ArticlePDF Available

Modelling the Prediction of the Session Rate of Perceived Exertion in Soccer: Unravelling the Puzzle of Predictive Indicators

Authors:

Abstract and Figures

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.
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 710 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 1315.
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
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
... Each item was rated on a 10-point scale. The individual items were subsequently categorised as muscle soreness, sleep quality, or energy levels using the average score of each item in the pertaining to the athletes individual characteristics improved the predictive accuracy of a gradient boosting machine in soccer with a RMSE of 0.93 although only 47.6% of predictions were correct [17]. ...
... The boundaries of these categories are reflective of the first and second ventilatory thresholds [24], and may be used to demarcate entry into the three distinct physiological exercise intensity domains [29]. This approach may be more practically rele- [11,17,18]. However, one study reported an accuracy of 47.6% for single RPE values which was increased to 91.7% when using a 'loose accuracy' approach, where predictions within a range of ± 1 unit of the actual value were marked correct [17] wellness scores, and training workloads alongside the absolute and relative ELI at 67.0% and 65.2% accuracy, respectively. ...
... This approach may be more practically rele- [11,17,18]. However, one study reported an accuracy of 47.6% for single RPE values which was increased to 91.7% when using a 'loose accuracy' approach, where predictions within a range of ± 1 unit of the actual value were marked correct [17] wellness scores, and training workloads alongside the absolute and relative ELI at 67.0% and 65.2% accuracy, respectively. The accuracy of the RF and BS models for the low, moderate and high RPE categories are presented in a confusion matrix in Figure 1. ...
... Finally, well-performing models provide insight into the most important factors, by looking at which of them have the greatest influence in the decision-making process of ML models. In this context, a growing number of studies have focused on the prediction of both injuries and performance through ML approaches (Geurkink et al., 2019;Alessio;Rossi et al., 2019). Although a previous review reported that there is no consensus regarding the best way to predict sports injuries and performance using ML, each method is able to predict a specific variable in different contexts (Kotsiantis et al., 2007) such as the Decision Tree (Rossi et al., 2018) and the Artificial Neural Network (Claudino et al., 2019). ...
... IL was assessed with RPE proposed by Foster (Foster, 1998;Foster et al., 2021), S-RPE (derived multiplying RPE x duration) (Foster et al., 2017(Foster et al., , 2021, and/or HR, while EL using GPS (Rossi et al., 2018) and/or accelerometer indicators. Twenty-one studies (84%) used standard statistical analysis (Alemdaroğlu, 2020;Derbidge et al., 2020;Enes et al., 2021;Gaudino et al., 2015;Gomez-Piriz et al., 2011;Grünbichler et al., 2020;Lacome et al., 2018;Lu et al., 2017;Marynowicz et al., 2020;Maughan et al., 2020Maughan et al., , 2021Montini & Rocchi, 2020;Rago et al., 2019;Silva et al., 2018;Suarez-Arrones et al., 2015;Torreño et al., 2016;Wiig et al., 2020) and only four (16%) (Geurkink et al., 2019;Alessio;Rossi et al., 2019;Vallance et al., 2020) used ML approaches to analyse data. Only four studies regarding injuries were included: three used IL and EL to understand their association with injuries (Jaspers et al., in these workloads increased IR. ...
... Five studies aimed to predict IL using EL indicators (Geurkink et al., 2019;Lacome et al., 2018;Alessio;Rossi et al., 2019;Silva et al., 2018). Two studies predicted HR with standard statistical analysis (Lacome et al., 2018;Silva et al., 2018) and three studies predicted RPE from EL variables using ML (Geurkink et al., 2019;Alessio;Rossi et al., 2019). ...
... A total of 29 studies (48.3%) included internal workload monitoring. Heart rate monitoring was utilized by five articles (Bunn et al. 2021;Carey et al. 2016;Crouch et al. 2021;Di Credico et al. 2021;Geurkink et al. 2019;Jelinek et al. 2014), 14 studies included Rated Perceived exertion (RPE) before/after training sessions to measure internal load of players and nine articles included RPE is a subjective form of monitoring an athlete's physical activity (Foster et al. 2001). Reliability of the RPE tool has shown moderate positive correlations and has been identified in previous studies investigating wellness and RPE in male Australian footballers (r=0.59) ...
... This may be why an emphasis is placed on the evaluation of physical demands during training sessions in this sport using both internal and external assessments of load (Gaudino et al. 2015). GPS tracking, although still in its infancy, was the most common external workload monitoring tool used when investigating soccer ( (Klemp et al. 202;Philp et al. 2020;Zumeta-Olaskoaga et al. 2021), three studies combined two methods (Goes et al. 2021;Jaspers et al. 2018;Oliver et al. 2020;Perri et al. 2021) and 11 studies utilized more than three tools (Bruce & Wilkerson 2021;Di Cesare et al. 2020;Dijkhuis et al. 2021;Fanchini et al. 2018;Gasparini & Alvaro 2020;Geurkink et al. 2019;Jaspers et al. 2018;Jauhiainen et al. 2019;Mandorino et al. 2021;Mandorino et al. 2021;Vallance et al. 2020) in which no similar predictive tools were included in each methodology. Therefore, although several prediction and modelling strategies have been proposed for injury ...
... ;Colby et al. 2017; Colby et al. 2017; Colby et al. 2018; Dijkhuis et al. 2021; Gasparini & Alvaro 2020;Geurkink et al. 2021; Guerrero- Calderon et al. 2021;Klemp et al. 2021;Thorton et al. 2017;Windt et al. 2017), 13 articles utilized GPS and accelerometers(Bartlett et al. 2016; Bunn et al. 2021; Carey et al. 2016; Carey et al. 2018; Chambers et al. 2018; Crouch et al. 2021;Gastin et al. 2019;Gaudino et al. 2015;Geurkink et al. 2019; Gimenez et al. 2020;Rossi et al. 2018;Rossi et al. 2019;Vallance et al. 2020) and five included accelerometers(Di Credico et al. 2021;Gabbett et al. 2011;Jaspers et al. 2018;Peek et al. 2021;Wilkerson et al. 2018). Three studies utilized a combination GPS, accelerometers, magnetometers and gyroscopes(Bunn et al. 2021; Chambers et al. 2018; Crouch et al. 2021).Gaudino ...
Article
Full-text available
Training load (TL) is frequently documented among team sports and the development of emerging technology (ET) is displaying promising results towards player performance and injury risk identification. The aim of this systematic review was to identify ETs used in field-based sport to monitor TL for injury/performance prediction and provide sport specific recommendations by identifying new data generation in which coaches may consider when tracking players for an increased accuracy in training prescription and evaluation among field-based sports. Data was extracted from 60 articles following a systematic search of CINAHL, SPORTDiscus, Web of Science and IEEE XPLORE databases. Global positioning system (GPS) and accelerometers were common external TL tools and Rated Perceived Exertion (RPE) for internal TL. A collection of analytics tools were identified when investigating injury/performance prediction. Machine Learning showed promising results in many studies, identifying the strongest predictive variables and injury risk identification. Overall, a variety of TL monitoring tools and predictive analytics were utilized by researchers and were successful in predicting injury/performance, but no common method taken by researchers could be identified. This review highlights the positive effect of ETs, but further investigation is desired towards a ‘gold standard” predictive analytics tool for injury/performance prediction in field-based team sports.
... To date, the most common approaches using machine learning algorithms in soccer have focused on the relationship between the RPE and different internal and external loads during training sessions or matches for senior players (Geurkink et al., 2019;Jaspers et al., 2018;Rossi, Perri, Pappalardo, Cintia, & Iaia, 2019). Recently, machine learning algorithms have also been applied for predicting RPE based on GPS-derived external measures in youth soccer training (Marynowicz, Lango, Horna, Kikut, & Andrzejewski, 2022). ...
... The sample size contributed to the findings. The current analysis included a total of 787 individual observations, similar to previous studies (Geurkink et al., 2019;Marynowicz et al., 2022), but lower than other ones (Jaspers et al., 2018;Rossi et al., 2019). However, the model accuracy decreased when it was conducted with a reduced sample size (differentiating between player status) and spectrum of variables. ...
... However, the model accuracy decreased when it was conducted with a reduced sample size (differentiating between player status) and spectrum of variables. On the other hand, findings with regard to variable importance imply limited comparability with previous results due to the different machine learning algorithms and variables selected for the prediction of RPE (Geurkink et al., 2019;Jaspers et al., 2018;Marynowicz et al., 2022;Rossi et al., 2019). ...
Article
The aim of this study was to explore how situational variables affect youth soccer players' perceived exertion (RPE) after official matches. Thirty-five elite youth male players (14.33 ± 0.86 years; 173.49 ± 6.16 cm; 63.44 ± 5.98 kg) who belonged to two different teams of a professional club participated in this study. Data collection was conducted during two seasons (2016-2017, 2017-2018) and included 60 official matches (30 official matches per team). Ten minutes after each match players rated their RPE and using a modified Borg CR-10 scale. A Random Forest Regression was used to quantify the importance of match-related situational variables in RPE. Afterwards, a linear mixed model analysis was applied to identify the variability in RPE among the situational variables. The game-playing time, the player status (starter or substitute) and the player identity were the strongest predictors of RPE. Moreover, the match outcome and the final scoreline showed significant effects on both starter and substitute players but the main effect of the quality of the opponent was only identified in starter players (p < 0.05). These results allow practitioners to know how situational variables interact and modulate RPE after official matches and help them to prescribe and adapt the players' training content and load before and after matches.
... In terms of continuous on-field monitoring of athletes, the most commonly captured physiological measurements consist of heart rate (HR) and various HR-derived metrics, partly thanks to a high level of accessibility to commercially available validated sensor systems (5,6). HR-based metrics are often more difficult to interpret than RPE and require greater physiological expertise to analyse (8). In pursuit of a continuous monitoring metric which is easy to interpret, researchers have been attempting to estimate RPE using physiological measurements that can be continuously monitored via sensor systems (9)(10)(11)(12)(13)(14)(15). ...
... For the RFR model tuning, the number of estimators was tuned between 50 to 200 at an increment of 50, and 100 estimators were found to be the most optimal. For the MLP model, the number of dense layer units (32, 64, 128), the learning rate (1eÀ3, 1eÀ4, 1eÀ5), and the epoch number (5)(6)(7)(8)(9)(10)(11)(12)(13)(14)(15)(16)(17)(18)(19)(20) were tuned, and the final model was built with 128 units per dense layer, trained at 1eÀ3 learning rate for 15 epochs. Table 2 shows the detailed test results of the leave-one-subjectout cross-validation. ...
Article
Full-text available
Introduction The purpose of this study is to investigate the importance of respiratory features, relative to heart rate (HR), when estimating rating of perceived exertion (RPE) using machine learning models. Methods A total of 20 participants aged 18 to 43 were recruited to carry out Yo-Yo level-1 intermittent recovery tests, while wearing a COSMED K5 portable metabolic machine. RPE information was collected throughout the Yo-Yo test for each participant. Three regression models (linear, random forest, and a multi-layer perceptron) were tested with 8 training features (HR, minute ventilation (VE), respiratory frequency (Rf), volume of oxygen consumed (VO2), age, gender, weight, and height). Results Using a leave-one-subject-out cross validation, the random forest model was found to be the most accurate, with a root mean square error of 1.849, and a mean absolute error of 1.461 ± 1.133. Feature importance was estimated via permutation feature importance, and VE was found to be the most important for all three models followed by HR. Discussion Future works that aim to estimate RPE using wearable sensors should therefore consider using a combination of cardiovascular and respiratory data.
... Bongiovanni et al. [14] showed high accuracy of the extra tree regression model in predicting football skills (change of direction and vertical jump tests), as well as performance (YYIRT1) and endurance (10 and 20 m sprint) based on measurements of anthropometric parameters such as adjusted arm muscle area, arm muscle circumference, and suprapatellar girth) [14]. Geurkink et al. [46] used gradient-boosting machine learning models to predict the optimal workload level during a match based on heart rate [46]. In addition, player performance during a match was estimated by the time that athlete would have run at a constant pace on grass from video footage analysis of the match, according to a study by Manzi et al. [47], where they used K-Nearest Neighbors, Decision Tree, Random Forest, and Support-Vector Machine classifiers for the predictions based Sandamal et al. Results in Engineering xxx (xxxx) 102592 on body height and weight, percentage of body fat, VO2 max, and heart rate [47]. ...
... Bongiovanni et al. [14] showed high accuracy of the extra tree regression model in predicting football skills (change of direction and vertical jump tests), as well as performance (YYIRT1) and endurance (10 and 20 m sprint) based on measurements of anthropometric parameters such as adjusted arm muscle area, arm muscle circumference, and suprapatellar girth) [14]. Geurkink et al. [46] used gradient-boosting machine learning models to predict the optimal workload level during a match based on heart rate [46]. In addition, player performance during a match was estimated by the time that athlete would have run at a constant pace on grass from video footage analysis of the match, according to a study by Manzi et al. [47], where they used K-Nearest Neighbors, Decision Tree, Random Forest, and Support-Vector Machine classifiers for the predictions based Sandamal et al. Results in Engineering xxx (xxxx) 102592 on body height and weight, percentage of body fat, VO2 max, and heart rate [47]. ...
Article
Full-text available
This study is the first attempt to use explainable machine learning (ML) to investigate the factors influencing physical development and the athlete's performance in the Karakalpakstan and Khwarazm regions of the Republic of Uzbekistan. Three physical fitness tests (Dribbling Shuttle, Goal Accuracy, and Yo-Yo intermittent), anthro-pometrics, and hematology data (33 subject-specific factors) of 60 young soccer players were used in this study. Results reveal that the explainable ML algorithms were effective in prediction with high model accuracy measured from mean absolute error and mean squared error (0.004 and 0.052 for Dribbling Shuttle test with Random Forest model, 41.783 and 5 for Yo-Yo Intermittent Recovery Test-level with XGB model and 0.027 and 0.135 for Goal Accuracy test with XGB model). In contrast, model bias was estimated using mean absolute percentage error (3.6 % for the Dribbling Shuttle test with the Random Forest model, 0.3 % for the Yo-Yo Intermittent Recovery Test with the XGB model and 36.7 % for the Goal Accuracy test with the XGB model). Furthermore, Shapley additive explanations analysis exhibited that the anthropometric parameters such as body mass, mesomorph, and ankle diameter are dominant in physical development. These results confirm the generally accepted concept of the dominant anthropometric profile of early maturing professional soccer players. Additionally, it was understood that the hematology parameters of red blood count and mean corpuscular volume significantly influence physical development. Overall, explainable AI-based methods are more effective in predicting young athletes' physical development and performance. Therefore, the findings can effectively be incorporated with fitness assessment interventions and enhance related regional policy development.
... The factors interacting with RPE are TD, PWvh, and ACh, and likewise, with RPE dur , the interacting factors are DUR, TD, and PWvh. Other studies underline the importance of using more than one EXLD parameter to forecast more precisely the perceived exertion [36,37,66]. TD and DCm (per minute) can predict RPE, and likewise, player load, HSR, and ACm can predict REP dur [36]. ...
Article
Full-text available
Perceived exertion (RPE, RPEdur) and wellness scores (Hooper) are common methods to assess the training load and readiness in football. However, in professional football, there is a lack of data concerning the application of these tools in microcycles of different durations. Therefore, the purpose of this study was to explore the correlations between (a) RPE, RPEdur, and weekly external load (EXLD) and (b) Hooper and EXLD within microcycles of different durations. Independent of the microcycle’s length, RPE and RPEdur correlated with the EXLD. In MIC5, RPE and RPEdur correlated with the EXLD, except for accelerations > 3 m/s2 (ACh) and distance > 25.5 km/h (SPvh). In MIC6, RPE and RPEdur correlated with the EXLD, except Ach (with RPE). In MIC7 and MIC9, RPE and RPEdur correlated with the EXLD, except RPE with ACh. For Hooper variables, independent of the microcycle duration, SRN and FTG had more correlations in comparison to SLP; the same finding was observed in MIC7 and MIC9, but SRN and SLP in MIC5 and MIC6, respectively, had no correlations. Multiple regression analysis showed that for RPE, the main predictors are total distance (TD), metabolic power distance > 50 w/kg (PWvh), and ACh, which can explain 31% of the variance, but for RPEdur, 58% can be explained by DUR, TD, and PWvh. For SLP, the predictors FTG and SRN can explain a very small percentile of the variance (1%, 4%, and 6%, respectively). The main findings were the multiple correlations between RPE, RPEdur, Hooper, and microcycles of different durations; most of them were found for long durations. This tendency is similar to that described in the literature for microcycles of the same duration. (a) EXLD has a stronger correlation with RPEdur than with RPE, and (b) FTG and SRN are the best predictors of the EXLD during the upcoming training session. It is recommended that fitness coaches use these tools to assess and forecast the training load; thus, they will be able to prepare more effective sessions during microcycles of different durations.
Article
Sports data analysis can revolutionize how coaches and athletes train, leading to enhanced skills and improved team outcomes by providing valuable insights into performance metrics, enabling personalized training programs, and fostering a data-driven approach to decision-making. Sports data analysis has evolved in tandem with the increasing availability of data and the widespread adoption of data-driven practices in sports. Futsal stands out as one of the most challenging team sports to analyze, particularly in the case of female futsal, which has received limited research attention. The high-paced nature of the game, the smaller playing area, and the emphasis on close ball control necessitate a specialized approach to data collection and analysis. This situation presents a significant opportunity for in-depth exploration. By recognizing a gap in utilizing nutrition and physical activity data for female futsal players, researchers embarked on the journey to design and develop a recommendation system based on diet and training data. The study enlisted 14 talented female futsal players, and data was collected using an advanced LPS (Local Positioning System) device. Despite having limited player information, researchers successfully addressed the well-known “cold start” challenge. They created a content-based filtering recommendation system that accurately predicts the caloric expenditure of futsal players, achieving an impressive determination coefficient of 0.94. This innovative system has the potential to revolutionize the training methods of female futsal players, paving the way for advancements in sports data analysis and opportunities to enhance the visibility of women’s futsal on a broader stage.
Article
Monitoring fatigue during resistance training is essential to avoid injuries caused by overtraining. Fatigue can be comprehensively quantified by the external and internal load, where the external load is the work done by the athlete, and the internal load is the psychological and physiological response to the external load. This paper proposes a computer vision method to continuously monitor fatigue during resistance training by predicting external and internal parameters, namely the generated power and the rating of perceived exertion. We utilize the human pose estimation from two Microsoft Azure Kinect cameras to capture the movement of athletes while performing stationary exercises. Our method processes the obtained kinematic data, computes skeleton features to train traditional machine learning algorithms, and constructs feature maps to train convolutional neural network-based models to predict the load parameters. For evaluation, we recorded a dataset of 16 subjects who performed squat exercises on a Flywheel and rated their perceived exertion after each set. A measuring unit integrated into the Flywheel provided power readings for each repetition. The results show that our method achieves good results in predicting both parameters. Gradient Boosting Regression Trees best predicted perceived exertion with a mean absolute percentage error of 8.08% and a Spearman's = 0.74. Multi-layer Perceptron performed best in predicting power with a mean absolute error of 23.13 Watts and = 0.79. Our findings show that our approach delivers promising external and internal load quantifications for fatigue, with great potential to provide external feedback to coaches or athletes.
Conference Paper
Full-text available
The use of RPE as a measure of Internal load has become a common methodology used in team sports owing to its low cost. The aim of this study was to build a machine learning process able to describe the players' RPE by the external load extracted from the GPS. In this paper, we propose a multidimensional approach to assess the RPE in professional soccer which is based on GPS measurements and machine learning. By using GPS tracking technology, we collect data describing the training workload of players in a professional soccer club during a season. We show that our Ordinal predictor is both accurate and precise in medium RPE value (i.e., between 4 and 7) but it is not consistent in etreme value (i.e., below 4 and above 7). Our approach is a preliminary study that suggest that it is possible to predict players' RPE from GPS training and match data. However, these are not the only information needed to understand the players' effort perceived after a trainings or matches.
Conference Paper
Full-text available
In order to achieve optimal gains from training while reducing the risk of negative effects, athletes must perceive an ideal Internal Training Load. This load is invoked by the training content and individual characteristics. Training sessions are often team-based, although each athlete has his/her own characteristics, making it hard for coaching staff to optimize the training load for each athlete individually. In this paper, we propose a methodology for a decision support system that predicts three different sRPE scores: an average group score and an individualized score prior to and after training. This system can be used to personalize training sessions and to follow-up the training in a real-time fashion. We report results on a dataset collected from association football (soccer) players of a Belgian club. The reported results are better than current state-of-the-art from literature in the Australian football domain, with a smaller dataset and less invasive features.
Article
Full-text available
The ability of machine learning techniques to predict athlete ratings of perceived exertion (RPE) was investigated in professional Australian football players. RPE is commonly used to quantifying internal training loads and manage injury risk in team sports. Data from global positioning systems, heart-rate monitors, accelerometers and wellness questionnaires were recorded for each training session (n=3398) from 45 professional Australian football players across a full season. A variety of modelling approaches were considered to investigate the ability of objective data to predict RPE. Models were compared using nested cross validation and root mean square error (RMSE) on RPE predictions. A random forest model using player normalised running and heart rate variables provided the most accurate predictions (RMSE ± SD = 0.96 ± 0.08 au). A simplification of the model using only total distance, distance covered at speeds between 18-24 km·h−1, and the product of total distance and mean speed provided similarly accurate predictions (RMSE ± SD = 1.09 ± 0.05 au), suggesting that running distances and speeds are the strongest predictors of RPE in Australian football players. The ability of non-linear machine learning models to accurately predict athlete RPE has applications in live player monitoring and training load planning.
Article
Full-text available
We read with great interest the recent letter, “Time to bin the term ‘overuse’ injury: is ‘training load error’ a more accurate term?”1 and in particular its associated PostScript correspondence, “Are rolling averages a good way to assess training load for injury prevention?”2 We are currently investigating the association between training loads and injury risk,3 and so we have also been considering the best way to model this relationship. We share Dr Menaspa's concerns regarding the use of rolling averages for the calculation of ‘acute’ and ‘chronic’ loads. Namely, that they fail to account for the decaying nature of fitness and …
Article
Full-text available
Background: In professional senior soccer, training load monitoring is used to ensure an optimal workload to maximize physical fitness and prevent injury or illness. However, to date, different training load indicators are used without a clear link to training outcomes. Objective: The aim of this systematic review was to identify the state of knowledge with respect to the relationship between training load indicators and training outcomes in terms of physical fitness, injury, and illness. Methods: A systematic search was conducted in four electronic databases (CINAHL, PubMed, SPORTDiscus, and Web of Science). Training load was defined as the amount of stress over a minimum of two training sessions or matches, quantified in either external (e.g., duration, distance covered) or internal load (e.g., heart rate [HR]), to obtain a training outcome over time. Results: A total of 6492 records were retrieved, of which 3304 were duplicates. After screening the titles, abstracts and full texts, we identified 12 full-text articles that matched our inclusion criteria. One of these articles was identified through additional sources. All of these articles used correlations to examine the relationship between load indicators and training outcomes. For pre-season, training time spent at high intensity (i.e., >90 % of maximal HR) was linked to positive changes in aerobic fitness. Exposure time in terms of accumulated training, match or combined training, and match time showed both positive and negative relationships with changes in fitness over a season. Muscular perceived exertion may indicate negative changes in physical fitness. Additionally, it appeared that training at high intensity may involve a higher injury risk. Detailed external load indicators, using electronic performance and tracking systems, are relatively unexamined. In addition, most research focused on the relationship between training load indicators and changes in physical fitness, but less on injury and illness. Conclusion: HR indicators showed relationships with positive changes in physical fitness during pre-season. In addition, exposure time appeared to be related to positive and negative changes in physical fitness. Despite the availability of more detailed training load indicators nowadays, the evidence about the usefulness in relation to training outcomes is rare. Future research should implement continuous monitoring of training load, combined with the individual characteristics, to further examine their relationship with physical fitness, injury, and illness.
Article
Full-text available
Purpose: The aim of this study was to quantify and predict relationships between RPE and GPS training load variables in professional Australian Football (AF) players using group and individualised modelling approaches. Methods: Training load data (GPS and RPE) for 41 professional AF players was obtained over a period of 27 weeks. A total of 2711 training observations were analysed with a total of 66 ±13 sessions per player (range; 39 to 89). Separate generalised estimating equations (GEE) and artificial neural network analyses (ANN) were conducted to determine the ability to predict RPE from training load variables (i.e. session distance, high-speed running (HSR), high-speed running %, m·min-1) on a group and individual basis. Results: Prediction error for the individualised ANN (root mean square error [RMSE]; 1.24 ±0.41) was lower than the group ANN (RMSE; 1.42 ±0.44), individualised GEE (RMSE; 1.58 ±0.41) and group GEE (RMSE; 1.85 ±0.49). Both the GEE and ANN models determined session distance as the most important predictor of RPE. Further, importance plots generated from the ANN revealed session distance was most predictive of RPE in 36 of the 41 players, whereas, HSR was predictive of RPE in just 3 players and m·min-1 as predictive as session distance in just 2 players. Conclusions: This study demonstrates that machine learning approaches may outperform more traditional methodologies with respect to predicting athlete responses to training load. These approaches enable further individualisation of load monitoring, leading to more accurate training prescription and evaluation.
Chapter
Full-text available
Article
Purpose: Machine learning may contribute to understanding the relationship between the external load and internal load in professional soccer. Therefore, the relationship between external load indicators and the rating of perceived exertion (RPE) was examined using machine learning techniques on a group and individual level. Methods: Training data were collected from 38 professional soccer players over two seasons. The external load was measured using global positioning system technology and accelerometry. The internal load was obtained using the RPE. Predictive models were constructed using two machine learning techniques, artificial neural networks (ANNs) and least absolute shrinkage and selection operator (LASSO), and one naive baseline method. The predictions were based on a large set of external load indicators. Using each technique, one group model involving all players and one individual model for each player was constructed. These models' performance on predicting the reported RPE values for future training sessions was compared to the naive baseline's performance. Results: Both the ANN and LASSO models outperformed the baseline. Additionally, the LASSO model made more accurate predictions for the RPE than the ANN model. Furthermore, decelerations were identified as important external load indicators. Regardless of the applied machine learning technique, the group models resulted in equivalent or better predictions for the reported RPE values than the individual models. Conclusions: Machine learning techniques may have added value in predicting the RPE for future sessions to optimize training design and evaluation. Additionally, these techniques may be used in conjunction with expert knowledge to select key external load indicators for load monitoring.
Article
Purpose: The aims of this study are to determine match exertion, subsequent recovery and to investigate to what extent the coach is able to estimate players' match exertion and recovery. Methods: Rate of perceived exertion (RPE) and Total quality of recovery (TQR) of 14 professional basketball players (age 26.7±3.8 y, height 197.2±9.1 cm, weight 100.3±15.2 kg, body fat 10.3±3.6 %) were compared with observations of the coach. During an in-season phase of 15 matches within 6 weeks, players gave RPE after each match. TQR scores were filled out before the first training session after the match. The coach rated observed exertion (ROE) and recovery (TQ-OR) of the players. Results: RPE was lower than ROE (15.6±2.3 and 16.1±1.4; p=0.029). Furthermore, TQR was lower than TQ-OR (12.7±3.0 and 15.3±1.3; p<0.001). Correlations between coach' and players' exertion and recovery were r=.25 and r=.21, respectively. For recovery within 1 day the correlation was r=.68 but for recovery after 1-2 days no association existed. Conclusion: Players perceive match exertion hard to very hard and subsequent recovery reasonable. The coach overestimates match exertion and underestimates degree of recovery. Correspondence between coach and players is thus not optimal. This mismatch potentially leads to inadequate planning of training sessions and performance decrease during fixture congestion in basketball.
Article
Research describing load monitoring techniques for team sport is plentiful. Much of this research is conducted retrospectively, and typically involves recreational or semi-elite teams. Load monitoring research conducted on professional team sports is largely observational. Challenges exist for the practitioner in implementing peer-reviewed research into the applied setting. These challenges include match scheduling, player adherence, manager/coach buy-in, sport traditions, and staff availability. External load monitoring often attracts questions surrounding technology reliability and validity, while internal load monitoring makes some assumptions about player adherence as well as having some uncertainty around the impact these measures have on player performance This commentary outlines examples of load monitoring research, discusses the issues associated with the application this research in an elite team sport setting, and suggests practical adjustments to the existing research where necessary.