The reference standard for sleep classification uses manual scoring of polysomnography with fixed 30-s epochs. This limits the analysis of sleep pattern, structure and, consequently, detailed association with other physiologic processes. We aimed to improve the details of sleep evaluation by developing a data-driven method that objectively classifies sleep in smaller time intervals. Two adaptive segmentation methods using 3, 10 and 30-s windows were compared. One electroencephalographic (EEG) channel was used to segment into quasi-stationary segments and each segment was classified using a multinomial logistic regression model. Classification features described the power in the clinical frequency bands of three EEG channels and an electrooculographic (EOG) anticorrelation measure for each segment. The models were optimised using 19 healthy control subjects and validated on 18 healthy control subjects. The models obtained overall accuracies of 0.71 ± 0.09, 0.74 ± 0.09 and 0.76 ± 0.08 on the validation data. However, the models allowed a more dynamic sleep, which challenged a true validation against manually scored hypnograms with fixed epochs. The automated classifications indicated an increased number of stage transitions and shorter sleep bouts using models with smaller window size compared with the hypnograms. An increased number of transitions from rapid eye movement (REM) sleep was likewise expressed in the model using 30-s windows, indicating that REM sleep has more fluctuations than captured by today's standard. The models developed are generally applicable and may contribute to concise sleep structure evaluation, research in sleep control and improved understanding of sleep and sleep disorders. The models could also contribute to objective measuring of sleep stability.