the absolute benefit of dimethyl fumarate versus natalizumab was 2% and 3% for LASSO and ‘pre-specified’, respectively.

Our flexible approach can be extended to any number of RRMS treatments and other outcomes of interest. Importantly, it has the potential to inform patients and their doctors about the most appropriate treatment, thus contributing to the progression of personalized medicine.