HRV features, Recovery Points, and more

Blog post by Marco Altini.

In this post I'll cover a few aspects related to HRV features and the different methods that can be used to assess parasympathetic activity from RR intervals data. In lay terms, this means trying to determine physiological stress load from our HRV measurements. In particular, I'll go over the following:

• Features most representative of parasympathetic activity
• Issues with frequency domain features
• HRV4Training Recovery Points

Features most representative of parasympathetic activity

HRV can be quantified in many different ways (check out this post for a primer). As a result, we need to make sure we are looking at the right numbers when analyzing our data.

Making a few oversimplifications, since the autonomic nervous system regulates the heart beating, we can use HRV as a proxy to autonomic function, and therefore use HRV as a way to measure how we react to stressors (e.g. a workout). 
​
In the context of monitoring training load, but also more in general when looking at the effect of different stressors on our physiological stress, both in the short term and chronically, we are interested in quantifying parasympathetic activity. The parasympathetic branch of the autonomic nervous system is the one in charge of rest functions and recovery, and therefore we are interested in monitoring it because reduced parasympathetic activity is a clear sign of increased stress and poor recovery.
​
Most apps use a feature called rMSSD to quantify parasympathetic activity (or a transformation of rMSSD, see later). rMSSD is computed as the square root of the mean squared differences of successive RR intervals. When computing rMSSD, we look at beat to beat differences, thus the rMSSD feature is associated with short term changes in the heart. Since the parasympathetic activity works at a faster rate (e.g. < 1 second) compared to sympathetic activity, rMSSD is considered a solid measure of vagal tone and parasympathetic activity, similarly to the high frequency power (HF). HF is often used as another alternative to capture these faster, high frequency variations. Not surprisingly, rMSSD and HF are typically highly correlated.

While many research studies in the context of using HRV to measure training load, recovery and fitness investigate HF as the main HRV feature of interest, recent work seems to have settled on rMSSD or ln rMSSD (see papers by Buchheit, Plews, Esco, Flatt, etc.). 

There are several advantages in using rMSSD. Being easy to compute, values can be compared across studies. Something which is almost impossible when looking at frequency domain features (see later). Also, rMSSD is time invariant, so using a shorter or longer time window still provides comparable results, as I've also shown in another blog post. Finally, multiple studies validated the reliability of rMSSD for measurements as short as 60 seconds (or even shorter), therefore making it a very practical alternative for consumer products.

Issues with frequency domain features

One of the most frequent (...) questions I receive is why I do not use frequency domain features or how we should interpret the infamous LF/HF ratio. 

What's wrong with frequency domain analysis?

HF, LF and the LF/HF ratio
While HF represents parasympathetic activity, similarly to rMSSD, things are much less clear around the low frequency power (LF). Some say LF represents sympathetic activity, others say a mix of both (most likely).

The principle behind the use of the LF/HF ratio is that since HF represents parasympathetic activity, when HF reduces the ratio increases and the higher ratio therefore the more stressed we are supposed to be. However since the function of LF is not really clear, looking at the ratio might be misleading as well.

There's another great explanation from James Heathers on ithlete's blog on why frequency domain features (especially LF) should not be used to monitor training load and recovery, check it out for more details.

Computation and comparison between studies
Frequency domain features are computed differently by everyone. There are guidelines  (see Berntson et al, 1997) but unfortunately there is no clear consensus on the actual implementation. Features are sometimes normalized and sometimes not, the ratio is obviously affected by normalizations, and most importantly since LF tries to measure low frequency components it does require a longer time window (at least 2 minutes). 

Given these issues, I typically do not even suggest using HF since rMSSD can capture the same information, consistently over time, and allow us to compare between studies and individuals, making it a better choice.

​If you are interested in knowing more about how frequency domain features are computed in HRV4Training, I reported my processing steps here.

HRV4Training Recovery Points

Given the explanations above, it should be no surprise that HRV4Training uses rMSSD to compute the HRV4Training Recovery Points. For advanced users, all features are available, but by default rMSSD is displayed in all screens except for the home screen, where only Recovery Points are available.

rMSSD values are transformed so that the values are a bit more readable and user friendly. The result is a number approximately on a scale between 1 and 10, with higher values representing higher parasympathetic activity, lower stress, better recovery (in general).

When transforming rMSSD, we do not simply scale the value, but perform a logarithmic transformation, as often reported in literature (see again a great body of work by Buchheit, Plews, Esco, Flatt, etc.). The logarithmic transformation has the additional advantage that makes measurements within a person "closer together". This trick also helps in assimilating the information, since the high measurement to measurement variation in HRV is one of the most difficult aspects to grasp (i.e. two consecutive measurements will never yield the exact same scores).

Finally, Recovery Points are also age-corrected, meaning that the well known reduction in HRV with age will not be reflected on the Recovery Points. The idea behind this correction is two-fold. First, if you were to use the app for a very long time, you should be able to compare values across years without much of the influence due to aging. Secondly, when correcting for age, values are easier to compare between individuals.

Follow @marco_alt