You are on page 1of 6

Comparison of methods to evaluate

decoupling in cycling
Konstantin Varbenov
Graz University of Technology
konstantin.varbenov@gmail.com

May 3, 2020
Abstract

Decoupling is used as a way of objectively measuring cardiac drift in cycling. It is used for assessing
fatigue and also, as a way of determining whether the combination of intensity and duration of a session
have been enough to provide a training stimulus. In this short paper, I examine different quantitative ways
of calculating decoupling. I start with implementing Seiler’s method to real-world data and, based on the
method’s weaknesses, develop two further ways of calculating decoupling. Also, the inherent limitations of
using decoupling are analysed.

I. Introduction and
P
Prel = , (2)
he basic idea of decoupling is compara- CP6

T tively simple. The ratio of power, mea-


sured in [W ], and heart rate in bpm,
changes over time in such a way that with
where
HRR
HR
= Heart rate reserve in %
= Current heart rate in bpm
enough time, a higher heart rate is required for HRrest = Resting heart rate
the same power. HRmax = Maximum heart rate
Let us begin by implementing Seiler’s method CP6 = Critical power for 6 minutes
[Seiler, 2020] on some real world data. P = Current power in [W ]
The idea behind the method is the following: Prel. = Relative power in %.
The theory is that as the ride progresses,
• Split the ride into 20 minute sections. HR will increase for the same power and thus
• Calculate the average heart rate and power efficiency will drop. This will become visible
for each section. in the plot: positive decoupling will occur.
• Normalize these values as shown in 1 and
2, so that both are dimensionless.
Implementation
• Calculate the ratio of HRR/Prel.
• Re-normalize these values such that the The method described above was implemented
first ratio (of the first 20 minutes) equals 1. on a real ride. During the ride, power and
• plot the efficiency over time (the ratios heart rate data were collected using consumer
mentioned above). products. The ride itself was 3:30h long, pre-
dominantly flat and without any time (only
The normalization is given by about 30s) spent above zone 2. Figure 1 shows
the power and heart rate data for the ride.
HR − HRrest The ride was analysed as described above,
HRR = (1) see Table 1. Analysis showed one weakness
HRmax − HRrest

1
Quantitative methods for decoupling assesment • May 2020

of the method: it relies very heavily on the Figure 2: 20’ Decoupling


first 20 minutes of data! This is a significant
issue in real-world data as the first 20 minutes
are usually very noisy (they include, above all,
warm-up and ride through traffic). Also, a 20
minute normalization period appears to be too
short.
To remedy these issues, I discarded the first 20
minutes and normalized over the following 1h.
That is, the efficiency in the timespan of 0:20h -
1:20h is defined as := 1.

Figure 1: Ride data

physiological pathways lead to higher effi-


ciency and alter the result.
• This method was based on indoor train-
ing. As such, it is inherently incapable
of dealing with the large variability of
power when riding outside. In particular,
downhill sections are hard to handle be-
cause power drops to 0, while HR doesn’t
go back to resting HR. Computationally,
this can be a problem, because the ratio of
It is obvious that the general trend of the power to HR will require a division by 0
decoupling 2 is indeed increasing. The second if there is a sufficiently long coasting. The
half of the ride (escpecially towards the end) same applies to pauses.
shows some noticeable decoupling of up to • Splitting the ride into 20 minute segments
20%. is a very coarse resolution and makes
However, a few limitations are visible: analysis very hard. Splitting it into 5’ sec-
tions instead introduces huge, unrealistic
• This method only applies to steady, Z1/Z2 fluctuations (decoupling of up to 70%),
rides. Because it relies on the ratio of rendering this method useless for fine-
power and heart rate, it cannot be extrap- grain analysis. Instead, the method can
olated to higher zones, where different be modified to be coninuous by using a
moving average instead of fixed segemnts.
Table 1: Seiler decoupling

Start End øPow. [W] øHR [bpm] øPow. [%] øHR [%] Eff. Norm. eff. Dec. [%]
0:20 0:40 106 126 0,29 0,50 1,74 1,17 17,14
0:40 1:00 154 136 0,42 0,56 1,35 0,91 -8,77
1:00 1:20 144 136 0,39 0,56 1,43 0,97 -3,23
1:20 1:40 156 139 0,42 0,58 1,38 0,93 -6,98
1:40 2:00 147 140 0,40 0,59 1,48 1,00 -0,42
2:00 2:20 134 139 0,36 0,58 1,59 1,07 7,37
2:20 2:40 146 143 0,40 0,60 1,52 1,03 2,78
2:40 3:00 131 141 0,35 0,59 1,68 1,13 13,26
3:00 3:20 109 136 0,29 0,56 1,90 1,28 28,31

2
Quantitative methods for decoupling assesment • May 2020

II. Methods the regression, only the data points between


the 20th and 80th minute are used: this is the
I have devised a new method of analysis that reference period, which means that for this
aims to address some of the shortcomings men- hour, decoupling is defined as zero.
tioned above. The main considerations while For the rest of the ride, HRR is predicted and
developing it were the following: compared to the real HRR. The difference is
the decoupling. Finally, decoupling is plotted
• Seiler’s method normalizes heart rate
as a 10 minute moving average1 to negate the
based on the assumption that an athlete’s
effects of short-term noise.
HRR (Heart rate reserve) ranges between
The analysis of the same ride as in the previous
the resting and maximum heart rate. How-
section is described below, see figures 3 and 4.
ever, the resting heart rate is never reached
during rides. There exists a practical lower
Several important observations should be
bound for HR which can be determined
made.
by stopping during the ride for 5 minutes
and looking at your heart rate monitor. • Firstly, 1 minute average power and heart
This heart rate will be called HR0 . When rate (at the end of this minute) correlate
normalized this way, the new HRR met- well (r = 0, 88, R2 = 0, 78).
ric (in %) will now be within the range of • The values for HRmax and HR0 don’t mat-
[0; 1]. ter mathematically when regression is per-
• For reasons that will be apparent below, formed. They are used, here, solely for
we normalize power by dividing it by CP1 , the purpose of plotting both heart rate
or the power that can be sustained for and power in the range from 0 to 1. Of
a minute. This ensures that Prel is also course, when normalization is in the form
within the range of [0; 1]. described in equation 1, the denominator
• A finer analysis than the 20’ sections shouldn’t be equal to 0 and also the frac-
is desired, that is, decoupling should tion should be positive, so that the result-
be a continuous metric, such as imple- ing decoupling also has the correct sign.
mented in the analysis software intervals • The main reason this method generalizes
[intervals.icu] better than Seiler’s is that when regression
• The metric needs to cope better with noisy is applied, information from data points
data: pauses, stopping; and also be robust with high heart rate and power are better
with its comparison period. As we saw, the preserved. When a period of 20 minutes
first 20 minutes are not an ideal choice. is averaged over, much more of this infor-
mation is lost: this is the price one pays
The method I devised goes as follows: first,
when binning before regression.
HR and power are normalized, similar to
• The inherent weaknesses of decoupling
Seiler’s method:
become more clear, they will be discussed
HR − HR0 in section III.
HRrel = (3)
HRmax − HR0

and
P
Prel = . (4)
CP1
Power is then averaged over 1 minute to mini-
1 The value of 10 minutes is somewhat arbitrary and
mize noise (hence the normalization in 4). The
subject to change. It has to be long enough to be able to
data points obtained, HRrel and Prel , are plot-
account for short descents etcc. where power is low. Also,
ted against each other and a linear regression being the same magnitude as Seiler’s 20 minutes, we get
is applied to determine their dependence. For comparable absolute values.

3
Quantitative methods for decoupling assesment • May 2020

Figure 3: Relative power vs. relative heart rate

Figure 4: Decoupling over time

4
Quantitative methods for decoupling assesment • May 2020

Several additional techniques can be applied be physically nonsensical if it contains such


to try and deal with the the major shortcoming jumps.
of the analysis: low wattage. As is visible from Another issue is that analysis is now non-
the plot in figure 4, low power quickly results intuitive: the x-axis doesn’t represent physical
in momentarily higher decoupling - note that time 2 . Finally, the physiological process of
there is always a slight offset. To try and rem- recovering during coasting/pauses cannot be
edy this, I experimented with two approaches: captured and measured.
• Filter power below a certain wattage con- As expected, a generally lower absolute
sidered to be "coasting" - here, 80 Watts. value comes as a result: when low power is
• Multiply the decoupling value with the excluded, this is a natural consequence.
averga power for the last minute. The idea
behind this was that low power will then III. Discussion
counteract the decoupling it caused.
Implementing a computational method to asses
decoupling in real world, outside cycling can
Unfortunately, both methods did not work out
be a useful tool. However, it exposes some of
to a satisfactory degree. Multiplying by the
the method’s weaknesses. They include:
average power actually only introduces noise
and doesn’t really compensate for coasting, so • the inability to properly cope with large
the method is not further pursued. power variations. This mainly includes
Filtering for power below 80 Watts is more coasting and pausing, but also any riding
interesting, so I will present it quickly. above Z2 influences the HR response and
Generally, regression works just as above, but makes data interpreting more difficult.
the data points with Prel < 0, 16 are removed. • To automate this analysis is not easy. An
This even has the effect of reducing correlation algorithm that detects the appropriate cut-
slightly to an r of 0,86. The main issue is that off point (to discard useless data at the
now, with data removed, the ride is sliced into beginning) and also the appropriate refer-
many separate chunks, which introduces an ence period (used to define the decoupling
otherwise non-existent discontinuity. This is of 0) is not easy to develop. Alternatively,
a substantial problem because parts of the ride the user of the software may individually
which were not next to each other are now
neighbours and the averaging over time may 2 Here, 20% of the data was filtered out

Figure 5: Relative power vs. relative heart rate Figure 6: Decoupling over time

5
Quantitative methods for decoupling assesment • May 2020

select those values based on some guide-


lines, as in intervals.
• Which averaging durations are used affect
the absolute values, and so making con-
clusions that inform training depends on
them. As such, these durations need to be
tailored to each athlete based on his/her
history.

The practical significance of decoupling as


a metric to asses field data will ultimately de-
pend on the success of these points.

References
[Seiler, 2020] Seiler, S. (2020). Low intensity
endurance training.
Youtube video

[intervals.icu] Cycling analysis software.


go to webpage

You might also like