Part 2- DFA a1 and Zone 1 limits - the effect of Kubios artifact correction
Part 3 - HRV artifact effects on DFA a1 using alternate software
Part 3 - HRV artifact effects on DFA a1 using alternate software
HRV analysis is a heavily researched subject which has been touched on in some of the previous posts on this blog. However, one aspect of the RR interval interpretation sometimes overlooked is the way artifacts are corrected. In this post we will briefly examine artifact types, the impact on HRV interpretation but most importantly will come to realize that avoidance of artifacts is much more preferable than correcting them. In the same way as a master chef attempts to compensate for inferior recipe ingredients, the HRV software will try to compensate for various RR artifacts to still produce a superior end product. Unfortunately, as with the chef not being able to produce the high quality meal from inferior material, the HRV analysis may also be tainted by artifact, despite correction algorithms.
Types of artifacts:
The top tracing shows a missed RR interval, middle tracing shows a misinterpreted RR peak and the bottom shows a section of noise masking any usable data.
Another study classified them in a more detailed fashion:
- The type 4 artifact is the missed RR peak so the RR length is effectively doubled.
- The missed beat (type 4) is also the most common seen according to that study:
Does the missing of beats matter?
Before attempting to answer this question, lets first look at the "hierarchy" of RR recording device quality. The most basic heart rate monitor attribute we generally think of is one of simple heart rate tracking. Does the monitor actually measure the heart rate under challenging conditions? As discussed in a previous post, optical, wrist worn HRM are not even able to track basic heart rate with precision during intense exercise. Some forehead based optical devices track fairly well (Moov sweat), but some do not (Polar OH1). Since we need absolute precision in where the RR peak is, the "Plethysmographic" sensor approach with optical methods is not going to work for us.
That leaves us with chest belts like the Polar H10, ECG capable garments like the Hexoskin or using a dedicated ambulatory ECG monitor.
One recent paper pitted the mobile Schiller AR12 ECG system against the Polar H10 during various activities.
They found that the AR12 ECG system had many more artifacts than the Polar H10:
The total RR interval artifacts while jogging were a whopping 12% with the AR12 system and .72% with the Polar. One take away lesson from this is that the AR12 system can't handle activity. The major question one is left with is if the H10, with near 1% artifacts is ideal for HRV purposes. There is no doubt that it is fine for rate tracking, but does even a low artifact rate effect HRV?
Avoidance vs Correction:
Probably the most definitive evidence that correction of artifacts can't fix some HRV parameters is the study by Giles and Draper. I would strongly recommend reading their paper.
The purpose of the study:
However, the methods that are used forThe subjects performed a VO2 max test on a treadmill wearing both a Polar H7 and a set of ECG leads.
the correction of artefacts in recent literature vary considerably.
Although several studies have used various interpolation
techniques (8,12), others simply deleted offending
intervals (20) or, most frequently, do not mention correction
at all (13,14,31). Equally, research that has specifically examined
the validity of HRMs has used several methods (4,32).
As such, there is a need for standard practice in the collection
and processing of RR interval data recorded using
HRMs in research as, currently, there are no standard criteria.
Therefore, this study aimed to examine current methods
for error correction, along with the extent of alterations in
artefact occurrence during exercise, in comparison with
simultaneously recorded ECG
Artifact correction techniques:
Various methods were used including the Kubios formulas -
The methods used for artefact correction were (a) Uncorrected,As noted in the table above, type 4 errors were the most prevalent, almost 97% of the total.
no correction applied to any intervals; data were left as
recorded. (b) Deletion, erroneous RR interval(s) were simply
deleted from the time series. Deletion can have a significant
effect on HRV parameters because of changes in the length
of the signal, particularly in short-term recordings and
frequency-domain parameters (23,26). Deletion introduces
step-like shapes into the RR interval time series resulting
in changes in variability as well as decreasing the length of
the signal, producing false low frequency (LF) and high frequency
(HF) components (26). Interpolation: Interpolation
methods, in contrast to the deletion, replace erroneous nonnormal
RR intervals with interpolated intervals. Critically,
interpolation allows for the length of the recording to remain
the same, mitigating the issue of reduced signal length. (c)
Degree Zero (average), substitution of artefact(s) with a mean
value that is calculated from surrounding RR intervals. On
longer sections of artefacts, degree zero interpolation results
in the same averaged value over a whole segment, resulting
in a flat shape, introducing false trends, and increasing LF
and very low frequency (VLF) power (5,26). (d) Degree One
(linear), a straight line is fitted over the irregular intervals to
obtain new values. As with degree zero, on longer sections of
artefacts, slope-like shapes occur, introducing false trends
and potentially increasing LF and VLF components (5,26).
(e) Cubic, cubic interpolation uses 4 datum points to compute
the polynomial; there are no constraints on the derivatives.
Cubic interpolation does not result in flat sections of
data. However, as nonlinear analysis is concerned with the
complexity and irregularity of heartbeat series, the introduction
of a potentially falsely correlated signal is of concern,
particular if there is a significant number of erroneous
intervals (23,26). (f ) Spline, cubic spline interpolation.
Smooth values are estimated through a number of datum
points by fitting a third-degree polynomial. Cubic spline
interpolation computes a third-order polynomial from only
2 datum points with the additional constraint that the first
and second derivatives at the interpolation points are continuous.
As with cubic interpolation, spline interpolation
may also introduce false correlations into the signal. (g)
Kubios HRV software (Version 2.2), Kubios software provides
options for the detection and correction of ectopic
beats and artefacts
How did the error correction methods effect the HRV parameters?
For the most part pretty well as long as the effort was below 50% VO2 max and the indexes were not the non linear variety (entropy, SD1):
However as noted below, artifact correction of any kind has major difficulty in accurately correcting non linear methods like SD1, sample entropy.
Linear interpolation produced correctedThey did not look at DFA a1, but I will have more on that shortly.
intervals with the lowest bias and ES. However, values
of RMSSD, LF:HF ratio, SD1, and SampEn at 60 +
VO2max exercise intensities showed large bias and ES, and
increased LoA and reduced ICC, regardless of correction
Some points from the discussion:
Briefly, artefacts increased relative to exercise intensity, to a peak of 4.46%
during recordings made at 80–100% VO2max. Artefact correction
was necessary, with large percentage bias and ES of
HRV parameters in all but supine and standing recordings;
correction resulted in reduced bias and ES for resting and 60% VO2max recordings with all methods.
Caution should be given to the interpretation of RMSSD, LF:HF ratio,
SD1, and SampEn at high (60% + VO2max) exercise intensities,
as, even when correction methods were applied, large
amounts of bias were still present.
As most recent HRM validation studies were
conducted either at rest or during an exercise that did not
involve upper body movement (e.g., cycle ergometry), the
observed occurrence of artefacts has previously been very
low, at less than 1% (12,20,31). The criterion for the identification
of artefacts also varies considerably among studies
and, as such, they may also be significantly underreported.
Inserting linearly interpolated
intervals can create flat sections with little or no variability,
and it is advised that large sections of ectopic beats should
not be edited using these techniques
Kubios HRV software provides options for the automatic
detection and correction of ectopic beats and artefacts.
However, despite Kubios appearing to accurately detect
most artefacts, the results showed larger bias and ES
compared with both the ECG and the manually corrected
cubic spline interpolation. Both correction methods used the
same interpolation calculation in Matlab. Figure 2 shows
a short section of data with a single type 4 artefact present;
it is apparent that although Kubios software does accurately
identify the artefact (Figure 2a), the erroneous beat is only
replaced with a single interval, rather than the required 2
Personal note - Hopefully Kubios has corrected this.
As the results of thisFinally:
study demonstrate, artefact correction is necessary for RR
intervals obtained from HRMs. Correction of artefacts
with a simple linear interpolation reduced bias and ES
and increased ICC, in most but not all cases: caution
should be given to RMSSD, LF:HF ratio, SD1, and
SampEn at high (60 + VO2max) exercise intensities
Where possible, select sections of RR intervals that are
Recently the Hexoskin team has released a new improved shirt design. It is essentially a simplified version of the Astroskin garment but containing only a single lead ECG (not 3). The elastic straps along the chest and abdominal areas are gone, replaced by an integrated adjustable elastic band:
In the past I have had good luck with the Hexoskin as far as heart rate tracking is concerned. I never looked at the artifact percent of either the Hexoskin or Polar H10 but let's do so now.
The conditions of use included pre wetting the Hexoskin/H10 and applying conductive cream to all sensors.
My goal here was to do about 3 minutes at my maximal aerobic power, then reduce to just under the MLSS (lactate steady state) for training both VO2 max as well as lactate disposal. The heart rate was high, between 140 and 170 bpm. According to some previous measurements, DFA a1 complexity should have stayed near the white noise value of .5 throughout.
Here is the Polar H10 tracing with the corresponding DFA a1 (rolling 60 second window tracking, update graph points every 10 seconds of the activity). This is from Kubios V3.3, premium version with their automatic artifact correction.
The vertical lines in the top tracing are the artifact corrections
- The loss of RR complexity (DFA a1 drop below .5) occurs during the VO2 max/MAP interval of 360 watts (as it should).
- There is regain of complexity, DFA a1 rise, during the subsequent moderate to high power portion (it should not have done this).
- Artifacts (circled in black) were 1.81%.
The Hexoskin shows a much different picture:
- Although the initial DFA a1 drop is similar, the values never regain complexity throughout the complex interval. I drew a line along the .5 limit (white noise), which is far different than the Polar tracing above.
- Almost no vertical lines - perhaps 4 artifacts:
- Despite a relatively low artifact rate on the H10, automatic correction by Kubios premium V3.3, there is a major discrepancy in the non linear DFA a1 parameter.
Updated data - 11/26/19
I decide to compare the Hexoskin to the Polar H10 worn at the same time over two 7 minute intervals. The first was at 173 watts, immediately followed by 7 min at 195w.
Here is a table of both SampEn and DFA a1, along with the artifacts and correction methods:
- The DFA a1 at 195 watts derived from a zero artifact tracing (Hexoskin) is below .5, yet the moderate artifact Polar file yields values .8 or higher. The discrepancy is present at the low watt interval as well (.8 with Hexoskin vs 1.4 with Polar). I tried both the standard medium correction method as well as the auto method that comes with the premium version.
- The artifact levels encountered are within literature accepted methods. I will continue to work on getting this issue defined and resolved. The potential of failure to reproduce study results as well as having athletes train incorrectly is the obvious result.
Here the tracings are also very disparate (as pointed out by the Giles and Draper's work):
- Sample entropy is well below 1 throughout the complex interval.
- Sample entropy is close to 2, well above and almost double what the Hexoskin showed.
HF peak frequency:
Some data indicates that HF peak can be helpful as a marker of exercise load and respiratory rate. Does a small artifact rate effect this as well?
Hexoskin HF peak:
Polar H10 HF peak:
- They are both about .8 to 1.0, with the Polar having some drop outs.
- The values are much closer here indicating that artifacts for this metric are more correctable.
Theoretically, the H10 should produce comparable RR interval data but in real life, on the road the Hexoskin has been designed to be a stable 1 lead ECG monitoring device. With proper sensor contact the newer Hexoskin garment produces a noise free tracing.
Here are 2 raw extracts from the session I did for the above Kubios figures:
This was at the end of the 3 minute, 360 watt beginning segment of the above figures
This was at the end of a Wingate 60 (maximal all out 60s) later in the morning. This has always been the most challenging to get right. There is tremendous back and forth body motion, arm and torso muscular activity with a rapidly changing ECG:
- Both tracings look as if they were made from a patient at rest with a good quality ECG machine.
- If there was an "artifact" we would be able to classify it.
For the older athlete:
Perhaps the "artifacts" were not noise or missed RR intervals. Masquerading in the mass of artifacts could be a potentially severe arrhythmia, not something we would want to miss especially in the realm of aging athletes. Part of my post ride analysis is a beat by beat review of the intense intervals, looking for any sign of arrhythmia. In time perhaps, more leads will become available with the Hexoskin and ischemic change could be looked for as well.
- Artifacts in the recording of RR intervals are common.
- The most common type of artifact is that of a missing beat.
- Many correction algorithms exist to "compensate" for these artifacts.
- Most formulas will do a reasonable job at correction producing accurate HRV results with the exception of the non linear indexes.
- According to the study by Giles and Draper, RMSSD, LF:HF ratio, SD1, and
SampEn at high (60 + VO2max) exercise intensities will be problematic.
- My data indicates that DFA a1 will also be misrepresented at an artifact rate below 2%.
- The Polar H10 is an excellent device for heart rate tracking but in real world usage is prone to a low but significant artifact rate. This should be examined carefully by the researcher or end user.
- The new Hexoskin garment with integrated straps produced almost artifact free tracings even under the most challenging conditions. This device should be considered for those interested in artifact free recordings.
- It is better to avoid the artifact than correct it!
Heart rate variability during dynamic exercise
- Firstbeat VO2 estimation - valid or voodoo?
- Heart rate variability during exercise - threshold testing
- Exercise in the heat and VO2 max estimation
- DFA alpha1, HRV complexity and polarized training
- HRV artifact avoidance vs correction, getting it right the first time
- Review of Moov Sweat Forehead HRM vs Polar OH1 vs Hexoskin
- Moov HR Sweat - part 2
- Polar OH-1 accuracy on the forehead
- Analysis of Hexoskin binary RR interval and respiratory .wav data
- DFA a1 and Zone 1 limits - the effect of Kubios artifact correction
- DFA a1 vs intensity metrics via ramp vs constant power intervals