Session PB5.3

Classification of Obstructive and Central Sleep Apnea Using Wavelet Packet Analysis of ECG Signals

J Gubbi*, A Khandoker, M Palaniswami

The University of Melbourne
Melbourne, Australia

Obstructive sleep apnea (OSA) causes a pause in airflow with continuing breathing effort. In contrast, a central sleep apnea (CSA) event is not accompanied with breathing effort. The aim of this study is to differentiate characteristics of CSA and OSA using wavelet packet analysis of the ECG signal over 5 second periods and support vector machines. Six patients were used in the study, they had both CSA and OSA events. The dataset contained ECGs sampled at 256 Hz with 1591 CSA clips and 28786 OSA clips of five seconds duration. Only 5-second ECG (which contains at least one cycle of inspiration and expiration) clips were extracted from pre-scored apnea events. Eight level wavelet packet analysis was performed on each 5 second clip using mother wavelet Daubechies (DB3). Two features, namely the best tree and the entropy of the best wavelet tree, were extracted from each clip. The maximum feature number was 510. However, as only the best tree was being used, approximately 25 features per clip were used in classification. Analysis of the best tree for CSA and OSA events indicated that OSA is predominantly in the low frequency region. However, the best tree of CSA clips appeared in both the high frequency and low frequency regions, specifically between 64 to 96 Hz. This was the distinguishing feature between OSA and CSA. Both sets of features were used separately for classification using the support vector machines. Radial basis kernel with parameters C = 100 and gamma of 0.1 was chosen for final testing after parameter grid analysis. Leaving one patient out, the error was calculated. Using only the best tree as feature resulted in an 11% average error. The entropy of the best tree resulted in a better performance with about an 8% error. These results indicate the possibility of noninvasively classifying CSA and OSA events based on shorter segments of ECG signals.

(Abstract Control Number: 72)