The goal of the frequency analysis of RR Interval data Is to identify periodic components and estimate their frequency and power. There is no single best way to accomplish this task. Choosing the most appropriate methods depends on the characteristics of the particular data under study (for example, short periods of time, longer periods of time, normal volunteers with little ectopic activity, patients with congestive failure and frequent ventricular ectopic activity) and the particular information required (e.g., entire spectrum or specific frequency bands). We will discuss some important considerations for several methods of frequency domain analysis, and also the specific advantages and disadvantages of the common analytic techniques.
The input signal and the importance of proper QRS annotation
Quality of the Input Signal. The quality of the frequency spectrum obtained is critically dependent on the quality of the input data. Often the sequence of RR intervals is obtained from an ambulatory ECG recording. High frequency noise due to poor skin-electrode contact, broken leads, or cable movement are likely to be detected and automatically labeled as QRS complexes by arrhythmia detection algorithms in computerized Halter systems.
Recorder Artifacts. Wow and flutter due to tape transport errors represent another potentially important confounding factor. This problem can be avoided by using Halter recordings containing a phase-locked timing track. When this is not possible, it is Important to analyze a test set of tapes. Halter tapes from a completely paced patient or from a QRS simulator to detect any errors introduced by the recording equipment. Often wow and flutter contribute error so minor that it can be neglected. In some instances, the spectral pattern of wow and flutter will be sufficiently distinct to allow correction of the data even if the error is substantial (16).
Annotation of QRS's. The requirement for accurate QRS annotation is more demanding for RR variability than that required when processing Halter recordings to quantify ventricular arrhythmias. It is especially important to correctly annotate QRS complexes when RR variability is small because a few ECG deflections due to noise or ectopic complexes misclassified as normal QRS complexes will produce large artifacts in either periodograms computed with fast Fourier transform algorithms or periodograms computed with autoregressive methodology. When a Halter recording is analyzed for ectopic activity, ventricular premature complexes (VPC's) usually are carefully annotated by automatic arrhythmia detectors and by technicians, but aerial premature complexes (Ape's) are often considered of less importance. For analysis of RR variability data, it is equally important to avoid labeling Ape's as normal QRS's to avoid errors in power spectra. Recognition of Ape's by QRS morphology alone is not possible, and consequently template-based analysis systems may miss them. Many Ape's labeled as normal QRS's can be corrected during review of short or long RR intervals by a technician who can use additional information, e.g., P wave morphology, to correct QRS labels ( 15). Some Halter processing software, e.g., the Columbia University editing tools, presents questionable RR intervals to technicians for review. It is possible that fully automatic algorithms can be developed to "clean" the RR Interval data by examining the clustering of RR intervals and iteratively excluding extreme outliers. Of course, such algorithms will have to be tuned to prevent It from filtering out marked sinus arrhythmia.
Dealing with Missing Data and Ectopic Complexes
Rarely will the sequence of RR intervals consist entirely of normal-to-normal (N-N) intervals, but measurements of RR variability generally are only concerned with N-N intervals. In time domain analysis, non-N-N intervals (and sometimes the subsequent N-N interval) can just be excluded from computations. However, for frequency domain analysis this is not appropriate because it is important to maintain the temporal ordering and spacing of the sequence of N-N intervals. Albrecht and Cohen (17) compared two methods of dealing with missing values (ectopic complexes) in the sequence of RR intervals prior to estimation of the HR power spectrum. One method computed the autocorrelatlon functions without making explicit assumptions about the missing RR intervals. The second "naive" method used linear interpolation (splines) to fill in the sampled RR sequence when missing data were present. They compared the ability of these two methods to provide reliable estimates of the low and high frequency power when variable (0-45%) amounts of the RR data were missing.
Surprisingly, the low and high frequency power estimates derived using the naive (splining) approach substantially outperformed the estimates obtained using the по-assumption method in simulation studies. They discuss in detail the theoretical aspects underlying this result, which we will simply summarize here: splining simply reduces the high frequency components for the duration of the spline, without substantially perturbing the low frequency components. The amount by which high frequency power is attenuated Is, in practice, quite small In most cases. An advantage of the splining approach to missing data is that the resulting data are suitable for efficient computational methods based on the fast Fourier transform (FFT). CHRONOS uses the methods developed by Albrecht and Cohen for correcting for ectopic complexes and noise prior to - computing the RR interval periodogram with FFT algorithms (17).
