Autoregressive spectral models of heart rate variability. Practical issues

Autoregressive time series model-based spectral estimates of heart period sequences can provide a parsimonious and visually attractive representation of the dynamics of interbeat intervals. While a corollary to Wold's decomposition theorem implies that the discrete Fourier periodogram spectral estimate and the autoregressive spectral estimate converge asymptotically, there are practical differences between the two approaches when applied to short blocks of data. Autoregressive spectra can achieve good frequency resolution and excellent statistical stability on short segments of heart period data of sinus origin. However, the order of the autoregressive model (number of free parameters to be estimated) must be explicitly chosen, a decision that influences the trade-off of frequency resolution with statistical stability. Akaike's Information Criterion (AIC), an information-theoretic rule for picking the optimum order, is sensitive to the aggregate amount of data in the analysis. Thus, the best model order for estimating the spectrum of a 4-minute segment of data will generally be lower than the best order for estimating an hourly spectrum based on averaging 15 4-minute spectra. A major advantage of the autoregressive model approach to spectral analysis is the ease with which it can be extended to handle messy data frequently seen in heart rate variability studies. A number of autoregressive-based robust-resistant techniques are available for the analysis of heart period sequences that contain a high volume of nonsinus and other unusual beats intervals. A theoretically satisfying framework is also available for spectral analysis of unevenly sampled data and missing data.