-
Sport training consists of systematically performed exercises, in order to improve
physical abilities and acquire skills connected to a speci c sport. It has been
shown that when an athlete achieves the optimal training load, the performance
can improve while minimizing the risk of negative training e ects such as
injury [
7,15
]. The total load can be described by an External Training Load (ETL)
which is de ned by the content of the training, designed by the coaching sta .
This ETL, combined with individual characteristics such as oxygen uptake,
invokes an internal physiological stress, called an Internal Training Load (ITL) [
14
].
While the ETL can be quanti ed using di erent statistics that describe the
content of the training, such as the duration, it is not straight-forward to quantify
the ITL. Foster et al. [
9
] showed that this ITL can somewhat be assessed through
quanti cation of the training sessions' duration and intensity. The former can
be quanti ed in time, while the latter can be quanti ed using di erent methods
such as heart rate (HR) monitoring or the Rate of Perceived Exertion (RPE) [
5,
14
]. The disadvantages of HR monitoring is that on the one hand it requires
physiological expertise in order to interpret the longitudinal data produced by
the sensors [
1
]. On the other hand, it tends to underestimate the intensity during
intermittent activities, which can place a greater demand on the cardiovascular
system than steady state exercise [
15, 16
]. Therefore, RPE could serve as an
interesting alternative or addition, as it requires minimal e ort from the athlete.
This RPE is expressed as a score on a validated scale and is a simple, practical
tool that represents the athletes' own perception of training stress [
13
]. When
the RPE is assessed from a complete training, it is called the session Rate of
Perceived Exertion (sRPE).
The athlete's individual characteristics can have a impact on the perceived
ITL [
14
]. Thus, individuals within the same team may perceive an equal ETL
di erently. As an example, when an athlete has a high maximal oxygen uptake
(VO2max), the athlete may perceive a training stimulus to be lower than an
athlete with a lower VO2max [
4
]. Therefore, the coach should take the athletes'
individual characteristics into account when prescribing the ETL. As there are
often quite a lot of di erent athletes with varying characteristics to consider,
it can be very complex to personalize trainings. Furthermore, coaches tend to
underestimate the athlete's sRPE [
6
]. Therefore, a decision support system that
helps the coaches in assessing both the ETL of their prescribed training sessions
as the ITL of each of the athletes individually, through the form of an sRPE
score, can be very useful. A schematic overview of such a system is depicted in
Figure 1. This system is composed of two large components. First, there is an
application that allows the coach to design the next training in an intuitive and
user-friendly manner. Second, a machine learning module is present that deploys
three di erent predictive models:
{ A model that predicts the group sRPE, given statistics concerning the
content of a training, such as the duration and an estimated distance that the
athletes will need to traverse. With this model, the coach can get a general
idea of the training load of his/her prescribed training before it occurs.
{ A model that predicts the sRPE of each of the athletes individually
prior to training, given their individual characteristics and the content of
the training, allowing for coaching sta to tailor their training sessions on
an individual level to the characteristics of each player.
{ A model that predicts the sRPE of each athlete individually after
or during training, given their individual training statistics (such as the
actual distance the athlete traversed, how long the athlete participated to
the training, etc.) and their individual personal characteristics (such as their
age or oxygen uptake). This model allows to follow-up the training of each
athlete in a real-time fashion and could serve as some form of anomaly
detection. When we assume that the model is quite accurate in predicting
the individual sRPE scores of athletes, a high positive deviation between the
reported sRPE and predicted sRPE score could give an indication of injury.
The main ow of this system is the following. First, the coach plans his training
using an application with a user-friendly interface. Based on this plan, di erent
estimates of statistics are calculated, such as the expected duration and the
distance that the athletes will need to run. These estimates are combined with
weather forecasts and fed to the machine learning module in order to generate a
prediction of the group sRPE, which gives a general overview of the training load.
The coach can then adapt his planned training until he or she gets the expected
training load, allowing for the coaching sta to better control the intensity factor
of the ITL. Moreover, individual characteristics of each athlete, such as their
VO2max and deviation from group sRPE in the past, can be fed to another
predictive model in order to generate individual sRPE predictions, enabling
better personalization of training sessions. Finally, after training, personal training
statistics, such as the e ective distance traversed and the amount of time an
athlete participated to the training can be fed to a nal predictive model to
generate a better individual sRPE prediction. This prediction can be compared
to the score reported by the athlete. When the absolute di erence between the
two scores exceeds a certain threshold, an alert can be generated for the coach to
enable some form of anomaly detection. Moreover, this ne-grained individual
sRPE prediction can also be generated during the training, allowing for the
coaching sta to follow up the di erent athletes in a real-time fashion.
The focus of this paper is on the machine learning module of the proposed
system. The training planner application is considered out of scope. Our
contribution is two-fold: on the one hand we propose a system that can support
the coaching sta in designing their training sessions and increase the e ciency
of these sessions for athletes individually; on the other hand we present, to the
extent of our knowledge, results that are better than current state-of-the-art on
predicting the individual sRPE scores, from a di erent sport domain but with
a smaller dataset and less invasive features. The outline of the remainder of
the paper is as follows. In Section 2, related work concerning the prediction of
sRPE scores in sports, along with corresponding results is given. In Section 3, a
methodology and preliminary results for the three discussed predictive models
are discussed. Finally, we conclude our paper and give future research directions
in Section 4.
2
Related work
While quite some studies have already investigated the factors impacting the
RPE-score in di erent sport domains [10{12], only a limited amount of studies
have already attempted to predict the sRPE-scores for athletes. These studies
are solely situated in the Australian football sport domain, while our study
concerns association football data. In Bartlett et al. [
3
] a labeled dataset was
constructed consisting of 2711 training observations corresponding to 41 AFL3
players. The data consisted of six features derived from GPS sensor
measurements: the training duration, session distance, high-speed running (HSR), ratio
of HSR and distance, duration divided by session distance and PlayerLoad™ [
2
].
The labels are the sRPE-scores on Borg's CR-10 scale, indicated by the athletes
themselves [
5
]. Both Generalised Estimating Equations (GEE) and Arti cial
Neural Networks (ANN) were evaluated to predict sRPE values in combination
with two methodologies: constructing one global model, trained on all collected
samples, and constructing a model for each player individually. For the GEE, a
Root Mean Squared Error (RMSE) of 1:85 0:49 and 1:58 0:41 were reported
for the global and individual models respectively. For the ANN, a RMSE of
1:42 0:44 and 1:24 0:41 were reported. Unfortunately, our study was performed
on retrospective data, and does not include all the variables used in the study
of Bartlett et al., such as PlayerLoad™. This makes it impossible to create a
baseline with the same variables on our association football dataset.
In Carey et al. [
8
] a lot of di erent regression and classi cation models were
evaluated on a dataset of 3398 records, corresponding to 45 professional AFL
players. The labels were scores on the session RPE scale, which adapts the
descriptions of the Borg CR-10 scale [
5,9
]. A lot of di erent features have been used
derived from global positioning systems, heart-rate monitors, accelerometers and
di erent questionnaires, increasing the e ort required from athletes. The most
accurate model was a Regression Random Forest with a reported RMSE of
0:96 0:08.
3
3.1
Methodology & Results
Data Collection & Pre-processing
In order to collect labeled data, 61 di erent training sessions from one of the top
football teams in the Belgian Jupiler Pro League4 were attended from the 21st of
June 2015 until the 4th of May 2017. During each training, players were wearing
3 http://www.afl.com.au/
4 http://www.sport.be/nl/jupilerproleague/
Team Pro Polar sensors5, which measures a multitude of di erent variables, such
as speed and heart rate. Within thirty minutes after training, the sRPE (using
a slightly adapted scale from Borg, introduced by Foster et al. [
9
]) was asked
to each of the athletes individually, each time presenting the scale. In total,
913 records were collected from 45 di erent eld players. A distribution of the
di erent sRPE-values can be found in Figure 2. In order to generate a dataset for
the group sRPE prediction model, means were taken for each training, resulting
in a dataset of 61 records. The di erent variables used in the three predictive
models can be found in Table 1. We also experimented with extracting two kinds
of features from the heart rate and speed time series, generated by the Polar
Team Pro sensors. The rst kind of features were just generic features, such
as the number of peaks or the maximum value, extracted from the time series
using the tsfresh6 package in Python. The second kind of features we extracted
from (a subset of) the timeseries were shapelets, which identi es subseries that
posses a lot of discriminative power for a speci c class [
18
]. Unfortunately, the
predictive performance of the model that predicts the individual sRPE scores
after training did not improve with the incorporation of both feature categories.
This could be due to the fact that we did not have the time series for all samples,
resulting in a smaller dataset.
As depicted in Table 1, the data for the predictive model that predicts the sRPE
score for a training was generated by grouping the original dataset by training
and taking the mean of the di erent variables and sRPE values. As only 61
training sessions in total have been attended, the dataset used for this predictive
model was rather small. We evaluated di erent machine learning techniques,
5 https://www.polar.com/en/b2b_products/team_sports/team_pro
6 https://github.com/blue-yonder/tsfresh
Feature name
Temperature
Humidity
Windspeed
Visibility
Distance
Duration
Number of sprints
Training sRPE prediction
Position
Player name Age sRPE deviation stats
Clustered location
Model
group
group
group
group
Description
the temperature at the start of
the training
measured at start of training
measured at start of training
measured at start of training
group Calculated by taking the mean
of distances from a training
group Calculated by taking the mean
group Calculated by taking the mean
prior indiv generated by the training group
sRPE model
prior indiv 5-dim vector summing to four:
fcentral def, wing def/mid,
central mid, central atk, wing
atkg
prior indiv one-hot encoded
prior indiv
prior indiv the mean, std, max and min of
deviations from reported sRPE
values from the player and the
mean sRPE of that training (the
training of the current sample is
not included)
prior indiv the continent of the player
(onehot encoded)
Category
r
e
h
t
a
e
w
s
t
a
t
s
g
n
i
n
i
a
r
t
VO2max prior indiv
Speed aerobic threshold prior indiv
Speed anaerobic threshold prior indiv
Time on 5m sprint prior indiv measured at start of season
Time on 10m sprint prior indiv measured at start of season
Time on 5x10m shuttle run prior indiv measured at start of season
Height countermovement jump prior indiv measured at start of season
Muscle ber type prior indiv
Duration deviation
post indiv the di erence between the mean
duration of each player
participating to the training and the
player itself
# sprints deviation post indiv di . overall mean and own value
Distance deviation post indiv di . overall mean and own value
Average speed post indiv
Time in di erent HR zones post indiv divided in ve equally spaced
zones (HR50-60 until
HR90100+) where HR50-60
represents the time the athlete had
a heart rate between 50 and
60 percent of his/her maximal
heart rate, normalized
Dist. di erent speed zones post indiv divided in four equally spaced
zones (3-7 km/h until 19+
km/h), normalized
# pos. and neg. accelerations post indiv divided in di erent zones and
normalized
using a leave-one-out cross-validation strategy: Random Forest Regression (RF),
Decision Tree Regression (CART), Multivariate Adaptive Regression Splines
(MARS), Linear Regression (LR), Generalized Additive Models (GAM) and
Support Vector Regression (SVR). Results for each of these techniques can be
found in Figure 3. While the Support Vector Regression model performs the best
overall with a Mean Absolute Deviation (MAD) score of 0.89, the performance
of the models on each of the samples individually uctuates a lot. Therefore,
a stacking strategy was applied. In other words, the predictions from each of
the models were fed as features to a second-level model, a Random Forest. This
increased the predictive performance signi cantly (MAD and RMSE score of
respectively 0.66 and 0.81), as can be seen in Figure 3. The prediction from this
second-level model was then fed as an extra feature to the two other models that
predict the sRPE scores individually.
The coaching sta may not only be interested in getting a global indication of
the load of their planned training, but potentially would like an individualized
indication as well to allow for training personalization. One simple baseline
would be to take the average di erence between the mean training sRPE and
the individual sRPE values of a player (indicating whether or not an athlete
systematically scores higher or lower than his/her peers) and add this to the
predicted sRPE value for the training:
RP^ Eind = RP^ Egroup + dev
Applying this formula to our dataset already results in a RMSE score of 1:47,
which is competitive to the best performing global model of Bartlett et al. [
3
]
with a much smaller dataset. Of course, the model can be improved even more
by including more variables, which can be found in Table 1. It's straight-forward
that for the model that predicts individual sRPE values prior to training, only
variables can be used that are available at that time. Predictions were generated
using Generalized Additive Models (GAM). We applied 5-fold cross-validation to
the complete dataset, which is the same validation strategy used in the study of
Carey et al. [
8
]. The achieved RMSE, averaged over the ve folds, is 1:0176 0:06.
The MAD is equal to 0:7435 0:06. These results are already better than the
currently best reported results from Bartlett et al. [
3
], with a smaller dataset
and without features describing the individual ETL. It is important to note that
this comparison should be taken with a grain of salt, since the results stem from
two di erent sport domains.
3.4
Individual Training sRPE Prediction: after training
A predictive model that can accurately predict an individual sRPE score after the
training can enable some form of anomaly detection by calculating the absolute
di erence between the predicted sRPE score and the reported sRPE score of the
athlete. If this di erence exceeds a certain threshold, an alert can be generated
for the coaching sta . Moreover, this model can also be deployed during training
in order to generate sRPE scores for the players individually in a real-time
fashion, allowing the sta to better follow-up athletes. The features used in
this predictive models are the same features from the model that predicts the
sRPE score prior to training combined with features describing the ETL of each
athlete individually, such as the times in di erent heart rate zones. Again, we
used a GAM as the predictive model and 5-fold cross-validation was the chosen
validation strategy. The achieved RMSE and MAD are respectively 0:8548 0:09
and 0:6086 0:06, which is better than the currently best reported results in
literature from a di erent sport domain [
8
]. While still requiring more thorough
investigation, we believe that the alerts generated by our system could sometimes
indicate injury. For three of the eight samples were the MAD was higher than 3,
an injury occurred on the day itself or the next day. As an example, the sample
with the highest MAD, equal to 7.58, corresponded to a player that fell sick the
next day. Another example, with a MAD of 4.20 corresponded to a player who
had an injury to the knee on a training later that day. These samples should be
removed from the dataset, since they confuse the model, which would result in
an even lower error value. The results of the three di erent predictive models
are summarized in Table 2.
4