A.N. Kalinichenko, O.D. Yuryeva

The technique of heart rate variability analysis is based on statistical and spectral analysis of NN intervals (time intervals between adjacent heart beats of background rhythm). The signal formed by the series of NN intervals is nonstationary. Nevertheless some traditional techniques of random signals analysis, such as FFT based and autoregressive spectral analysis, are commonly used for the investigation of heart rate regulation physiological mechanisms. The use of the mentioned above methods suggests fulfillment of signal stationarity condition within the frames of the analyzed signal fragment. Otherwise the obtained results can not be considered as statistically consistent. The most well known methods of random process stationarity assessment are runs test and reverse arrangements test. However this approach provides only an estimation of the signal nonstationarity in the whole and gives no possibility of detecting the signal sections that can be considered as locally stationary. More efficient way to solve this problem is to use more complicated techniques such as correlation, spectral analysis and parametric modeling of processes. The results of experimental examination of different approaches to the signal stationarity assessment with regard to the heart rate variability spectral parameters estimation are presented. The following three approaches were considered: autoregressive parameters monitoring; analysis of detrended signal and assessment of generalized likelihood ratio. The goal of the algorithms examination was estimation of the stationary segments borders detection accuracy. The stationarity violation was considered as detected, if the used nonstationarity index value in the vicinity of the corresponding time instant was at least two times greater than maximum value of this index at the preceding stationary segment. The value of the detection error at the crossing point of type I (no stationarity violation was detected) and type II (false detection of stationarity violation) error curves was used as a criterion for the algorithms parameters optimization procedure. The above methods were examined with the use of both artificially modeled signals and a set of real signal recordings. A combination of parameters providing least values of nonstationarity detection error was determined for each method. It was shown that the three methods demonstrate similar performance, while the obtained errors can be reduced by their joint use and also by taking into account some additional statistical indexes (such as signal mean value and variance).