|
C. Figueroa, A. Suhrbier, T. Penzel, N. Wessel, J. Kurths
Polysomnographic recordings can be transformed into a sequence of sleep stages which are displayed as a hypnogram. The rules that have been conventionally applied to obtain such hypnograms are those provided by Rechtschaffen and Kales. The procedure involves a time consuming visual analysis of the polysomnographic recordings by expert scorers. More recently the American Academy of Sleep Medicine (AASM) has proposed new rules with less sleep stages (only one deep sleep stage rather than two and no movement time) thereby easing the development of automatic sleep scoring. We contribute to the aim of automation by applying the technique of linear discriminant function analysis to extract information from the polysomnographic signals that best reproduce the hypnograms provided by the expert sleep scorers. In addition to variables from the electroencephalogram (EEG), we examine also properties of the electrocardiogram (ECG), electromyogram (EMG) and respiratory signals in relation to their ability to represent the sleep stages as maximally separable categories. We achieve this separation with a linear discriminant function, which we apply to a group of ten healthy subjects and fifteen patients of sleep apnoea (both data sets from the Siesta Database). We estimate the accuracy in the prediction of the hypnogram from our variables and assess the sources of error in our results. We have performed the study by selecting several groups of variables (EEG alone, EEG plus combinations of ECG, respiration and EMG). A subset of twenty variables chosen by an iterative search results in accuracies of 85% and 76% for the healthy and patient groups, respectively. Larger sets of variables including also EMG properties give 86% accuracy for the healthy group and 76% for the patient group. Including respiratory and ECG variables increased the accuracy to 87% for the healthy group and 79% for the patient group. We conclude that only 20 EEG variables are sufficient to obtain an accuracy higher than the approx. 70% agreement achieved by independent human scorers. We propose to investigate alternative methods for variable selection and fuzzy logic as a contrast to the present classification. |
|