This paper proposes a methodology to the feature selection problem of pattern classification problems. For this purpose, pattern recognition or signal processing involves two major tasks: clustering transformation and then, feature selection. The concept of clustering reduces the dimensionality of the measurement space and generates a set of features. However, there is so far no covering theory how to select discriminative and biologically important features from the pool of generated features. This paper describes a hybrid approach in which an unsupervised learning method namely, a modified Self-Organizing Feature Map is used for feature extraction and then, a smaller linearly independent set of features are selected using Step-Down Logistic Regression. Experimental results using Doppler umbilical artery waveforms indicate that classification results using such features shows a very high correlation with the actual output. Also, Modified Self-Organizing Feature Map followed with the Step-Down Logistic Regression show to be a very powerful approach to discovering useful feature space in complex data.