Feature subset selection is a central issue in a vast diversity of problems including classification, function approximation, machine learning and adaptive control. On a wide variety of applications, especially when using real data, input features may be not independent and output variable depends on the relationship among inputs rather than on input values themselves. Feature selection methods that assume independence of attributes will fail on these cases. On the other side, most of alternative approaches are quasi-exhaustive, requiring large CPU processing time. In this paper, an alternative methodology based on sensitivity analysis of trained artificial neural networks (ANN) is analyzed. Results so far attained on illustrative toy examples and on real data support the validity of the developed approach.