Session S82.5

Editing RR Series and Computation of Long-Term Scaling Parameters

R Sassi*, LT Mainardi

Università di Milano
Milano, Italy

Long term analysis of Heart Rate Variability (HRV) obtained from 24-hour ambulatory ECG recordings was shown to provide prognostic information, in particular for post-infarction risk stratification. But movements, sweating and electrode detachments may corrupt the ECG signal and introduce long gaps in the RR sequences. While ectopic beats are usually treated with linear interpolation, in 24-hours recordings, gaps a few minutes long can only be avoided by excluding completely the corrupted ECG from the study, unless editing techniques are employed to replace the missing intervals. While the wealth of studies on time series correction for the subsequent application of linear parameters is large, when considering long-term scaling exponents, the impact of corrections is still debated and often overlooked.
In this work, three methods and their correspondent long term scaling parameters were considered: DFA, alpha (the slope of the spectrum at low frequency) and Dispersional Analysis (DA). Starting from a collection of 68 high quality Holter recordings collected from normal subjects (age range: 20 to 90 years), we first built the probability density function of the size of the gaps typically encountered. Most of the gaps involved two consecutive RR intervals but about 0.9% of the runs were longer than 20 consecutive intervals. Then, the 20 longest NN segments fulfilling the following two criteria were identified: i) the series was composed only of NN intervals; ii) each NN interval didn’t change more than 20% with respect to the preceding NN interval. The mean length of the 20 reference sequences was 19213±7017 beats.
In each of the reference series, gaps were artificially inserted drawing the correspondent lengths from the estimated density function. The gaps were then corrected via i) substitution of an average value (M); ii) linear interpolation (LI) and iii) deletion (D, i.e. the gap was removed and subsequent beats were shifted down in the sequence) .The scaling exponents were computed before and after the cancellations took place ("bias"), recording the difference between the two values. The procedure was then repeated to obtain a statistical relationship between the bias and the percentage of missing points. For corrections up to 5% of the total number of points, the mean difference in the value of the scaling exponents was small (<< 0.5% for M and LI and < 1 % for D) and in most cases not significantly different from zero (paired t-test). Finally, the variability of the estimators on gaps-free series was estimated by repeating the same computational procedure but removing samples only on the head or tail of the reference series. While M and LI seemed to perform better than D, once we considered the variability of the estimators, the three methods appeared equivalent. The simulations suggest a negligible effect of the corrections on the value of the scaling parameters, at least as long as the number of points edited is relatively small.

(Abstract Control Number: 216)