Content uploaded by Youri Geurkink
Author content
All content in this area was uploaded by Youri Geurkink on Mar 11, 2021
Content may be subject to copyright.
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Note. This article will be published in a forthcoming issue of the
International Journal of Sports Physiology and Performance. The
article appears here in its accepted, peer-reviewed form, as it was
provided by the submitting author. It has not been copyedited,
proofread, or formatted by the publisher.
Section: Original Investigation
Article Title: Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer:
Unravelling the Puzzle of Predictive Indicators
Authors: Youri Geurkink1, Gilles Vandewiele2, Maarten Lievens1, Filip de Turck2, Femke
Ongenae2, Stijn P.J. Matthys3, Jan Boone1, and Jan G. Bourgois1
Affiliations: 1Department of Movement and Sports Sciences, Ghent University, Ghent,
Belgium. 2Department of Information Technology, Ghent University, Ghent, Belgium.
3Performance Department, KAA Gent (Football Club), Ghent, Belgium.
Journal: International Journal of Sports Physiology and Performance
Acceptance Date: November 26, 2018
©2018 Human Kinetics, Inc.
DOI: https://doi.org/10.1123/ijspp.2018-0698
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Full title:
Modelling the Prediction of the session Rate of Perceived Exertion in Soccer: Unravelling the
puzzle of predictive indicators.
Submission type:
Original investigation.
Author details:
1. Youri Geurkink
Affiliations: Department of Movement and Sports Sciences, Ghent University, Ghent, Belgium
2. Gilles Vandewiele
Affiliations: Department of Information Technology, Ghent University, Ghent, Belgium
3. Maarten Lievens
Affiliations: Department of Movement and Sports Sciences, Ghent University, Ghent, Belgium
4. Filip de Turck
Affiliations: Department of Information Technology, Ghent University, Ghent, Belgium
5. Femke Ongenae
Affiliations: Department of Information Technology, Ghent University, Ghent, Belgium
6. Stijn P.J. Matthys
Affiliations: Performance Department, KAA Gent (Football Club), Ghent, Belgium
7. Jan Boone*
Affiliations: Department of Movement and Sports Sciences, Ghent University, Ghent, Belgium
8. Jan G. Bourgois*
Affiliations: Department of Movement and Sports Sciences, Ghent University, Ghent, Belgium
* These authors share last authorship.
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Corresponding email address:
Name: Jan G. Bourgois
Affiliations: Department of Movement and Sports Sciences, Ghent University, Ghent,
Belgium
Postal Adress: Watersportlaan 2, 9000 Ghent
Telephone number: 003292646297
Email address: jan.bourgois@ugent.be
Running title: Modelling the prediction of the session Rate of Perceived Exertion in soccer.
Abstract word count: 235
Test-only word count: 3380
Number of figures: 2
Number of tables: 1
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Abstract
Purpose: This study aimed to predict the session Rate of Perceived Exertion (sRPE) in
soccer and determine the main predictive indicators of the sRPE. Methods: A total of 70
External Load Indicators (ELIs), Internal Load Indicators (ILIs), Individual
Characteristics (ICs) and Supplementary Variables (SVs) were used to build a predictive
model. Results: The analysis using Gradient Boosting Machines showed a mean absolute
error (MAE) of 0.67 ± 0.09 AU and a Root Mean Squared Error (RMSE) of 0.93 ± 0.16
AU. ELIs were found to be the strongest predictors of the sRPE, accounting for 61.5%
of the total normalized importance (NI), with total distance as the strongest predictor.
The included ILIs and ICs accounted only for respectively 1.0% and 4.5% of the total
NI. Predictive accuracy improved when including SVs such as group-based sRPE-
predictions (10.5% of NI), individual deviation variables (5.8% of NI) and individual
player markers (17.0% of NI). Conclusions: The results showed that the sRPE can be
predicted quite accurately, using only a relatively limited number of training
observations. ELIs are the strongest predictors of the sRPE. It is however useful to
include a broad range of variables, other than ELIs, because the accumulated importance
of these variables account for a reasonable component of the total normalized
importance. Applications resulting from predictive modelling of the sRPE can help the
coaching staff to plan, monitor and evaluate both the external and internal training load.
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Introduction
Soccer training consists of structurally and systematically performed general and
specific exercises in order to improve physical abilities and acquire skills. The adaptations
imposed by a training stimulus take place on an anatomical, physiological, biochemical,
molecular and functional level 1. The content of a training session is usually prescribed by the
coach and can be defined as the external training load. The total external training load
comprises all of the players’ actions during a training session 2 and is generally quantified using
tracking technology 3. The external training load elicit an internal physiological stress, or the
internal training load. The internal training load is, however, not only dependent on the imposed
external training load, but also on the players’ individual characteristics (ICs) 2.
It is possible to assess the internal training load through quantification of a training
session’s duration and intensity 2. Duration is quantifiable in time and relatively easy to
measure. Intensity, on the other hand, can be quantified using different methods, such as heart
rate (HR) monitoring, blood lactate concentrations and the (session) Rate of Perceived Exertion
(sRPE) 2. HR monitoring is widely used in soccer, but it has been suggested that HR monitoring
underestimates or overestimates the intensity during intermittent activities 4,5. Furthermore, HR
monitoring requires both technical and physiological expertise to make an appropriate analysis
6. The blood lactate concentration is not often monitored during soccer training for practical
reasons 2. The sRPE, on the other hand, is a simple and practical tool that represents the players’
own perception of training stress, which includes both physiological and psychological stress
7. The sRPE has shown to be a valid indicator of intensity in soccer 7–10 and a more valid marker
of exercise intensity in soccer over a broad range of activities than HR monitoring 5.
The external training load is the first factor influencing the internal training load 2 and
has previously been investigated by Gaudino et al. 9. They confirmed that the sRPE is related
to external load indicators (ELIs) such as high-speed distance, impacts and accelerations.
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Secondly, the players’ ICs influence the internal training load. Because of the variability in
adaptations to a training stimulus 11, coaches should take the players’ ICs into account when
prescribing the external training load. It has been suggested that fitter athletes within a team
may not receive the optimal stimulus for physiological adaptations through extensive use of
group training exercises 12. On the other hand, players with inferior fitness may be overstressed
through group training sessions 2. To provide an appropriate training stimulus, it is important
to establish a level of agreement between the athlete’s sRPE and the coach’s sRPE. However,
previous research has shown that, in several sports, coaches tend to underestimate or
overestimate the athlete’s sRPE 13–15.
Based on the previous statement, it seems that coaches can experience difficulties
controlling the external training load and may put athletes at risk of maladaptive responses to
training 16, such as fatigue, injury and a reduction in performance 2. To further understand the
underlying indicators of external training load, internal training load and ICs contributing to
the sRPE, this study aims to predict the sRPE and identify the main predictors of the sRPE.
Methods
Subjects
Forty-six players (age: 25.6 ± 4.2 year) from an elite Belgian soccer team participated
in this study. Prior to the start of the study, players were informed about the study protocol and
the criteria set to assess the sRPE. Data were obtained between June 2015 and March 2017. To
ensure a higher level of homogeneity concerning the experimental data, goalkeepers were
excluded from this study, given the differences in external training load of goalkeepers
compared to the other playing positions. This study was approved by the ethical committee of
the Ghent University Hospital. Before the start of the study, all participants signed an informed
consent.
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Design
During 61 training sessions, a total of 913 individual training observations were
obtained. The analysed training sessions were prescribed and guided by the coaching staff,
without interference of the research staff. Strength and recuperation sessions were excluded to
ensure a greater similarity between training sessions. During the training sessions, players’ HR
was assessed through 20 Hz portable HR monitors and GPS (Polar Team Pro, Kempele,
Finland). GPS-units with a frequency of 10 Hz were used, the accelerometers inside the GPS-
units had a frequency of 100 Hz. No control subject participated in this observational study.
Methodology
session Rate of Perceived Exertion
The sRPE was assessed 15 minutes after each training session. Players were asked to
rate the training session between 1 and 10 using the session-RPE scale 17. All players were
familiarised with the protocol.
External Load Indicators
Using the GPS-units, the following variables were measured during the training
sessions: total distance (m), training duration (s), distance (m) in 5 speed zones (3.00 – 6.99
km/h; 7.00 – 10.99 km/h; 11.00 – 14.99 km/h; 15.00 – 18.99 km/h; > 19.00 km/h), the number
of accelerations (m/s2) (0.50 – 0.99; 1.00 – 1.99; 2.00 – 2.99; 3.00 – 50.00), the number of
decelerations (m/s2) (0.50 – 0.99; 1.00 – 1.99; 2.00 – 2.99; 3.00 – 50.00) and the number of
sprints (> 25 km/h). Average speed (m/s) was derived using distance and time. The proportion
of workload in every acceleration-, deceleration- and speed zone relative to the total workload
was also included into the predictive model. Lastly, individual deviations of several ELIs (total
distance, total time and number of sprints) compared to the group mean, based on historical
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
training data, were derived. This way, the model could take differences in external workload
for players compared to the group mean into account.
Internal Load Indicators
Data concerning HR was subdivided into five different HR-zones relative to their
maximum HR (50% - 60%, 60% - 70%, 70% - 80%, 80% - 90%, 90% - 100%) (s). The
maximum HR was determined during an incremental test protocol on a treadmill. The
proportion of the time spent in each HR-zone relative to the total time spent in the 5 HR-zones
and the Edwards’ TRIMP were also included into the model. Edwards’ TRIMP is expressed as
the product of the accumulated training duration in each HR-zone with a coefficient relative to
each HR-zone (50% - 60% = 1, 60% - 70% = 2, 70% - 80% = 3, 80% - 90% = 4, 90% - 100%
= 5) 18.
Individual Characteristics
The individual characteristics can be subdivided into physiological and personal
characteristics. The physiological characteristics consisted of assessment of conditional
parameters, sprinting speed, acceleration, anaerobic power and muscle fiber composition.
Assessment of conditional parameters was conducted on a treadmill using an incremental test
protocol, starting at 8 km/h with an increase in speed of 2 km/h every 3 minutes. The sprinting
tests consisted of two different exercises, namely a 10-meter sprint and a 5 times 10-meter
shuttle run. The 10-meter sprint was used as a measure of starting speed and acceleration,
whereas the shuttle run provided information about the players’ agility in combination with
speed. The jumping test was used as a measure of anaerobic power and consisted of a counter
movement jump with arms. The complete test battery was performed on the same day. The
participants did not perform strenuous exercise 24 hours before testing. The tests were
conducted every 6 months and changes in test performances were registered. Because of
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
injuries or decreased fitness status, 5 out of the 46 players were unable to perform the sprinting
tests. For the complete description of the tests, the experimental protocol from Boone et al. 19
can be consulted.
Muscle fibre composition was estimated in a non-invasive way, through measurements
of the carnosine content in the gastrocnemius and soleus by proton magnetic resonance
spectroscopy. For a complete description of the protocol, Baguet et al. 20 can be consulted. The
muscle fibre distribution was only measured once, since muscle fibre is largely dependent on
genetics 21 and substantial changes in the distribution of muscle fibres because of training are,
in our case, unlikely. Because of unavailability of the MRI scanner, muscle fibre type was not
determined in 21 out of the 46 players.
The personal characteristics consisted of age, playing position and nationality. Players’
age was determined during each training observation. Playing position was subdivided into 5
categories (central defender, full back or wide midfielder, central midfielder, winger and
central attacker). Nationality was defined using the latitude and longitude of the players’
country of birth.
Supplementary variables
Several variables and techniques were used in an attempt to improve predictive
accuracy and were categorized as ‘supplementary variables’ (SVs). First, group-based sRPE-
prediction variables were included. Based on the average external and internal workload,
several predictive models were used to predict a group-based prediction of the sRPE. The
following models were used to predict the group-based sRPE: Generalized Additive Models,
Multivariate Adaptive Regression Splines, Decision Tree, Random Forest, Linear Regression
and Support Vector Regression. The predicted group-based sRPE-values were then included
into the model. Players’ individual deviance from the group mean (mean deviation, standard
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
deviation, maximum deviation, minimum deviation) was also included, since individual
differences in the interpretation of the sRPE-scale can influence the sRPE-value provided by
the players. To further account for differences between players, the effect of every player on
the predictive accuracy was quantified using one-hot encoding. One-hot encoding is defined as
mapping of a variable to a binary vector of length equal to the number of categories, in our
case the number of players (n = 46). All elements in the vector are converted to zero except at
the index corresponding to the category of that sample. This technique created 46 features for
the model but is in this study regarded as one variable. Lastly, 4 variables regarding weather
(temperature, humidity, visibility and windspeed) were included to account for the effect of
weather on the perceived training load.
Statistical analyses
Gradient Boosting Machines were used to identify the main predictive indicators of the
sRPE and to predict the sRPE. This machine learning technique creates a large number of
different decision trees, resulting in an integrated model predicting the outcome. Missing
values for muscle fibre type, sprinting times and jumping performance were replaced with the
group mean. Data was analysed using Python 3.5.
To evaluate predictive accuracy, two different standard statistical metrics were used,
namely the mean absolute error (MAE) and the Root Mean Squared Error (RMSE). It was
opted to express model accuracy also in the number of correct classifications. The predicted
value was rounded number, from which could be evaluated if the model predicted correctly.
To provide more information about the deviation from the mean, a ‘loose accuracy’ approach
was used. This approach classified a difference within a range of 1 as correct, when the
predicted value was compared with the observed sRPE. For example, if the observed value was
4, all predicted values between 3 and 5 were classified as correctly.
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Measurements were taken using 5-fold cross-validation, in which the data was
partitioned according to training identifiers. In other words, on average 50 and 13 training
sessions were used in respectively the training and test set in each fold. This was done to avoid
contamination between the training and test set. The influence of each included indicator was
expressed as a normalized importance value. The importance of each variable is a measure of
the magnitude by which the model-predicted sRPE-value alters for various values of a variable.
In our case, the normalized importance was determined through the number of expressions of
a variable in the created decisions trees.
Results
The average training sRPE was 4.34 ± 1.06 AU, the distribution of sRPE-values can be
found in figure 1. The median number of records per player was 18 (ranging from 2 to 46).
Other descriptive values of regularly reported ELIs were: duration 90 ± 20 minutes and total
distance covered 5977 ± 1893 m.
Predicting the sRPE
The predictive model showed a MAE of 0.67 ± 0.09 AU and a RMSE of 0.93 ± 0.16
AU. In total, 47.6 ± 7.17% of the cases were correctly classified. The ‘loose accuracy’
approach, classifying all predictions within a range of plus or minus 1 relative to the observed
sRPE-score as correct, resulted in 91.7 ± 3.45% correctly classified cases. The model did not
predict sRPE-values of 8, 9 or 10, because of the limited number of observed high sRPE-values.
In figure 2, a confusion matrix is depicted showing the observed versus the predicted sRPE-
values.
Predictive indicators of the sRPE
Most of the strongest individual predictors of the sRPE can be regarded as ELIs, since
total distance, total time and number sprints are among the strongest predictors. The normalized
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
importance of all ELIs together accounted for 61.5%. The sRPE showed a positive relationship
with most measures, presenting an increased sRPE when external workload increased. Players
ICs accounted only for 4.5% of the total normalized importance. HR showed to be a poor
predictor of the sRPE since Internal Load Indicators (ILIs) only accounted for 1.0% of the total
normalized importance.
The SVs were subdivided in machine learning variables (10.5% of NI), individual
deviation variables (5.8% of NI), individual player markers (17.0% of NI) and weather
variables (0.2% of NI), which accounted for 33% of the total normalized importance. An
overview of the NI of all indicators can be found in table 1.
Discussion
The aim of this study was to predict the sRPE and determine the main predictive
indicators. The sRPE was predicted quite accurately, using a broad range of variables. Our
study in soccer is, to our knowledge, the first to incorporate a large set of predictive indicators
other than ELIs, namely ICs, ILIs and SVs. The findings demonstrate that ELIs are the
strongest predictors of the sRPE. It is also useful to improve predictive accuracy by including
machine learning variables, individual deviation variables and individual players markers. The
included ILIs and ICs showed poor predictive value.
The findings from this study shows higher predictive accuracy than previous studies
conducted in soccer 22,23. Despite a smaller dataset, our results showed a lower MAE and
RMSE. This may be explained using different machine learning techniques and/or other
predictive indicators. Our findings showed that even on a relatively small dataset, the sRPE
can be predicted quite accurately. This may offer solutions to (sub-)elite teams, especially when
teams consist of large numbers of players. The sRPE is regularly monitored in team sports,
however, it can take considerable time and effort 24. The model used in our research could help
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
to predict players’ sRPE, which might make daily assessment of the sRPE unnecessary 25. The
coaching staff may choose to closely monitor the external and internal training load for a
limited period, to build a model that is appropriately predictive. After this period, the coaching
staff can choose to monitor the predicted sRPE, instead of daily collection of the sRPE. The
constructed model may also form a basis for live monitoring of the sRPE during a training
session 25,26. This would allow coaches to adapt players’ training content, possibly preventing
negative training effects. Live monitoring of the sRPE may however not only provide live
information but could also provide information regarding the process of accumulating external
training load on the sRPE. Currently, the sRPE only provides a global indication about the
‘product’ of the internal training load, rather than provide information about the dynamical
‘process’ in which a certain internal training load is achieved. Since different combinations of
external training load can elicit a similar sRPE, more insight into temporal changes of the sRPE
during training sessions could help the coaching staff to prescribe and adapt the external
training load.
Previous research already showed that ELIs can be used to predict the sRPE in soccer
22,23. Our results showed that total distance was the strongest predictor of the sRPE, while total
duration and the number of sprints were also strong predictors. The importance of these factors
in relation to perceived training load can be explained by the fact that these measures capture
information about both session volume and intensity 25. It should however be noted that
findings from our research show some differences in regard to previous research 22,23, since the
importance assigned to each ELI varies between studies. Differences between studies may be
explained by differences in training content, the use of different devices to register players’
activity or different modelling approaches. Although several findings with regard to variable
importance can be generalized, the differences in variable importance between studies may
imply limited transferability of predictive models. Previous studies did however not include
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
variables other than ELIs. Since these variables account for 38.5% of the NI, it is useful to
include a broad range of variables with a possible impact on the perceived exertion. The ICs,
ILIs and weather variables showed however only small predictive value. A set of variables
which showed promise are the individual deviation variables and the individual players
markers (respectively 5.8% and 17.0% of the total NI). The normalized importance values
assigned to these variables shows that there are differences in interpretation of the sRPE-scale,
since some players consistently rate a training session higher or lower than the group mean.
This corresponds with a previous statement by Bartlett et al. 27, endorsing the importance of
personal interpretation of the sRPE-scale and an individualized interpretation of the perceived
exertion.
There are some limitations and opportunities to be stated. A limitation of our study is
limited number of extreme sRPE-values, especially on the higher end of the continuüm. For
that reason, the model was adapted to predict sRPE-scores between 1 and 7, which explains
the error for sRPE-values of 8, 9 or 10. Extreme sRPE-values could however provide very
useful information to both the model as well as the coaching staff. Furthermore, the error was
larger when predicting lower sRPE-values. Future studies should include more extreme sRPE-
values. The results showed that is useful to include a broad range of variables into a predictive
model. Including more variables may improve model accuracy, offering opportunities for
future research. In our study, the external training load was not registered over a prolonged
series of training sessions, so changes in external training load over time, which can be
quantified using measures such as the acute/chronic workload ratio 29 and exponentially
moving weighted averages 30 are not taken into account. Lastly, information about the training
modalities could be quantified to provide useful information to the predictive model. When
including more factors into a predictive model, researcher should consider methodological
quality, but also practicality of the predictive factor. Baseline feeling or lifestyle involving for
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
example diet or rest probably may have an influence on the sRPE. However, these factors are
practically more difficult to quantify. Subjective ratings may at the moment be a suitable
alternative when attempting to get quantitative insight into lifestyle, since previous research
showed that morning measured ratings of fatigue, sleep quality and delayed onset of muscle
soreness (DOMS) are sensitive markers to daily fluctuations in the sRPE in elite soccer players
28. Although our research may already be regarded as fairly comprehensive given the number
of individual characteristics, variables such as the aforementioned can also be incorporated to
further improve model accuracy and explore the role of the different predictive indicators.
Practical Applications
Predictive modelling of the sRPE may aid the process of planning and adjusting training
load. In our model, regularly monitored ELIs such as distance, training duration, the number
of sprints and deratives of these ELIs were found to be strong predictors, showing that it is
important to carefully consider these ELIs when planning and monitoring training load.
Furthermore, differences between individuals on a physiological and personal basis, but also
differences when interpreting the sRPE-scale, can influence the sRPE-values. Our study
provides further confirmation for an individualised approach when planning and monitoring
training load.
Conclusions
Our predictive model showed that the sRPE can be predicted quite accurately using
only a relatively small number of training observations. ELIs are the strongest predictors of the
sRPE, with total distance as the strongest predictor. Including a broad range of variables, other
than ELIs, is useful since the accumulated importance of these variables account for a
reasonable component of the total normalized importance. Applications resulting from
predictive models may help a coaching staff to plan, monitor and evaluate training sessions.
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Future research could focus on including more variables into the model, such as factors of daily
status and changes in external training load over time, to create an even more comprehensive
model predicting the sRPE.
Acknowledgements
The authors would like to thank the coaches and players for their cooperation.
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
References
1. Viru A, Viru M. Nature of training effects. Exerc Sport Sci. 2000;6(7):67-95.
2. Impellizzeri FM, Rampinini E, Marcora SM. Physiological assessment of aerobic
training in soccer. J Sports Sci. 2005. doi:10.1080/02640410400021278
3. Cummins C, Orr R, O’Connor H, West C. Global positioning systems (GPS) and
microtechnology sensors in team sports: A systematic review. Sport Med. 2013.
doi:10.1007/s40279-013-0069-2
4. Algrøy EA, Hetlelid KJ, Seiler S, Pedersen JIS. Quantifying training intensity
distribution in a group of norwegian professional soccer players. Int J Sports Physiol
Perform. 2011. doi:10.1123/ijspp.6.1.70
5. Little T, Williams AG. Measures of Exercise Intensity during Soccer Training Drills
with Professional Soccer Players. J Strength Cond Res. 2007;21(2):367-371.
doi:https://doi.org/10.1519/00124278-200705000-00013
6. Alexiou H, Coutts AJ. A comparison of methods used for quantifying internal training
load in women soccer players. Int J Sports Physiol Perform. 2008.
doi:10.1080/02640424.2015.1088166
7. Impellizzeri FM, Rampinini E, Coutts AJ, Sassi A, Marcora SM. Use of RPE-based
training load in soccer. Med Sci Sports Exerc. 2004.
doi:10.1249/01.MSS.0000128199.23901.2F
8. Borresen J, Lambert MI. Quantifying Training Load: A Comparison of Subjective and
Objective Methods. Int J Sports Physiol Perform. 2008;3:16-30.
doi:10.1123/ijspp.3.1.16
9. Gaudino P, Iaia FM, Strudwick AJ, et al. Factors influencing perception of effort
(session rating of perceived exertion) during elite soccer training. Int J Sports Physiol
Perform. 2015. doi:10.1123/ijspp.2014-0518
10. Haddad M, Chaouachi A, Wong DP, et al. Influence of fatigue, stress, muscle soreness
and sleep on perceived exertion during submaximal effort. Physiol Behav. 2013.
doi:10.1016/j.physbeh.2013.06.016
11. Timmons JA. Variability in training-induced skeletal muscle adaptation. J Appl
Physiol. 2011. doi:10.1152/japplphysiol.00934.2010
12. Hoff J. Soccer specific aerobic endurance training. Br J Sports Med. 2002.
doi:10.1136/bjsm.36.3.218
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
13. Brink MS, Frencken WGP, Jordet G, Lemmink KAPM. Coaches’ and players’
perceptions of training dose: Not a perfect match. Int J Sports Physiol Perform. 2014.
doi:10.1123/IJSPP.2013-0009
14. Doeven SH, Brink MS, Frencken WGP, Lemmink KAPM. Impaired Player-Coach
Perceptions of Exertion and Recovery During Match Congestion. Int J Sports Physiol
Perform. 2017. doi:10.1123/ijspp.2016-0363
15. Murphy AP, Duffield R, Kellett A, Reid M. Comparison of athlete-coach perceptions
of internal and external load markers for elite junior tennis training. Int J Sports
Physiol Perform. 2014. doi:10.1123/IJSPP.2013-0364
16. Wallace L, Coutts A, Bell J, Simpson N, Slattery K. Using Session-RPE to Monitor
Training Load in Swimmers. doi:10.1519/ssc.0b013e31818eed5f
17. Foster C, Florhaug JA, Franklin J, et al. A New Approach to Monitoring Exercise
Training. J Strength Cond Res Natl Strength Cond Assoc J Strength Cond Res.
2001;15(151). doi:10.1519/00124278-200102000-00019
18. Edwards S. The heart rate monitor book. Med Sci Sport Exerc. 1994;26(5):647.
doi:10.1249/00005768-199405000-00020
19. Boone J, Vaeyens R, Steyaert A, Bossche L Vanden, Bourgois J. Physical Fitness of
Elite Belgian Soccer Players by Player Position. J Strength Cond Res. 2012.
doi:10.1519/JSC.0b013e318239f84f
20. Baguet A, Everaert I, Hespel P, Petrovic M, Achten E, Derave W. A New Method for
Non-Invasive Estimation of Human Muscle Fiber Type Composition. PLoS One.
2011;6(7):6. doi:10.1371/journal.pone.0021956
21. Simoneau J, Bouchard C. Genetic determinism of fiber human skeletal muscle type
proportion in. J Fed Am Soc Exp Biol. 1995;9(11):1091-1095.
doi:10.1096/fasebj.9.11.7649409
22. Jaspers A, Brink MS, Probst SGM, Frencken WGP, Helsen WF. Relationships
Between Training Load Indicators and Training Outcomes in Professional Soccer.
Sport Med. 2017. doi:10.1007/s40279-016-0591-0
23. Rossi A, Perri E, Trecroci A, Savino M, Alberti G, Iaia FM. GPS data reflect players’
internal load in soccer. In: IEEE International Conference on Data Mining Workshops,
ICDMW. New Orleans, LA, USA: IEEE; 2017:890-893.
doi:10.1109/ICDMW.2017.122
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
24. Burgess DJ. The Research Doesn’t Always Apply: Practical Solutions to Evidence-
Based Training Load Monitoring in Elite Team Sports. Int J Sports Physiol Perform.
2016:1-19. doi:10.1123/ijspp.2016-0608
25. Carey DL, Ong K, Morris ME, Crow J, Crossley KM. Predicting ratings of perceived
exertion in Australian football players: Methods for live estimation. Int J Comput Sci
Sport. 2016. doi:10.1515/ijcss-2016-0005
26. Vandewiele G, Geurkink Y, Lievens M, Ongenae F, De Turck F, Boone J. Enabling
Training Personalization by Predicting the Session Rate of Perceived Exertion (sRPE).
In: Machine Learning and Data Mining for Sports Analytics ECML/PKDD. ; 2017.
27. Bartlett JD, O’Connor F, Pitchford N, Torres-Ronda L, Robertson SJ. Relationships
between internal and external training load in team-sport athletes: Evidence for an
individualized approach. Int J Sports Physiol Perform. 2017. doi:10.1123/ijspp.2015-
0791
28. Thorpe RT, Strudwick AJ, Buchheit M, Atkinson G, Drust B, Gregson W. Tracking
morning fatigue status across in-season training weeks in elite soccer players. Int J
Sports Physiol Perform. 2016. doi:10.1123/ijspp.2015-0490
29. Hulin BT, Gabbett TJ, Lawson DW, Caputi P, Sampson JA. The acute:chronic
workload ratio predicts injury: high chronic workload may decrease injury risk in elite
rugby league players. Br J Sports Med. 2016. doi:10.1136/bjsports-2015-094817
30. Williams S, West S, Cross MJ, Stokes KA. Better way to determine the acute:chronic
workload ratio? Br J Sports Med. 2016. doi:10.1136/bjsports-2016-096589
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Figure 1: distribution of observed sRPE-values from 46 players, resulting in 913 training
observations.
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive
Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Figure 2: confusion matrix showing the observed versus the predicted sRPE-values. The
matrix provides insight into the predictive model by showing the number of correct and
incorrect prediction for each sRPE-value.
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18
“Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive Indicators” by Geurkink Y et al.
International Journal of Sports Physiology and Performance
© 2018 Human Kinetics, Inc.
Table 1: Overview of the normalized importance for all indicators.
Cat.
Indicator
NI
Description
ELIs
Total distance
0.250
Total distance (m) covered during a training session.
Total duration
0.067
Total time (s) during a training session.
Number of sprints
0.072
Total number of sprints (n) during a training session.
Deviation of total distance
0.061
Differences in total distance between players during a training session.
Deviation of total sprints
0.109
Differences in the total number of sprints between players during a training session.
Other ELIs (# 28)
0.058
Average speed, distance covered in 5 speed zones (km/h) (3.00 - 6.99, 7.00 - 10.99, 11.00 - 14.99, 15.00 - 18.99,
> 19.00), normalized proportion of distance covered in each speed zone, number of accelerations in each zone
(m/s2) (0.50 - 0.99, 1.00 - 1.99, 2.00 - 2.99, 3.00 - 50.00), number of decelerations (m/s2) (0.50 - 0.99, 1.00 -
1.99, 2.00 - 2.99, 3.00 - 50.00), normalized proportion of number of accelerations and decelerations in each
zone, differences in training duration between players during a training session.
All variables related to ELIs (# 33)
0.615
ICs
Physiological characteristics (# 8)
0.013
VO2max, speed at aerobic and anaerobic threshold, muscle fibre composition, sprint time (time at 5 meter and
10 meter), 5 times 10-meter shuttle run, counter movement jump with arms.
Personal characteristics (# 3)
0.032
Age, nationality, playing position (central defender, full back, central midfielder, winger, striker).
All variables related to ICs (# 11)
0.045
ILIs
Heart rate variables (# 11)
0.010
Time spent in 5 HR-zones relative to the maximum HR (50% - 60%, 60% - 70%, 70% - 80%, 80% - 90%, 90%
- 100%), proportion of time spent in each HR-zone, training load based on Edwards’ TRIMP.
All variables related to ILIs (# 11)
0.010
SVs
Group-based RPE-predictions (# 6)
0.105
The sRPE of every training session was predicted by the following models: GAM, MARS, DT, RF, LR, SVR.
Individual deviation variables (# 4)
0.058
Individual sRPE deviances from group mean (mean, maximum, minimum and standard deviation).
Individual player marker (# 1)
0.170
The players’ individual influence on predictive accuracy was quantified by through one-hot encoding.
Weather variables (# 4)
0.002
Weather circumstances (temperature, humidity, visibility, wind speed).
All other variables (# 15)
0.330
Abbreviations: Cat, Category; NI, Normalized Importance; #, number of; ELIs, External Load Indicators; ICs, Individual Characteristics; ILIs, Internal Load Indicators; SVs,
Supplementary Variables; HR, Heart Rate; sRPE, session Rate of Perceived exertion; GAM, Generalized Addictive Models; MARS, Multivariate Adaptive Regression Splines;
DT, Decision Tree; RF, Random Forest; LR, Linear Regression; SVR, Support Vector Regression.
Downloaded by UNIVERSITEITSBIBLIOTHEEK GENT on 12/23/18