Session M9.3

Multitaper Spectral Estimation Reveals Excess Power At 0.02 Hz Which Characterizes Apnea Patients

M.R. Jarvis, P.P. Mitra

Caltech
Pasadena, CA, USA

We have implemented a simple method for distinguishing between patients with apnea and normal patients. The method is based on a moving window spectrum (spectrogram) of the ECG raw signal squared and utilizes the fact that the power in the signal changes in a characteristic manner in apnea patients.
Apnea patients show an excess of power in a frequency band centered at approximately 0.02 Hz which is apparent during apnea episodes. This peak is absent in normal patients and may be used to identify both apnea patients and their episodes with a high degree of confidence. The success of our method stems from our use of multitaper methods for spectral estimation. The low value of this frequency means that good bias control is required to clearly separate this peak from zero. This is achieved in multitaper spectral estimation by tapering the data with a set of functions (Slepian sequences) which are explicitly chosen to minimize bias.
In addition we have investigated the point process extracted from the raw signal during apnea. The point process description is of interest since the only information it utilises is the location of the cardiac pulses regardless of shape. We find that periods of normal breathing in apnea patients tend to be associated with a particularly periodic heartbeat. This can also be used to identify apnea patients although with less confidence. We conclude that simply the times of the cardiac pulses is not complete information to extract information about apnea, though it does provide a moderately good classification.
Finally, we looked for possible artifacts of the nature of direct respiratory signals in the raw data (generated, say, by sensor motion). While these are present, indicated by peaks in the spectrum of the raw signal at the breathing frequency, they are not strong enough to provide the sort of performance we obtained by looking more directly at the cardiac signals. In conclusion, we have obtained a robust yet simple signature of sleep apnea based on ECG time series by looking at the low frequency fluctuations of the power in the signal. We look forward to understanding the physiological origins of this signal.
20000503.123923 Score :30/30 = 100.00%.