In [Slyadnev et al, 2020], we describe an algorithm for detecting isolated features. The user selects a series of capping faces, and the algorithm automatically finds all faces separated from a CAD part by this selection.
This algorithm gives a pretty powerful way of selecting faces for subsequent defeaturing. For example, one can easily separate the cavity feature residing between the selected inlet and outlet faces. Internally, the algorithm checks if the initial AAG of the model is separated into a greater number of connected components once the capping faces are eliminated from the graph. You will find more details in our paper mentioned above.