Networks are often used to represent the relations among the variables of a dynamical system. The properties of network topology are usually exploited to understand the organization of the system. Nevertheless, the dynamical organization of a system might considerably differ from its topological one. In this paper, we describe a method to identify “relevant subsets” of variables. The variables belonging to a relevant subset should be strongly integrated and should have a much weaker interaction with the other system variables. Extending previous works on neural networks, an information-theoretic measure is introduced, i.e., the Dynamical Cluster Index, in order to identify candidate relevant subsets. The method solely relies on observations of the variables’ values in time