A study on the optimum order of autoregressive models for heart rate variability

Abstract

Heart rate variability (HRV) has been used as a non-invasive marker of the activity of the autonomic nervous system and its spectrum analysis gives a measure of the sympatho-vagal balance. If short segments are used in an attempt to improve temporal resolution, autoregressive spectral estimation, where the mode] order must be estimated, is preferred. In this paper we compare four criteria for the estimation of the 'optimum' model order for an autoregressive (AR) process applied to short segments of tachograms used for HRV analysis. The criteria used were Akaike's final prediction error, Akaike's information criterion, Parzen's criterion of autoregressive transfer function and Rissanen's minimum description length method, and they were first applied to tachograms to verify (i) the range and distribution of model orders obtained and (ii) if the different techniques suggest the same model order for the same frames. The four techniques were then tested using a true AR process of known order p = 6; this verified the ability of the criteria to estimate the correct order of a true AR process and the effect, on the spectrum, of choosing a wrong model order was also investigated. It was found that all the four criteria underestimate the true AR order; specifying a fixed model order was then looked at and it is recommended that an AR order not less than p = 16, should be used for spectral analysis of short segments of tachograms.