Sleep apnea detection using wavelet analysis of ECG derived respiratory signal

The purpose of this study is to find a reliable and practical way for detecting the minute by minute occurrence of sleep apnea. For this aim, the time series of instantaneous respiratory rates (IRR) estimated with electrocardiography (ECG) derived respiratory (EDR) signals are analyzed by using wavelet decompositions. EDR signals are derived from the two sets of single-lead ECGs by 0.2–0.8 Hz band-pass filter implementations. ECG signals are obtained from apnea-ECG database on PhysioNet databank. Wavelet decompositions are implemented to the segments of 3-minutes length time series of IRRs in which the 2nd minute is accepted as deciding minute for apnea. According to the results obtained from wavelet analysis of the first data set consisting of 35 recordings, variances of 3rd, 4th, 5th and 6th detail components can be used as discriminative features which demonstrate the minute based real time apneas. The second data set consisting of 35 recordings is used for testing this consequence. For this aim, the feature vectors derived from the first data set is used for training a nonlinear auto-regressive (NARX) type artificial neural network (ANN) classifier. Due to the minute-based assessments of overnight sleep ECGs revealed that implementation of the NARX based classification following the wavelet decompositions has success level of %82.7 and %78.3 for the first (learning) and second (test) dataset, respectively. However, due to the subject-based assessment the success level is greater than %93,33.