Background: Using clinical data to model the medical decisions behind sequential treatment actions raises methodological challenges. Physicians often have access to many covariates that may be used when making sequentially adaptive treatment decisions for individual patients. They typically prioritize certain covariates over others, when looking at patient's characteristics and history, and use their own subjective weights when making treatment decisions. In the small sample size context, Bayesian variable selection methods can help finding which variables are really used in everyday practice and allow for expert information to be incorporated into prior distributions. The originality of this work is to combine routine care elicitation with the observed data. Motivated by clinical practice data involving repeated dose adaptation for Irinotecan in colorectal metastatic cancer, we propose a modification of the Stochastic Search Variable Selection (SSVS) method, which we call Weight-based SSVS (WBS), to incorporate experts’ medical opinion in prior settings through elicited clinical relevance weights. Methods: To model the relationship between repeated dose adjustment for Irinotecan by patient’s characteristics (age, weight loss, …) and toxicities, in colorectal metastatic cancer, we assumed a linear mixed effect model. Then physician experts were asked to specify numerical clinical relevance weights to express their beliefs about the importance of each variable in their decision making, assessed on a scale between 0 and 100. We use these weights to construct beta prior distributions for the model parameters which rule the inclusion of covariates. We analyzed the WBS model performance compared to the Lasso and SSVS ones through an extensive simulation study. We simulated a sample of patients treated for colorectal metastatic cancer by Irinotecan, following the distributions of our motivating data set. We also applied WBS to our case study. Results: Compared to the usual SSVS method, the WBS method showed better performance and produced the lower rates of false positives and false negatives. Lasso showed poor performance, confirming better performance of Bayesian approach compared to frequentist when dealing with small samples. Conclusion: We propose a highly performant Bayesian variable selection method to incorporate expertise into models when dealing with small sample size.