There is growing interest in sleep research and increasing demand for screening of circadian rhythms in genetically modified animals. This requires reliable sleep stage scoring programs. Present solutions suffer, however, from the lack of flexible adaptation to experimental conditions and unreliable selection of stage-discriminating variables. EEG was recorded in freely moving C57BL/6 mice and different sets of frequency variables were used for analysis. Parameters included conventional power spectral density functions as well as period-amplitude analysis. Manual staging was compared with the performance of two different supervised classifiers, linear discriminant analysis (LDA) and Classification Tree. Gamma activity was particularly high during REM (rapid eye movements) sleep and waking. Four out of 73 variables were most effective for sleep–wake stage separation: amplitudes of upper gamma-, delta- and upper theta-frequency bands and neck muscle EMG. Using small sets of training data, LDA produced better results than Classification Tree or a conventional threshold formula. Changing epoch duration (4 to 10 s) had only minor effects on performance with 8 to 10 s yielding the best results. Gamma and upper theta activity during REM sleep is particularly useful for sleep–wake stage separation. Linear discriminant analysis performs best in supervised automatic staging procedures. Reliable semi-automatic sleep scoring with LDA substantially reduces analysis time.