Session PC1.4

Signal Stationarity Assessment for the Heart Rate Variability Spectral Analysis

AN Kalinichenko*, MI Nilicheva, SV Hasheva, OD Yurieva, OV Mamontov

St. Petersburg State Electrotechnical University
St. Petersburg, Russia

The NN-interval series used as the input signal for heart rate variability (HRV) spectral analysis techniques represents nonstationary random process due both to its physiological nature and to external factors such as noises and changes in signal acquisition conditions. Nevertheless some standard spectral analysis techniques (the use of which requires fulfillment of the signal stationarity condition) are widely applied to this task. In practice the analyzed signal is considered as close to stationary that can be provided by the use of correct signal acquisition procedure and by human check of the date prior to their processing. But a number of tasks exists where the above conditions can not be satisfied: bed-side ECG processing, Holter recordings analysis, signal processing in portable ECG analyzers and some other. This paper is devoted to the examination of several methods of signal stationarity assessment as applied to the task of HRV spectral parameters estimation. The following techniques were considered: - run test; - spectral error measure; - autocorrelation function distance (ACFD); - generalized likelihood ratio (GLR). Artificially modeled signals (containing locally stationary fragments and transient fragments) as well as a set of real signal recordings obtained in the course of orthostatic test (each recording containing at least two distinctly different stationary fragments corresponding to different stages of the test) were used as experimental data. The ability of producing some stationarity (nonstationarity) indicators was examined for all listed above methods. It was shown that such wide spread method as run test does not provide satisfactory results. Among the other examined methods the best performance demonstrated ACFD and GLR techniques. The parameters optimization for these two methods was implemented and practical algorithms for their use were developed. Besides a generalized indicator of HRV signal nonstationarity was proposed in order to get a quantitative estimation of the signal quality. The developed methods can be used for the enhancement of HRV parameters estimation reliability both in real-time ECG analyzers and in computer-based systems for HRV investigations.

(Abstract Control Number: 218)