On the use of the Akaike Information Criterion in AR spectral analysis of cardiovascular variability signals: a case report study

In order to better understand the relationship between model order selection by the Akaike Information Criterion (AIC) and AR spectral fidelity in the analysis of cardiovascular variability signals, the authors selected a set of signals (pulse interval invasive and noninvasive systolic and diastolic arterial pressure) characterized by having major oscillations easily recognizable by eye and clearly detected by the Blackman-Tukey spectral estimator. After model order estimation by the AIC and AR spectral estimation, the authors checked if the main spectral features were actually reproduced. The sample records were chosen 3 min long. In most cases the AR estimator failed in resolving spectral components in the low frequency region (<0.15 Hz) of the spectrum; forcing the model order to large values (>18), all missed peaks were always resolved. The authors' results suggest that AR spectral estimation based on automatic selection of model order by the AIC may lead to oversmoothed spectra, masking some oscillatory characteristics of cardiovascular variability signals.<>