Personalized recommendations of medical treatments to patients should aim to optimize their expected outcome based on reliable evidence. The available evidence is often scarce as (1) medical practices are often implemented prematurely, before rigorous testing has been completed, and (2) randomized controlled trials often provide evidence for only a selective subgroup. These concerns prove the need for a treatment recommendation system that is informed by scientific evidence on treatment benefit for the considered patient (based on literature review, electronic health records (EHRs), …), but is also able to learn where the information is scarce. In this project, we will develop a data-informed system to assist personalized treatment decision-making based on predictions of what a patient’s outcome would be if one of multiple treatment options were considered. These predictions will be based on available EHRs, but unlike standard predictions, will consider the causal structure of the problem and account for confounding bias. In addition, we will develop a novel study design for optimizing these predictions, by supplementing the EHR data with data from small, targeted randomized experiments integrated into routine practice. Randomized experimentation enables the system to continuously learn and wash away possible (unmeasured) confounding bias present in routine EHRs. We will apply (part of) the methodology to develop an expert system for psoriasis based on the PsoPlus registry.