In many real–world applications, e.g., medical diagnosis, behavioral analysis, Bayesian networks are used to describe relationships between variables. However, such variables are not directly observable, but can be inferred through noisy but costly features. In this paper, our previously proposed framework of dynamic instance–wise feature selection and classification is extended to work with structured data instances, i.e., data instances where relationships between classification variables are represented using a known Bayesian network. The objective is to maximize classification accuracy while minimizing the total cost of selected features. To this end, starting from lowest degree nodes, the proposed method sequentially selects features for each variable in the Bayesian network and performs classification. The resulting classification decisions are propagated through the Bayesian network and used during the classification process of the remaining variables. The performance of the proposed method is illustrated on two datasets and its effectiveness is compared with existing methods.