2.1 ECG recording and processing

Fig 1 shows the basic steps of data collection, processing, and analysis. First 20-min 12-lead resting ECG (CardioControl Working Station, Welch Allyn, Delft, the Netherlands) was recorded in the supine position after a resting period of at least 10 minutes. Throughout the ECG recording (performed with a sampling frequency of 600Hz), subjects were asked to breathe at 15 respirations per minute (= 0.25Hz, guided by a visual metronome) to standardize the effects of respiratory rate on HRV.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Flow chart of the basic sequence of the individual processing steps for determining HRV values for MEANS and REF.

https://doi.org/10.1371/journal.pone.0304893.g001

All ECGs were processed by MEANS to obtain the location and type of the QRS complexes [11]. MEANS is a well-validated and widely used automated ECG processing and interpretation system [16, 20, 21]. A detailed description and comparison of the MEANS algorithm with other algorithms is outside the scope of our article. However, the MEANS algorithm has extensively been described and evaluated before [22, 23], showing its validity in correct classification of QRS complexes. For example, on the CSE multi-lead library containing 250 ECGs, all 2,081 QRS complexes were correctly detected, and 99.6% of the complexes were correctly typified [22].

As already mentioned, MEANS has a completely modular structure. Signal evaluation is mainly carried out serially by the modules. In addition to the input and output modules, 4 further module groups are responsible for the following tasks [11]:

• detection of QRS complexes and disturbances as well as the P waves
• typification according to waveform for the QRS or ST-T complexes
• determination of the dominant complexes for contour analysis
• segmentation of P and T waves and QRS complexes (recognition of start and end points)
• classification (rhythm, contour) of the ECG.

Afterward, all MEANS-processed ECGs were visually controlled for the correctness of determination of QRS location and type by two trained medical students supervised by an experienced cardiologist (in the following called "reference" = REF).

We differentiate three types of QRS complexes, normal sinus beats, supraventricular extrasystoles (SVES), and ventricular extrasystoles (VES). MEANS considers any abnormally typified QRS complex as ventricular, while the distinction between normal and atrial complexes is based on the timing and morphology of the preceding P waves. Artifacts were detected and classified by MEANS QRS detector, resulting from sudden baseline shifts or spikes. Finally, we group SVES, VES, and artifacts as non-sinus beats.

For each study participant, both the 20-min tachograms of the MEANS and the REF-controlled ECGs were then separately used for HRV calculation according to the following procedure. First, sixteen 5-min segments of each of the 20-min tachograms were produced by a sliding 5-min window moving over the entire tachogram in 1-minute steps. If in a 5-min MEANS or REF segment the proportion of non-sinus beats or artifacts was greater than 10%, both the MEANS and REF segments were removed. If all segments of an ECG contained more than 10% non-sinus beats, the ECG was excluded from this analysis. Additionally, the remaining segments were checked for stationarity of the RR intervals according to the reverse arrangement test at the 5% level to means and variances and removed when the stationarity z-score cut-off exceeded 1.96 [24]. Removal of the segments with non-stationary data did not lead to the exclusion of any of the study participants due to a lack of data. From the remaining segments, the segment of the MEANS-processed ECG with the lowest percentage of non-sinus beats and the corresponding REF segment was selected for HRV analysis. If several choices were possible, the earliest segment occurring in time was chosen.

HRV parameters were calculated for each of the MEANS-analyzed and REF-edited 5-min segments, independent of the selection procedure described above. In each of these segments, first, the non-sinus beats were replaced by interpolated sinus beats, inserted at 50% of the interval between the sinus beats preceding and following the non-sinus beats, thus creating a pseudo-sinus-rhythm interval series. Then, like in Bootsma et al., 2003 [25], the HRV spectrum was calculated from this non-equispaced interval data series, without resampling, as follows:

• normalization of the interval series by dividing the intervals by the average interval
• subtraction of the linear trend in the normalized interval series (creates a series of interval fluctuations around zero)
• cosine tapering of the initial and terminal 10% of this data series (creates a fade-in and fade-out effect, thus preventing an unwanted invasion of the spectrum by data-on and data-off transients)
• adding zeroes to the normalized, detrended, and tapered interval fluctuation series till the number of data in that series equals a power of 2 (a requirement for the FFT algorithm)
• computation of the FFT spectrum of this data series
• correction of the frequency spectrum for the energy loss as a consequence of the cosine tapering and zero padding, by multiplying it by a factor larger than 1 in order to make the total power of the frequency spectrum identical to the variance of the original time series (thus satisfying Parseval’s theorem).

Standard time-domain (SDNN in ms), pNN50 (percentage of successive RR intervals that differ by more than 50ms), RMSSD (root mean square of successive RR interval differences in ms) and frequency-domain parameters (LF (absolute power of the low-frequency band (0.04–0.15Hz) in ms²), HF (absolute power of the high-frequency band (0.15–0.4Hz) in ms²), LF/HF (ratio of LF-to-HF power)) were computed according to the method used in previous studies [25–28] and in compliance with the current guidelines for analysis of HRV [29].

2.2 Statistical methods

Performance analyses of MEANS regarding the correct detection of non-sinus beats were done using the entire 20-min ECGs. The sensitivity and specificity of the non-sinus beat detection performance of MEANS were determined according to the AAMI guidelines [8]. Two types of aggregated statistics are given: in the "gross statistics", all beats are weighted equally, regardless of whether they originate from one or the other ECG. The dependency of beats from one ECG is therefore not taken into account. This is in contrast to the average statistics, in which the dependency of beats from an ECG is taken into account and beat statistics are calculated individually for each ECG and then averaged.

Bland-Altman plots were used to visualize the differences between the HRV measures resulting from the different editing methods. As the differences did not seem to be normally distributed we took the logarithmically transformed data [30].

To evaluate the influence of misclassified beats on the HRV parameters, we used linear regression models with the percentage of false sinus beats as the independent variable and the relative HRV difference between MEANS and REF as the dependent variable (here the non-logarithmized data were used):

The association between selected CVD risk factors and the logarithmically transformed HRV values was assessed by linear regression models for both editing methods. Analyzed CVD risk factors were sex, age, previous myocardial infarction (MI), and prevalent type 2 diabetes mellitus. In the case of age, a quadratic term was additionally taken into account to reflect the non-linear relationship between age and HRV in our study [18]. Beta values and 95% confidence intervals (CI) for the correlation of the logarithmized HRV values of the respective editing method were obtained directly from the data, whereas the betas and the CI of the difference were obtained by bootstrapping beforehand, utilizing the non-parametric bootstrapping method with 10,000 resamples and a fixed seed. Nonparametric bootstrapping allows reliable estimation of statistical parameters, especially CI, without model assumptions. It is a kind of resampling procedure. Samples of the same size as the original sample are taken from the distribution of values of interest, with replacement. In each of these new samples, the statistics of interest are calculated, and the CI sought is estimated from the distribution of these results. The accuracy of this estimate depends on the quality of the initial sample. We use the percentile approach to calculate the CI, which is a very simple method but gives good results for approximately symmetric distributions [31].

Of the 1,779 participants of the CARLA study, 1,674 had an evaluable 20-min ECG (non-sinus rhythm or pacemaker subjects were excluded, n = 105), which was analyzed by both MEANS and REF. Further 33 ECGs were excluded because of more than 10% non-sinus beats in all 5-min segments, resulting in 1,641 participants remaining for evaluation of the 5-minute HRV.

All statistical analyses were performed using SAS 9.4 University Edition [32].

4.1 Performance of MEANS in discriminating sinus from non-sinus beats

We found that compared to the reference standard, MEANS classified 99% of all beats correctly. However, the average sensitivity of MEANS across all ECGs to detect SVES was low.

In their survey, Luz et al. [7] list results of sensitivity for SVES detection from different studies between 60 and 91%. In our study, we found a global sensitivity (calculated without considering individual ECGs) for SVES detection of 38%. It is therefore obvious that MEANS has some deficiency in detecting SVES (and artifacts) in single ECGs of our cohort. Whereas events classified as artifacts were mostly in reality sinus beats (in this case, the HRV values usually are not dramatically altered), the SVES were often recognized as sinus beats by MEANS (and could thus influence HRV values). Such misclassifications may have been caused by difficulties of MEANS in identifying deformed or low-amplitude P-waves. The recognition of SVES is one of the most compelling challenges for the developers of automatic algorithms at all [33].

As outlined by AAMI [8], we compared not only the global (i.e. gross statistics) but also the average performance data. In our study, average performance is noticeably higher than the global, but there is a lack of published data to compare our results with those in the literature. The difference between global and average statistics in our data is likely caused by the fact that the majority of false detections occurred in only a few ECGs. This means that in a study with participants from the general population, average performance may be more informative than global performance. For SVES, we observed average sensitivity (calculated first within each ECG and then averaged) of 61%. Huang et al. [33] reported an average sensitivity of 62,8%, and Chazal et al. [34] 25,7%, which can be interpreted as a result of differences in performance, but may also be caused by differences in the study population.

4.2 Impact of QRS misclassifications on HRV

The fact that misclassifications of beats in the ECG cause distorted HRV values is quite common [9, 10, 35]. However, the strength of the impact was little known until recently. Few recent studies are showing that time domain parameters are more robust with respect to ectopy and artifacts than frequency domain parameters [36–39]. In contrast, Bourdillon et al. found that RMSSD and SDNN are much more sensitive to a single artifact than LF and HF [40]. Our results support the finding, that time domain parameters are more robust. Reasons for different findings, among others, could be differences in the study population and its characteristics (e.g., numerical size, age, health status) and ECG recording conditions (e.g., supine or standing, breathing conditions, short- or long-term).

Our population is from an urban elderly general population and is quite large in number by comparison. For HRV, we examine 5-min segments, i.e., short-term. We do not look at the number of ectopy or artifacts itself in the ECG but at the amount (in % related to all beats of the segment) of beats falsely classified as sinus by our algorithm MEANS. Ectopic beats or artifacts are distributed quite unevenly across all ECGs. We consider each beat separately and do not take into account whether, for example, there is a burst of artifacts, as it is typical for motion artifacts.

Although the conclusion that in general frequency parameters are less robust to false sinus beats than time domain parameters, in our and other studies, the quotient of LF and HF seems less vulnerable to false sinus beats. However, despite that LF/HF is sometimes considered to be a measure of sympathovagal balance [41], current literature doubts this concept [25, 42–44]. It is still unclear to which physiological and psychological effects of the sympathovagal system LF/HF corresponds in detail. Furthermore, it seems to be problematic from a mathematical point of view to use a quotient of two measured quantities, which are separately very susceptible to false sinus beats.

Similar to Stapelberg et al. [38], we also observed a large spread in the relative change of HRV values. Only one or a few misclassified sinus beats can cause large changes in HRV values. In contrast, many false sinus beats may have a small effect on HRV. This could be the case if artifacts or SVES hardly disturb the sinus rhythm and thus have only a minor impact on the tachogram. On the other hand, there are examples where a single false sinus beat has strong effects, especially on RMSSD and HF.

4.3 Impact of QRS misclassifications on the association between HRV and CVD risk factors

We also investigated the influence of the beat classification method on the association of CVD risk factors with HRV. To our knowledge, this has never been done before. We found that using MEANS without an additional visual control does not substantially change the association between HRV (here SDNN, RMSSD, LF, HF, and quotient of LF and HF) and common CVD risk factors.

Assuming a linear relationship, it is easy to get a slope estimator for the single association. For the difference of these two estimators for both methods, we had no distributional information. Therefore, the non-parametric bootstrap and percentile method were applied [31]. To successfully implement these methods, some requirements have to be fulfilled. The most important assumption of the bootstrap procedure is that the sample must represent the population well. Since the bootstrap procedure uses only the information from the sample, the data must have enough variation in each variable to adequately represent the population. The percentile approach, which is used for computing CI, requires a not heavily skewed and approximately symmetrical probability distribution of the variable. The CARLA cohort fulfills these requirements because it is a large representative sample that reflects the characteristics of the underlying population with little selection bias, and the (logarithmized) HRV indices appear to be approximately symmetrically distributed.

4.4 Limitations

Our study has some limitations. Our reference was not based on the highest possible level of expertise. Instead, two trained students generated the reference standard (every ECG was assessed by one student only, without a cross-control) while being supervised by an experienced cardiologist. However, the agreement of the ratings of both students with MEANS was very similar, so that we can exclude substantial interrater differences. Furthermore, since our reference standard was a combination of a computer algorithm and visual reading, we expect even a higher validity of this reference than for an exclusively visual reading [45].

Furthermore, the within- and between-subject reliability of HRV estimates can be poor as it is affected by a range of factors during ECG recording [46]. However, in the present study, the conditions during the ECG recordings were highly controlled and standardized; thus, the measurement protocol is likely to have contributed to the improved reliability of the data. HRV parameters derived from ECGs with different durations and recording conditions are hardly comparable. Hence, we cannot extend our conclusions to setups with for example 24-h-ECGs. Additionally, we concentrated on the most crucial time- and frequency-domain parameters and did not consider others, like non-parametric parameters. Lastly, we have evaluated only one beat typification algorithm, and, therefore, our results in detail cannot be applied to every ECG processing program.