In the present paper, a comparative study of performance for three techniques of feature extraction is presented in order to classify the sleep stages using EEG signals. A multilayer feed forward neural network was used for classification. Six sleep EEG records for each of ten patients were selected from Cairo Center of Sleep Disorder. Three methodologies of analysis were utilized for feature extraction. These include: autoregressive modeling (AR), bispectral analysis, and discrete wavelet transform (DWT), where principle component analysis (PCA) was used to reduce feature dimensionality. The features derived from the three methodologies of signal analysis were used as input feature vectors to the classifier. Information fusion is very important task in pattern recognition as it is difficult to develop classifiers with a high
identification performance rate. The multilayer feed forward neural network gives higher classification rate using the data fusion at the feature extraction level. It reaches 83.4%.