Session PB5.2

Quantification of Obstructive Sleep Apnea in HRV Signals Using Quadratic Time-Frequency Distributions, Dynamic Features and PLS Analysis

AF Quiceno-Manrique, LD Avendano-Valencia, JL Rodriguez-Sotelo*, CG Castellanos-Dominguez

Universidad Nacional de Colombia
Manizales, Columbia

Obstructive sleep apnea (OSA) occurs when pharyngeal muscles loss tone in sleep stages, and the pharynx collapses recurrently, leading to breathing interruption, called apnea. This affection causes a reduction in blood oxygen saturation, which can predispose to diseases like hypertension, arteriosclerosis, left ventricular hypertrophy and systolic dysfunction. In order to perform a diagnosis of OSA, it is necessary to detect the presence of repetitive episodes of apnea and hypopnea during sleep. This presence is most reliably shown by attended overnight polysomnography in a sleep laboratory.
The heart rate is affected by changes in oxygen saturation, thus, it is possible to detect OSA by analyzing the heart rate variability (HRV) extracted from ECG signal. In this work, time-frequency analysis is implemented by means of quadratic transforms belonging to Cohen’s class, in order to analyze the frequency variations of the components of HRV signal.
The methodology used to extract features from time-frequency representations, includes the estimation of dynamic contours, like subband spectral centroids and cepstral coefficients. Furthermore, the dynamic contours are analyzed with partial least squares with the aim of excerpting the most relevant information contained in the time variant features to be used by a classifier, in order to detect OSA.
Finally, the classification stage is implemented with a k-nearest neighbor classifier, and a cross validation scheme is performed with 10 folds, using 70% of the observations in training, and the remaining 30% for testing purposes. The database used is available in Physionet, and it is formed by 70 recordings, labeled by one-minute intervals as normal or apnea. The results showed up to 88% of accuracy in the quantification of OSA, i.e., the classification of one-minute segments.

(Abstract Control Number: 71)