1 wc2009_canisiusetal_formatted.doc Heart Rate Spectrum Analysis for the automated Classification of Sleep Stages S. Canisius 1 , T. Ploch 1 , T. Penzel 2 , D. Krefting 3 , A. Jerrentrup 1 and K. Kesper 1 1 Philipps-University Marburg – Faculty of Medicine, Dept. for Internal Medicine, Sleep Disorders Centre, Marburg, Germany 2 Charité Universitätsmedizin Berlin, Interdisziplinäres schlafmedizinisches Zentrum, Berlin, Germany 3 Charité Universitätsmedizin Berlin, Abteilung medizinische Informatik, Berlin, Germany Abstract— Sleep disorders include a huge variety of diseases and their diagnosis very often requires complex and expensive biosignal recordings (polysomnography). Special interest lies in the visual classification of sleep stages in 30 sec windows (epochs) out of this biosignal data, requiring sufficient man- power and experience as well as time. Using the FFT of the heart rate signal to extract the LF/HF frequency power ratio as well as relative peak frequency power within the HF band in combination with the variability of peak frequency power in the HF band we were able to classify sleep stages with an au- tomated algorithm. As signal source we used the ECG signal recorded during the night included in polysomnography re- cordings from the SIESTA database. The visual 30 sec epoch classification of sleep stages is stored together with the signal data, thereby allowing the comparison of visual classifications to those of the algorithm. While the absolute accuracy of cor- rectly assigned epochs was only 57.5%, our approach can provide a rough overview of the distribution of sleep stages as the clinically relevant result. Therefore, this study represents a first step towards the sleep stage classification from the heart rate signal. Keywords— ECG, polysomnography, sleep stage, heart rate. I. INTRODUCTION Sleep disorders are characterized by a huge variety of diseases and represent an important disease entity within the modern 24-hour society. Major symptoms of these disorders are impaired sleep quality, daytime sleepiness, unintentional falling asleep during the day as well as headache, elevated blood pressure or problems in maintaining concentration. Furthermore, many sleep disorders are known to be associ- ated with an increased risk of cardiovascular complications like chronic arterial hypertension, myocardial infarction or stroke. For diagnosing sleep disorders, the investigation of the sleep structure during the night plays a key role. As many sleep disorders impose a disruption of the physiological sleep structure, it is important to investigate the distribution of sleep stages for an objective assessment of sleep quality. Up to now, the sleep stages are scored using different elec- trophyssiological signals being recorded during the night. These signals are at first hand the EEG, but also EOG (eye movements) and EMG (muscular activity). These signals are then scored visually in 30s epochs according to specific rules [1] (Rechtschaffen and Kales using frequency of the EEG, muscle tone, eye movements), resulting in a sleep structure curve (called the hypnogram) as can be seen in Figure 1. Fig. 1 Hypnogram of a healthy volunteer Recording the described electrophysiological signals dur- ing the entire night requires a complex and expensive 8-12- hour investigation (polysomnography) being performed within specialized sleep disorders centres. This means, the patient has to spend the night in a totally different environ- ment. Furthermore, he has to “suffer” extensive wiring, which very often disrupts or impedes sleep even more. Ap- plying visual analysis of the recorded signals as the gold standard, also the classification of sleep stages requires sufficient manpower and experience, not to mention the spent amount of time (app. 45min-1h for the analysis of one recording). It is self-evident, that there is a high demand for easier, cheaper and less obtrusive diagnostic approaches. The ECG signal (electrocardiogram) also represents an electrophysiological signal, but is much easier to conduct, requiring only 2-3 single-use electrodes, so that it can be performed also in the patients’ home. Furthermore, during the night ECG recordings result in a relatively stable signal, as movement artifacts (that occur e.g. during stress ECGs) are quite rare. Also extensive wiring as within a polysom- nography is not necessary. Using the ECG signal for determination of sleep struc- ture, i.e. the distribution of sleep stages during the night may therefore represent a much easier and cheaper approach to obtain the needed information about the patients’ sleep [2;3]. Taking these considerations into account, we developed a new signal processing algorithm which uses spectrum anal-