Diagnosis systems often use structured data. These data have a hierarchical structure related with the questions asked during the interview with the doctor or the survey taker in charge of verbal autopsies. The hierarchical nature of these questions leads to consider this aspect when analyzing medical data. Thus, it is recommendable to choose a similarity measure that takes into account this issue to better represent the reality. We propose the introduction of a kernel taking into account the hierarchical structure of the data interactions between sub-items in supervised binary classification methods. This kernel can integrate the knowledge from the application domain relative to how the features of the problem are organized. We focus on problems whose features can be hierarchically structured. These hierarchies are represented by trees on two levels. Our main contribution is the proposal of a kernel that simultaneously takes into account the hierarchical appearance and the interaction between variables. The proposed kernel has shown a good classification performance on a complex set of medical data including a high number of predictors and classes.