Interpreting HRV trends

Blog post by Marco Altini.

Heart rate variability (HRV) trends over long periods of time (e.g. from weeks to months) are one of the most interesting and complex aspects to analyze. While day to day acute changes reflect rather well training load in the day(s) before the measurement, in the long term things get much less linear.

Often for example, a bell-shaped HRV trend has been reported for athletes following a training program of 2-3 months before a competition. Thus, HRV does not simply increases with better physical condition and fitness, but typically increases up to a point (e.g. upon reaching functional overreaching), and then decreases (e.g. during tapering) before the competition. Results from research studies have shown that optimal performance was sometimes achieved after bigger reductions in rMSSD in the week preceding the competition. Thus, relations between HRV, training load, fitness and recovery get more complex to analyze when we move beyond day to day acute changes.

In this post, I will go over recent literature analyzing HRV trends in athletes preparing for a competition, and try to extract rules that can help you better understand longer term trends. I will show some examples on my own data and how HRV4Training will automate part of the analysis and incorporate more advanced insights, directly in the app.

​The importance of trends

Day to day HRV guidance is very useful to make small changes to our overall training plan. If today we are down, we can easily adapt and move our intense training by a day, or make some other small adjustments. Day to day guidance based on acute changes can certainly help avoiding overtraining, since we know when it's a good idea not to push it. However, this kind of analysis cannot say much on the big picture. How is our overall condition? Are we at risk of overtraining or non-functional overreaching? How are we adapting to a new training phase? Analyzing trends can help you answering these questions and provide a better understanding of your overall condition.

Previous research

​Analyzing trends is not easy. State of the art research often presents case studies, i.e. an analysis of one single athlete or very few athletes over a period of a few months [1, 2]. The reason being that each athlete is different and analyzing physiological changes with respect to training load and performance is already hard enough on a single person. Given the nature of these studies it's hard to extrapolate and generalize.

​However, in this post I will try to highlight the main findings and structure them in a way that can ease trends interpretation. ​This post builds on top of a great paper published recently by Buchheit [3], in which he already provided a generalization of the main trends seen in years of his research in sports and performance. I will combine his analysis to additional findings and considerations derived from the work of Plews [1, 6], Flatt [9] and others, and my own considerations of how all of these parameters can be analyzed together.

Definitions

First of all, some definitions and clarifications:

• Trends are changes or adaptations in physiological values (e.g. heart rate, HRV) over longer periods of time, typically weeks to months, during a training program. It is important to consider trends in the context of at least training load and training phase (e.g. load/deload) [5].
• When talking about HRV I am referring to rMSSD (or ln rMSSD) baseline values, i.e. 7 days moving averages. All other variables introduced later (coefficient of variation, HRV normalized by HR, etc.) are also considered as 7 days moving averages, since here we are not particularly interested in acute changes.
• Determining accurately the HRV baseline value requires daily recordings (or at least 5-6 recordings/week). So this is the input data you will also need if you want to run the same analysis.
• The HRV coefficient of variation (CV) is a measure of the variability in our baseline values. Analyzing both the baseline rMSSD value and its variation (CV) can provide more insights compared to looking at the baseline only.
• HRV normalized by heart rate will be computed as rMSSD / AVNN, where AVNN is the average of the RR intervals during the same time window, as reported by [1, 3].

Findings

Research up to today is clear on at least two points: first, for trend analysis a multiparameter approach is key. Looking at HRV (rMSSD) alone will not be sufficient to understand long term trends and overall physical condition. Secondly, the relation between HRV and fitness which is sometimes observed at the cross-sectional level over a population (i.e. higher HRV being associated with higher fitness), is much less obvious when we analyze HRV changes longitudinally within one person, especially for elite athletes following a specific training program.

A multiparameter approach:
Additional variables that can be valuable are: heart rate at rest (HR), HRV coefficient of variation (CV rMSSD), HRV normalized by heart rate (rMSSD/AVNN) and of course training-related parameters, for example training load and training phase [1, 3]. For example, analyzing both rMSSD and rMSSD/AVNN can be used together to better understand fatigue, since a decrease of rMSSD alone is not necessarily to be associated with increased fatigue [6]. On the other hand, a reduction in CV rMSSD has been associated with increased risk of non-functional overreaching [1]. This makes sense since less variation might mean that there is more stress. However, there is some controversy on this point since other authors interpret a reduction in CV rMSSD as a measure of better adaptation [9]. This also makes sense since we expect more day to day changes if we are not coping very well with the training and therefore rMSSD jumps around more (e.g. we start a new training phase with high intensity loads). I think a multiparameter approach should help in making this interpretation easier, by looking at rMSSD and HR together with CV rMSSD. In another study [8], coping well with training was associated with higher rMSSD and lower HR, while not coping was associated with the opposite. Finally, reductions in rMSSD and HR can be associated with parasympathetic saturation (i.e. parasympathetic activity is actually increasing even if we see reduced rMSSD) [3]. 

HRV, fitness and training load:
Making some oversimplifications, in the average population or in recreational runners negative adaptations to training (non-functional overreaching, overtraining) are generally associated with reductions in parasympathetic activity (rMSSD), while better fitness and performance is typically associated with higher values of rMSSD [4]. However, in elite athletes, when training load is close to an individual maximal load, HRV can either be unchanged or reduced [6], while after a few weeks of reduced load it can increase again [5]. Therefore reflecting training load more than fitness. Similarly, training phases of higher intensity reduced rMSSD in elite athletes, while training phases of reduced intensity increased rMSSD [7] (e.g. tapering). Finally, optimal competition performance is often associated with a reduction in rMSSD in the week prior to competition (bell-shaped curve mentioned above) [3, 8], while small variations in rMSSD or HRV/AVNN were sometimes associated with sub-optimal performance [8].

Key parameters + example analysis

We've seen in literature that looking at multiple parameters together is a must in the context of trying to understand long term trends. In particular, I will consider the following, together with a basic interpretation I derived from what is discussed above:

• Physiological parameters:
  • ln rMSSD: an increase is typically associated with coping well with training and improved fitness level. A reduction is not necessarily bad, it could be associated to parasympathetic saturation or tapering. Looking at HR and rMSSD/AVNN can help figuring out the situation. Most importantly, rMSSD should always be considered in the context of a specific training phase (i.e. reductions during intense blocks are normal, rMSSD should then re-increase with periods of reduced load).
  • HR: in general an increase is associated with more fatigue, unless it is occurring during tapering. A reduction is most of the times associated with coping well with training and better fitness.
  • ln rMSSD/AVNN: big increases associated with higher HR could show maladaptations to training, while no changes or small reductions are preferable. 
  • CV rMSSD: A decrease associated with higher rMSSD and lower HR can be representative of coping well with training, while a decrease associated with lower rMSSD is probably representative of risk of non-functional overreaching. An increase in CV might reflect some trouble in adapting to the a training block (higher intensity) and if associated with reduced rMSSD might be a warning sign of inappropriate training load.
• Training parameters:
  • Training load.
  • Performance.

Let's look at these parameters for two months of my own data, to get a better grasp of how they look like and how they change over time in response to training load and fitness. During these two months, I improved consistently my condition, while preparing for a half marathon.

The last plot shows a surrogate of my fitness level, i.e. the best time over 5 km during a given week. To determine my best 5 km time, I used GPX files downloaded from a Garmin watch, and automatically analyzed all 5 Km segments within each training to determine my best 5 km pace during a week (i.e. I did not actually perform any specific 5 km test). From the plot you can see my condition seems to be improving over time, as my time 5 km time reduces.

 
Baseline rMSSD shows quite some fluctuations during the first month, as I try to adapt to training cycles (load/deload), before flattening towards the end of the 2 months.



Baseline HR seems to be more representative of my fitness level, with a consistent decrease between February and March, after an initial phase of adaptation to training loads I was not used to.


The CV typically reduces in two cases: 1) fatigue 2) coping well with training. The reduction towards the end here is due to good adaptation to training, since it comes with higher rMSSD and lower HR.

rMSSD normalized by average RR (rMSSD/AVNN) does not show much changes, except for high initial values possibly representative of some trouble starting the training plan.

Training volume reflects my training cycle(s), with a main load phase between end of January and beginning of February followed by a deload the week after.

Training performance assessed as best 5 km pace over a week shows improvements over time (reduced time).

---

By looking at the rMSSD plot and the training volume plot, we can see how increased training volume is typically followed by reduced rMSSD. rMSSD increases afterwards as soon as training load is reduced.

We can also see how the CV is higher at the beginning, due to the fact that I am starting a new training plan and I need to adjust. Then, the CV lowers for two weeks and increases again during the central phase of the training plan. The central phase is the most intense, including more and longer interval trainings. I clearly seem to suffer during this part (higher HR, reduced rMSSD, higher CV). Finally, in the final part of the training program I reduced intensity and coped better (higher rMSSD, lower HR, reduced CV).

What can we derive by trying to automate this analysis?

Automating the analysis

The trends analysis can be automated by detecting non-trivial changes in trends in each parameter separately and then analyzing how the different parameters change with respect to the training program. 

To determine non-trivial changes, I computed the slope of the relation between time and the parameter of interest (e.g. rMSSD) over time windows of 2 weeks. Then I considered only slopes greater than 1 standard deviation from the mean, to make sure we are excluding non-trivial trends. Here is an example, where the arrows point to the actual change, while the colored segment is after the change has been detected:

We can see that after an initialization time the trend detection highlights a positive trend in rMMSD and negative in HR (yellow and red respectively). Then, a negative trend is detected also on the baseline rMSSD (in red), followed by an increase in HR (positive trend in yellow). The remaining changes in HR are trivial, and not highlighted, while another increase in rMSSD is detected around the week of February 15th. Note that the trends are annotated after they happened, as indicated by the arrows.

After automatically detecting the trends, we can combine them using a set of rules derived from literature, and try to understand more about our physiological responses to training programs in the long term.

In particular, the algorithm I implemented uses as input all parameters listed above (ln rMSSD, HR, CV rMSSD, ln rMSSD/AVNN), to determine if the trend belongs to one of the following conditions:

• No relevant trends: values are fluctuating normally but without consistent changes.
• Coping well with training: typically associated with unchanged or increased rMSSD and unchanged or decreased HR, together with lower CV.
• Maladaptation to training: associated with increased HR and CV.
• Accumulated fatigue: decreased rMSSD and increased HR, with reduced CV for the higher risk cases. Even more important to analyze in the context of a training program (similar trends could show up during tapering).
• Saturation: reduced HRV and HR, in particular for individual with low HR / athletes. Important to analyze in the context of a training program and training history.

Some of these conditions can have similar parameter values even across multiple parameters, which is why contextual information about training load and phases is necessary. The latest HRV4Training version lets you annotate training phase as load/deload, so that this analysis can be further automated or tuned by the user.

Let's have a look at the automated trends analysis output for the data above:

rMSSD for non-trivial trends is highlighted in yellow (increases) and red (decreases) and shows that over these 2 months of data three trends were automatically detected. The trends are associated to the data right before the highlighted segment starts (see arrows in previous image).

For HR data, an initial decrease as well as a subsequent significant increase have been detected.

The CV in this case can be used to discriminate between maladaptation to training, that we expect to be raising CV, and periods of fatigue, which might be associated with reduced CV. As we can see below in this case we associate the rMSSD, HR and CV trends to maladaptation to the very intense training phase.

The rMSSD/AVNN ratio does not tell us much in this specific example.

The automated trend analysis shows that for almost the entire training program I was adapting and coping well with training. This can be seen also by the performance plot, showing improvements in performance over time.

However, during the main training cycle consisting of more intense interval trainings, I had consistent physiological changes representative of maladaptation to training (higher HR, lower rMSSD, higher CV). This is rather normal for someone that is not doing intense trainings too often and did not cause major issues later on (no fatigue or decrease in performance). However, I probably should've slightly reduced the intensity of my trainings.

​In the final plot below we can appreciate better the detected trends (detected using physiological data only), in the context of all trainings performed during the two months considered:

Again, we can see how the system detects a maladaptation as soon as I start a higher volume and higher intensity training cycle. Including the periods of "no relevant trends", meaning that my physiology was probably more at risk of maladaptation even though not showing consistent trends yet, this phase lasted about 10 days. This kind of information, when provided timely, can be of great help to make additional adjustments to our training plan beyond what we can do already by analyzing day to day acute changes in HRV.

Summary

In this post we've introduced the main parameters and findings reported in literature in the context of analyzing physiological data (HR/HRV) for longer term trends. By performing tends analysis, we can better understand the big picture and how different training phases and loads are affecting our physiology beyond acute day to day changes in HRV.

I've also shown how non-trivial trends can be extracted automatically from the data for each parameter, and how by combining detected trends we can derive more insights from longitudinal HR and HRV data, for example in the context of better understanding if we are coping well with a specific training program/phase, or if we are having maladaptations or approaching fatigue.

This analysis is meant to be used as an additional aid for a user or coach, and should be considered in the context of all other non-autonomic (e.g. muscular fatigue) and subjective (e.g. stress, sleep) parameters. Most importantly, the specific training program and training phase need to be considered when analyzing trends. ​By tagging your trainings in HRV4Training (training intensity and phase) you will provide more context around your current training program, and help automating this analysis.

References

[1] Plews, D. J., Laursen, P. B., Kilding, A. E., & Buchheit, M. (2012). Heart rate variability in elite triathletes, is variation in variability the key to effective training? A case comparison. European journal of applied physiology, 112(11), 3729-3741.
[2] ​Stanley, Jamie, Shaun D’Auria, and Martin Buchheit. "Cardiac Parasympathetic Activity and Race Performance: An Elite Triathlete Case Study." IJSPP 10.4 (2015).
[3] Buchheit, M. (2014). Monitoring training status with HR measures: do all roads lead to Rome?. Frontiers in physiology, 5. Chicago
[4] Buchheit, M., Chivot, A., Parouty, J., Mercier, D., Al Haddad, H., Laursen, P. B., & Ahmaidi, S. (2010). Monitoring endurance running performance using cardiac parasympathetic function. European journal of applied physiology, 108(6), 1153-1167. 
[5] ​Pichot, V., Roche, F., Gaspoz, J. M., Enjolras, F., Antoniadis, A., Minini, P., ... & Barthelemy, J. C. (2000). Relation between heart rate variability and training load in middle-distance runners. Medicine and science in sports and exercise,32(10), 1729-1736.
[6] ​Plews, D. J., Laursen, P. B., Stanley, J., Kilding, A. E., & Buchheit, M. (2013). Training adaptation and heart rate variability in elite endurance athletes: Opening the door to effective monitoring. Sports Medicine, 43(9), 773-781.
[7] ​Plews, D. J., Laursen, P. B., Kilding, A. E., & Buchheit, M. (2014). Heart Rate Variability and Training Intensity Distribution in Elite Rowers. International journal of sports physiology and performance.
[8] ​Stanley, J., D’Auria, S., & Buchheit, M. (2015). Cardiac Parasympathetic Activity and Race Performance: An Elite Triathlete Case Study. IJSPP, 10(4).
[9] Andrew Flatt blog. http://hrvtraining.com/2015/05/19/hrv-stress-and-training-adaptation/