Saturday, December 28, 2019

HRV artifact effects on DFA a1 using alternate software

After writing the previous post on the issues of Kubios artifact correction, I decided to see if alternate HRV software would have the same effects on the DFA a1 index.  My thought was that perhaps Kubios just has some computational error in their calculations.  On the other hand, the artifact corrected rise in the DFA index may simply be the result of applying a given interpolation method.  We will evaluate a well documented software package called HRVanalysis.  This was developed by a group at a French university:

They have a good instructional manual and the initial paper linked above documents the various HRV methodology very well.  In regards to error correction, they use a cubic spline approach:
In order to replace invalid beats, the software therefore provides automatic correction of the RR series, inspired by the algorithm developed by Kamath and Fallen (Kamath and Fallen, 1995). First, false beats are detected using Cheung's algorithm (Cheung, 1981): a high and low threshold are set for the relative variation in successive RR intervals (+32.5% and −24.5%, respectively). Second, for each detected error, the number of missing beats is estimated by comparing the total time duration within the error period with the duration of the immediately preceding beat. If the number of beats to recalculate is 3 or less, a cubic spline interpolation is done. This configuration generally originates from a single supraventricular or ventricular ectopic beat, or from an isolated R-peak missdetection. For 4 or more successive errors, the missing beats are interpolated by copying and inserting the same number of previous RRs between the first and last valid RR Although, these corrections generally results in a clean RR signal, this type of automatic algorithm can produce inconsistent values when the original EKG or RR signal is too corrupted, with lengthy portions of successive false beats. It is therefore recommended to visually inspect and review each series before performing HRV analyses. Additionally to RR correction, the software enables excluding parts of the signal from analysis, and the number of corrections is displayed in the results.

Although I will not go through a full tutorial of using this software, a few things should be pointed out.  In the preferences, there is an option to automatically correct artifacts.  Under most circumstances this should be checked unless one was to do beat by beat correction (which can be done).


Here is what the RR series looks line in the program:
This was from a recording of my Polar H10 indoor cycling file from the last post.  The artifacts are in red.  The boxed areas were the 5 minute intervals at 170 vs 190w as well as the last 2 minutes of each section for computation.

Here is what a cleaner file looks like with only one artifact (arrow).  The RR intervals are all about 500 msec:

In the last post, I compared 5 minutes of cycling at about 170 vs 190 watts, looking at simultaneous tracings of artifact free (Hexoskin) vs Polar H10 (1-4% artifact)
A close up of the last 2 minutes of the190w section (20w above VT1) looks like this:
  • Each blue spike upward is an artifact.  The RR series is plotted over time.  If a beat is missed, the RR gap widens, hence the blue spike of RR time.  If this is a 2 minute window with a HR of 137 bpm, there are 274 beats.  Factoring in 6 artifacts (counting the 4th blue spike twice), the percent artifact would be about 2.2% - well in the "acceptable" range.

With no artifact correction the DFA a1 was about .9 with a HR of 137 bpm:
 
  • After automatic cubic spline artifact correction, the analysis changes.
  • The DFA a1 decreases to .759:


 A 30 second earlier window (same net length) shows similar results:


How does the same sequence look with no artifacts?

The same 2 minute block of 190w cycling (about 20w above VT1 for 5 minutes) -  Hexoskin tracing with no artifacts:

No RR spikes as expected, steady RR times.

The DFA a1 and other HRV parameters came out to this:
  • The DFA a1 is .498 with a mean HR of 137 bpm.
  • If the 2 minute window is shifted 30 seconds earlier the DFA a1 result is about the same:

Summary:
  • The bottom line is that artifact correction done by either Kubios automatic, Kubios threshold or a different HRV software using cubic spline interpolation will all lead to elevation of the DFA a1 index as compared to a recording with no artifact (Hexoskin).  This seems a limitation of the artifact correction methods and not an intrinsic bug in Kubios.  
  • The HRVanalysis software is a powerful tool and in many respects comparable to the free version of Kubios.  The team responsible for it's creation and maintenance should be proud of their work.
  • As discussed in the previous post, reliance on the DFA a1 change from correlated to uncorrelated white noise values to delineate training zone 1 to 2 transition (VT1) could be of questionable accuracy with current artifact correction methods.  
  • It is beyond the scope of this brief post to speculate on what limit of artifacts are reasonable to get accurate DFA a1 results.  However, the Polar tracings in this and the previous post had 1-4% artifacts which is generally deemed acceptable.  It is clearly not.
  • The effects of artifacts on higher power intensity associated DFA a1 decline may not be as critical since we are less interested in zone 3 effects. 
DFA a1 vs intensity metrics via ramp vs constant power intervals 

Sunday, December 22, 2019

DFA a1 and Zone 1 limits - the effect of Kubios artifact correction

In several prior posts, the use of the non linear HRV index, DFA a1 was discussed as a marker of zone 1 to zone 2 transition.  This of course coincides with the VT1/LT1, however since both parameters can be subject to error, an accurate assessment of DFA a1 may be an even better training zone marker.  Also in a prior post, a warning was given about the effects of correcting artifacts in the RR data by the Kubios software.  Although the literature review was sound, the window length I used for deriving DFA a1 was not optimal.  The recommended window of measurement for this index should be 2 minutes.  I decided to do another set of comparisons using the proper window time (2 min).  To simplify the issue only 2 cycling power intervals were done, VT1 power and 15-20 watts higher.  In a perfect scenario, the DFA a1 would be about .7 at VT1 power and suppress close to .5 (white noise) with just another 15 or 20 watts added.  

The question is: Does the presence of even minimal "acceptable" artifacts create an erroneous DFA a1 result that would yield the wrong zone 1-2 demarcation?  I am using the latest version of Kubios premium and extracted the DFA a1 results computed every 20 seconds of the past 2 minutes of each activity.  



The first test was done on an indoor trainer with very steady power and little body motion.  The second was done outside on a smooth road.  In both cases, the Polar H10 belt was worn with the Hexoskin "Pro" shirt and plenty of conductive gel was used.

Here is a listing of the first test indoors with artifacts, DFA a1 values for the Polar H10 using both medium and autocorrection.  The free version will only use the medium correction method.


Which graphs out as such.
The first half is at 168 watts and the second is at 186watts:

  • The big issue is the shift to higher DFA a1 complexity with artifact correction.  This occurs with either Auto or Threshold correction methods in Kubios.  I listed the raw data so you can see for yourself.
  • The blue dots are the zero artifact values which stay remarkably stable in the .7 to .8 range at the VT1 power, with prompt suppression to .5 (white noise) just 20 watts higher. 
  • The red dots are the Polar derived reading with low artifacts.  Although there is a change in DFA from low to higher power, it is substantially muted and well above the .7 cutoff at all times.
  • The Hexoskin had zero artifacts, so no correction is needed.
  • If I was attempting to stay in zone 1, or was testing myself for the DFA a1 transition point of .7, the Polar (with artifact correction) would have misrepresented the true values.
  • In my opinion this is extremely important to know.  Inappropriate (high) exercise efforts will occur if one is relying on the DFA a1 to delineate zone 1 to 2 transition.  Reproducing these findings across sessions and individuals will not be reliable either.

How does outdoor look?
Six minute intervals were done at 170 then 190 watts on a smooth road.  This was an hour after the indoor session.
Raw data:


With the DFA a1 comparison below:
Blue dots are the artifact free Hexoskin DFA a1 values, red from the Polar with Auto artifact correction:

  • Again we see that the DFA a1 is shifted higher in the face of artifacts with Kubios correction.  This occurs with either type of correction method.
  • In the artifact free tracing, there is a sharp transition from correlated to uncorrelated (white noise) in the DFA a1 tracing with just a 20 watt increment of power (blue circle).  This does not occur in the artifact corrected tracing (red circle).

Summary:
  • Both of these examples serve as a warning.  Given the findings here, the current artifact correction methods of Kubios will increase the DFA a1 complexity to enough of a degree to misrepresent zone 1 transition and VT1.
  • I am not implying that the Polar H10 is artifact prone.  However in my situation, I rarely get clean tracings.  On the other hand, the Hexoskin is generally near artifact free, even at heart rates near 180 bpm.  
  • If one is relying on getting an accurate DFA a1 during dynamic exercise it seems essential to avoid artifacts at this time (and not rely on correction).  
  • Perhaps in the future, Kubios or other HRV software will be able to correct RR artifacts without substantial effects on DFA a1. 
  • Finally, the use of Kubios artifact correction can lead to situations where both a given individual's findings or research investigation into DFA a1 will yield erroneous results.

Part 2 - HRV artifact effects on DFA a1 using alternate software 

DFA a1 vs intensity metrics via ramp vs constant power intervals