The conventional approach to the analysis of human sleep uses a set of pre-defined rules to allocate each 20 or 30-s epoch to one of six main sleep stages. The application of these rules is performed either manually, by visual inspection of the electroencephalogram and related signals, or, more recently, by a software implementation of these rules on a computer. This article evaluates the limitations of rule-based sleep staging and then presents a new method of sleep analysis that makes no such use of pre-defined rules and stages, tracking instead the dynamic development of sleep on a continuous scale. The extraction of meaningful features from the electroencephalogram is first considered, and for this purpose a technique called autoregressive modelling was preferred to the more commonly-used methods of band-pass filtering or the fast Fourier transform. This is followed by a qualitative investigation into the dynamics of the electroencephalogram during sleep using a technique for data visualization known as a self-organizing feature map. The insights gained using this map led to the subsequent development of a new, quantitative method of sleep analysis that utilizes the pattern recognition capabilities of an artificial neural network. The outputs from this network provide a second-by-second quantification of the sleep/wakefulness continuum with a resolution that far exceeds that of rule-based sleep staging. This is demonstrated by the neural network's ability to pinpoint micro-arousals and highlight periods of severely disturbed sleep caused by certain sleep disorders. Both these phenomena are of considerable clinical value, but neither are scored satisfactorily using rule-based sleep staging.

Original publication




Journal article


Journal of sleep research

Publication Date





201 - 210


University of Oxford, Medical Engineering Unit, UK.


Nerve Net, Humans, Electroencephalography, Wakefulness, Sleep Stages, Sleep, REM, Models, Biological, Adult, Female