Technical Note PR - TN 2014 / 00384 Issued : 08 / 2014

This report comprises two goals. The first goal is to investigate to what extent cardiorespiratory activity can provide information about sleep stages; the second goal is the analysis of overnight sleep with respect to finer micro-stages compared with the current definition of sleep stages. To extract information regarding sleep stages from respiratory effort and cardiac activity, we use autoregressive(AR-) based models. The use of cardiorespiratory signals is because they can be acquired unobtrusively. The first part of this thesis focuses on automatic sleep stage classification by incorporating a total of 9 new features extracted with AR models. To examine the performance of sleep stage classification, we use two classification algorithms including linear discriminant (LD) and hidden Markov model with unsupervised Gaussian mixture model (GMM-HMM). The second part is to explore possible microstages, which is done by clustering a total of 23 AR features and computing the between-subject agreement based on the clustering results. For sleep stage classification, results show that the new AR features are discriminative in distinguishing between different sleep stages. Adding AR features to an existing feature set can improve the classification performance. In addition, we found that the GMMHMM classifier performed better than the LD classifier when AR features are used. For the second part, some micro-stages are found, but some of the micro stages cannot be clearly explained, which suggest that further study is needed. PR-TN 2014/00384 Company Confidential until 2017-08 iv  Koninklijke Philips N.V. 2014 Conclusions: An exploration of sleep stages based on cardiorespiratory signals is presented in this report. We use AR models to extract physiological information from respiratory effort and ECG signals. The AR features show discriminative power among existing cardiorespiratory features, and give improvement in classification performances. Comparing LD and GMM-HMM classifiers, the performance of GMM-HMM is generally higher than the performance of LD when AR features are used. From this observation we speculate that AR features contain the information of sleep stage transitions. Apart from sleep staging, much emphasis has been put on proving that the R&K rules do not fully describe the sleep stages. The preliminary results from the exploration of micro-stages show that the clusters can be seen as microstages of the sleep structure, the meaning of some clusters can be explained, but for some clusters, their meanings are still unclear. This suggests that more investigations need to be done on exploring the physiological meaning of each cluster. Company Confidential until 2017-08 PR-TN 2014/00384  Koninklijke Philips N.V. 2014 v