Optical PPG and DFA a1 - an intriguing look

We have always assumed that optical PPG derived HRV (measured as pulse to pulse intervals, PPI, the analog to the RR interval seen with ECG) can't be used as a substrate for DFA a1 calculations. This has been based on both the (usual) low sample rate of the PPG sensor, along with the inherent noise from the optical pathway (wrist/arm motion). However, with the release of the Polar OH1 and now the Polar Verity Sense, we have a PPG device with a sample rate of 135 Hz. Before we dismiss PPG as a measuring technology for HRV, remember that it probably is the major player in resting HRV across platforms used for sleep dynamics and in HRV readiness/fatigue assessments (e.g., Apple, Garmin, Fitbit watch, etc.). There are innumerable reports published, validating PPG linear HRV (PPI vs RR ECG as SDNN or RMSSD) which are easy enough to search for. However, no study to date has looked at the Verity Sense in particular and ECG related HRV validation. The closest I could find was that of the OH-1, which appears to have the same sensor array and PPG sample rate. The agreement to standards (the H10) was decent for SDNN and RMSSD.

With that in mind, I decided to take a look at the Verity Sense vs ECG in terms of DFA a1 behavior and had some unexpected results.

Here is an example of the PPG signal recorded with Kubios mobile at rest on the forehead. We see the waveform is not "ECG sharp" but generally well defined, with little noise.

 

 

How did I test the PPG against the ECG?

The Movesense Medical ECG ("MS") in chest belt position, (sample rate 500 Hz) was used along with the Verity Sense (VS) in the "forehead" position. The advantage of the forehead is the absolute lack of any muscular artifact or distortion in the pulse waveform. Here is how the waveform looked while on a stationary bike at low-moderate intensity (using the Polar Sensor Logger app):

As a comparison, here is the waveform tracing with the sensor on the upper arm, tricep position, as recommended by Polar:

• Although it's not bad, there is definite variation in signal to noise and wavering baseline.

Initially, I did a 90-minute session in the high zone 1 intensity range, just under the LT1/VT1/HRVT1 power.

MS medical sensor over the 90 minutes (full 90 min interval in a single measurement window):

 

VS PPI over the identical timespan:

 

• What troubled me in particular was the very rapid drop in a1 with the VS recording, despite barely any effort in the early warmup. 
• This was followed by 90 minutes of sub 0.5 DFA a1 with only zone 1 intensity.

 

• We see very steady HR throughout, which did match the ECG data.
• Most of the a1 values are below 0.5 (the red line is a1 = 0.5)

Therefore, it appeared that values at rest modestly matched the ECG, but if any load was applied, DFA a1 would immediately drop to the "random/uncorrelated" range of 0.5 or lower (anticorrelated). Hence, I planned a session where DFA a1 would be near 1.0 (or higher) during the exercise (as measured by ECG). 

This next session was 50 minutes at an intensity well below LT1/HRVT1, somewhat higher than a brisk walk would entail.

Before presenting the DFA a1 results, let's look at the HR and interbeat intervals (RR vs PPI).

 

With a Bland-Altman plot of :

 

• These are very strong correlations and agreements, considering the differing base technologies involved.

Here is a comparison to our study of the H10 to a reference ECG:

 

 

• Both regression and Bland-Altman differences are very similar regarding HR and interbeat intervals.

Since many studies simply assess RR intervals for device validity, the above data would "validate" the Verity Sense as an HRV device (with a larger sample size). 

Here is another example of a published study looking at the Polar 810. They only looked at frequency HRV (no RMSSD, SDNN) and certainly no DFA a1. The RRs show close agreement, especially with correction.

 

• My point here is that RR intervals may "agree" and correlate quite well, but that does not guarantee DFA a1 precision.

 

Back to my data - does the VS actually hold up for measurement of DFA a1?

Time course of DFA a1 over the low intensity session:

 

• Oh boy, that is bad!
• We can see that the DFA a1 starts and ends, close to the ECG based values (1 to 1.5), but as soon as any load is applied, the Verity Sense PPI undulates around the "uncorrelated" range of 0.5 or below. 
• HR is hovering near 115 bpm. 

Here are the regressions for comparison (there is basically no relationship seen):

 

• It appears that the VS derived DFA a1 will always be near 0.5 despite the ECG yielding values of 1 to 1.5. The correlation coefficient reflects the poor (or no) relationship as well.

What are some take home messages from this example (and why I'm doing this post)?

• A PPI/RR monitoring device with an interbeat (IBI) validation study showing good agreement to an ECG does not guarantee validity in terms of DFA a1 during exercise. At this time only the H10 and Movesense Medical ECG sensor have been validated for DFA a1. Sure, others may work as well, but why chance it? 
• PPG/PPI in this current form, does not appear capable of discerning HRV correlation patterns (DFA a1) properly, even under minimal load and a near perfect waveform. Of course, much higher sample rates and on-device software may circumvent this issue in the future. 
• A word about upsampling: One (of many) advantages of using ECG is the ability to "upsample" the raw data. Kubios does this automatically (and can't be turned off) to any ECG stream. Therefore, when we say that a 125 Hz ECG tracing is enough for DFA a1, that should be qualified with the caveat that it was upsampled (by Kubios) to 2000 Hz. This upsampling can't be done with the RR or PPI output - the xxx (ms) timing is all we have, and nothing can be done about that to enhance it. Therefore, saying that a 135 Hz PPI (from the VS) is equivalent to the 133 Hz H10 ECG is an unfair comparison. For more details on upsampling, please see https://iopscience.iop.org/article/10.1088/0967-3334/26/5/013. Perhaps in the future I will do a post about it, as the subject is very intriguing.
• A word about "resolution": If one analyzes the PPI data output from the VS, it is apparent that the numerical values are not in line with a "resolution" of 7.4 ms. In fact, they are indicative of a resolution of 1 ms, which is what we would see with a waveform sample rate of 1000 Hz. This is determined by looking at the timing values and determining what the PPI stream is quantized to (e.g., a 1 ms grid). Here is a good explanation from ChatGPT:

 

• What's obviously happening here is that there is software interpolation of the PPI signal with a falsely "optimistic" resolution of 1 ms. This is another lesson to learn about on-device "tricks" to appear more precise than they actually are. This is not a negative critique of Polar in particular, but one should be aware of this in general when using a non validated device for DFA a1 measurement. 

 

A past example of the Polar H7 may also help in emphasizing this point. We looked at the H7 bias (to ECG), and there was definite bias in DFA a1:

 

• DFA a1 was lower than ECG in a simultaneous H7 recording. This is why I would not trust the H7 or Polar 800 series for DFA a1. As Polar mentions in their white paper, resolution was improved for the H10 - Whether this means the interpolation algorithms are better and/or actual sample rate was improved is unknown. Yeah, many folks are not interested in this level of detail, but remember garbage in = garbage out - it is important if we are basing training/performance assessments on these metrics.

• Aside from DFA a1, optical PPG HRV definitely has a role in resting metrics. The correlation/agreement is strong as far as IBI/HR. The Verity Sense in the forehead position provides a clean signal with classic waveform morphology. Even the upper arm is decent under many circumstances.
• Given this clear distinction, why use the OH-1 or VS over the H10? The H10 is cheaper, able to differentiate noise from arrhythmia, and has tried-and-true use case/validation for DFA a1. Bottom line - I don't see any reason, although it is a bit more comfortable to wear while sleeping, which has nothing to do with exercise related HRV.

 The blog index....