Context: Building decision support tools in medicine requires identifying relevant variables to model the medical decision. Currently, machine learning techniques are used to select variables used for decision making from patient care data. Along with data-driven analysis, eliciting experts opinion can be useful to model decision making. Thus, combining expert data to observed data into a new variable selection method could identify the whole set of relevant variables for decision making. We propose a method that introduces experts’ information into a Bayesian variable selection model, the Stochastic Search Variable Selection (SSVS) model. We consider the context of medical decisions regarding dose adjustment of Irinotecan for the treatment of metastatic colorectal cancer. Methods: Clinician first provide numerical clinical relevance weights to express their beliefs about the importance of each variable in their decision of dose adaptation. Then, a sample of simulated data is generated from these weights and combined with the observed data via the power prior method. We compare the performance of our method to the SSVS model in our case study. Results: The elicited weights deeply vary across clinicians. Performance does not depend on the amount of expert information. For the same amount of observed and expert information combined using the power method, we are able to select rare variables with high elicited weight. Conclusion: We present a Bayesian variable selection method incorporating elicited expert information and observed data. The method selects a set of relevant variables to model the medical decision process.