NM2 Using Randomized and Observational DATA to Predict Heterogeneous Treatment Effects

Abstract

Treatment effects that vary across different patients, known as heterogeneous treatment effects, are important to quantify for clinical decisions. A determinant of the effects variability is the baseline risk of patients, often reflecting the severity of their condition. As treatment options are often numerous, comparative effectiveness research is vital for clinical decision-making, and evidence from randomised and observational studies is increasingly available. We aim to combine observational and randomised evidence in a two-stage prediction model for heterogeneous treatment effects. The model integrates prognostic research and network meta-analysis methods by synthesizing several sources of data. We apply the model in estimating the heterogeneous effects of three drugs for patients with relapsing-remitting multiple sclerosis and ascertain the optimal treatment in each patient subgroup. First, we develop a prognostic model and we predict the risk of the outcome prior to treatment, using data from the Swiss Multiple Sclerosis Cohort - SMSC (935 patients). Then, we use this baseline risk as prognostic factor and effect modifier in a network meta-regression model with individual participant data from three randomized clinical trials to make personalized predictions under different treatment options. Finally, we use validation methods to evaluate the performance of the model. Our model indicates that baseline risk modifies the treatment effects of Natalizumab, Glatiramer Acetate and Dimethyl Fumarate. Several patient characteristics (such as age, gender, baseline disability status, etc.) influence the risk of relapse, and this in turn moderates the benefit within each one of the drugs. We will present an R-Shiny app that estimates the risk of relapse under all available treatments conditional on patients’ characteristics and consequently, indicates which treatment is preferable for which patient. Combining observational with randomised data and bridging methods from prognosis research and evidence synthesis provides a generic approach to estimate heterogeneous treatment effects.