Session P93.2

Recognizing Central and Obstructive Sleep Apnea Events from Normal Breathing Events in ECG Recordings

AH Khandoker*, GL Jayavardhana Rama, M Palaniswami

The University of Melbourne
Melbourne, Australia

Obstructive sleep apnea (OSA) causes a pause in airflow with continuing breathing effort. In contrast, central sleep apnea (CSA) event is not accompanied with breathing effort. The aim of this study is to differentiate CSA and OSA events from normal breathing events using wavelet based features of ECG signal over 5 second period and two-stage feed forward neural network. Total 19000 clips(each of 5-second duration) from normal breathing events, 2055 clips from 430 CSA and 7379clips from 1262 OSA events were selected from single lead ECGs (sampling rate=256 Hz) collected from 39 patients’ overnight sleep studies. Only 5-second ECG (which contains at least one cycle of inspiration and expiration) clips were extracted from pre-scored apnea events. Each ECG clip was decomposed to eight levels of detailed coefficients using mother wavelet Symlet (sym3). From each level of coefficients, Shannon’s entropy, log energy entropy, arithmetic mean and geometric mean were calculated. In total 32 features were extracted. Among them 18 best features were selected by add-remove method with Mahalanobis distance measures. The best feature was found to be Shannon’s entropy value of detailed coefficients at level 8 (0.5~1.0 Hz) (Receiver operating Characteristics (ROC) area=0.79 between normal events and apnea (OSA+CSA) clips) and level 6 (2~4 Hz) (ROC area=0.84 between OSA and CSA clips). At the first stage of classification, apnea events were classified from normal breathing events and at the second stage, apneas were classified into CSA and OSA types. Two-stage Feed forward Neural network(NN) models with 5 hidden layers was trained using the selected best features. Ten-fold cross validation (randomly 70% for training and remaining 30% for testing events by combining clips) results of the first stage NN model show the average accuracy of 93.0% in recognizing apnea events from all clips. As for the performance of second stage NN model using the same set of best features, average accuracy of 94.96% in classifying OSA and CSA events was obtained. These results indicate the possibility of noninvasively recognizing OSA/CSA events from normal breathing events based on shorter segments of ECG signals.

(Abstract Control Number: 84)