R-R Interval Outlier Processing for Heart Rate Variability Analysis using Wearable ECG Devices

Electrocardiograms (ECGs) captured by wearable ECG devices readily contain artifacts due to measurement faults. Since artifacts and R waves have quite similar frequency characteristics, R wave misdetection or R-R interval (RRI) miscalculation may result. Aiming at accurate analysis of heart rate variability (HRV), this paper proposes a new RRI outlier processing method consisting of three steps: evaluating RRI reliability, excluding RRI outlier, and complementing missing RRI. In the first step, the method evaluates the measurement status of all detected R waves and calculates RRI reliability based on the measurement status of a combination of the measurement status of two R waves. Since we target wearable ECG devices used in non-medical environment, the method evaluates R waves based on the threshold electric potential for left ventricular hypertrophy, and determines those exceeding the threshold as artifacts. The method accordingly sets lower reliability to RRIs containing R waves evaluated as artifacts. In the second step, the method excludes all RRIs with low reliability as outliers. These steps may be effective for HRV measures in the time domain, but are not sufficient for analyzing HRV measures in the frequency domain. Resampling the time series RRI data, which is essential for analyzing HRV in the frequency domain, may produce outliers if the target RRIs contain missing values. Our method accordingly complements missing RRIs before data resampling based on RRI characteristics. We postulate that consecutive changes in RRIs follow a simple formula consisting of three components: direct current, low frequency, and high frequency. Our method complements missing values according to the formula, which is calculated from RRIs time series regarded as having been properly measured. To confirm the effectiveness of the method before applying it to ECGs recorded by wearable devices, we evaluated all the steps using pseudo-ECGs generated artificially by adding noise and artifacts to open ECG data. Initial evaluation results showed that the proposed method outperformed conventional method regarding the precision of both time and frequency domain measures of HRV.