ArticlePDF Available

Abstract and Figures

Injuries have a great impact on professional soccer, due to their large influence on team performance and the considerable costs of rehabilitation for players. Existing studies in the literature provide just a preliminary understanding of which factors mostly affect injury risk, while an evaluation of the potential of statistical models in forecasting injuries is still missing. In this paper, we propose a multi-dimensional approach to injury forecasting in professional soccer that 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 then construct an injury forecaster and show that it is both accurate and interpretable by providing a set of case studies of interest to soccer practitioners. Our approach opens a novel perspective on injury prevention, providing a set of simple and practical rules for evaluating and interpreting the complex relations between injury risk and training performance in professional soccer.
Content may be subject to copyright.
Effective injury forecasting in soccer with GPS
training data and machine learning
Alessio Rossi
*, Luca Pappalardo
, Paolo Cintia
, F. Marcello Iaia
, Javier Fernàndez
Daniel Medina
1Department of Computer Science, University of Pisa, Pisa, Italy, 2ISTI, National Research Council, Pisa,
Italy, 3Department of Biomedical Science for Health, University of Milan, Milan, Italy, 4Sports Science and
Health Department, FC Barcelona, Barcelona Spain, 5Athletic Care Department, Philadelphia 76ers,
Philadelphia, Pennsylvania, United States of America
Injuries have a great impact on professional soccer, due to their large influence on team per-
formance and the considerable costs of rehabilitation for players. Existing studies in the liter-
ature provide just a preliminary understanding of which factors mostly affect injury risk, while
an evaluation of the potential of statistical models in forecasting injuries is still missing. In
this paper, we propose a multi-dimensional approach to injury forecasting in professional
soccer that 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 soc-
cer club during a season. We then construct an injury forecaster and show that it is both
accurate and interpretable by providing a set of case studies of interest to soccer practition-
ers. Our approach opens a novel perspective on injury prevention, providing a set of simple
and practical rules for evaluating and interpreting the complex relations between injury risk
and training performance in professional soccer.
Injuries of professional athletes have a great impact on the sports industry, due to their influ-
ence on the mental state of the individuals and the performance of a team [1,2]. Furthermore,
the cost associated with a player’s recovery and rehabilitation is often considerable, both in
terms of medical care and missed earnings deriving from the popularity of the player himself
[3]. Recent research demonstrates that injuries in Spain cause about 16% of season absence by
professional soccer players, corresponding to a cost of around 188 million euros per season
[4]. It is not surprising, hence, that injury forecasting is attracting a growing interest from
researchers, managers, and coaches, who are interested in intervening with appropriate actions
to reduce the likelihood of injuries of their players.
Historically, academic work on injury forecasting has been deterred by the limited availabil-
ity of data describing the physical activity of players. Nowadays, the Internet of Things have
the potential to change rapidly this scenario thanks to Electronic Performance and Tracking
Systems (EPTS), new tracking technologies that provide high-fidelity data streams extracted
from every training and game session [5,6]. These data depict in detail the movements of
PLOS ONE | July 25, 2018 1 / 15
Citation: Rossi A, Pappalardo L, Cintia P, Iaia FM,
Fernàndez J, Medina D (2018) Effective injury
forecasting in soccer with GPS training data and
machine learning. PLoS ONE 13(7): e0201264.
Editor: Jaime Sampaio, Universidade de Tras-os-
Montes e Alto Douro, PORTUGAL
Received: February 8, 2018
Accepted: July 10, 2018
Published: July 25, 2018
Copyright: ©2018 Rossi et al. This is an open
access article distributed under the terms of the
Creative Commons Attribution License, which
permits unrestricted use, distribution, and
reproduction in any medium, provided the original
author and source are credited.
Data Availability Statement: The owner of the
data is an elite soccer club in Italy which wants to
remain anonymous and did not give the
permission to make the original data publicly
available. The club has the right to choose which
information, results and data can be made public
and has granted the access to these data to the
authors only for research aims. In accordance with
SoBigData Ethical Committee, we can provide upon
request transformed data that are processed in
such a way that it is not possible to re-identify the
subjects involved in the study. The transformed
data reflect the same data distribution of real data
players on the playing field [5,6] and have been used for many purposes, from identifying
training patterns [7] to automatic tactical analysis [5,8,9]. Despite this wealth of data, little
effort has been put on investigating injury forecasting in professional soccer so far [10,11,12].
State-of-the-art approaches provide just a preliminary understanding of which variables affect
the injury risk, while an evaluation of the potential of statistical models to forecast injuries is
still poor. A major limit of existing studies is that they are mono-dimensional, i.e., they use just
one variable at a time to estimate injury risk, without fully exploiting the complex patterns
underlying the available data.
Professional soccer clubs are interested in practical, usable and interpretable models as a
decision making support for coaches and athletic trainers [13]. In this perspective the creation
of injury forecasting models poses many challenges. On one hand, injury forecasters must be
highly accurate, as models which frequently produce “false alarms” are useless. On the other
hand, a “black box” approach (e.g., a deep neural network) is not desirable for practical use
since it does not provide any insights about the reason behind the injuries. It goes hence with-
out saying that injury forecasting models must achieve a good tradeoff between accuracy and
In this paper, we consider injury prediction as the problem of forecasting that a player will
get injured in the next training session or official game, given his recent training workload.
We observe that existing mono-dimensional approaches are not effective in practice due to
their low precision (<5%), and we propose a multi-dimensional, easy-to-interpret and fully
data-driven approach which forecasts injuries with a better precision (50%); we validate this
result by simulating the usage of our forecaster over a season, with new training data available
as the season goes by. Our approach is entirely based on automatic data collection through
standard GPS sensing technologies and can be a valid supporting tool to the decision making
of a soccer club’s staff. This is crucial since the decisions of managers and coaches, and hence
the success of soccer clubs, also depend on what they measure, how good their measurements
are, the quality of predictions and how well these predictions are understood.
Related work
The relationship between training workload and injury risk has been widely studied in the
sports science literature [14,15,16,17,18]. For example Gabbett et al. [14,15,17,19] investi-
gate the case of rugby and find that a player has a high injury risk when his workloads are
increased above certain thresholds. To assess injury risk in cricket, Hulin et al. [20] propose
the Acute Chronic Workload Ratio (ACWR), i.e., the ratio between a player’s acute workload
and his chronic workload. When the acute workload is lower than the chronic workload,
cricket players are associated with a low injury risk. In contrast, when the acute/chronic ratio
is higher than 2, players have an injury risk from 2 to 4 times higher than the other group of
players. Hulin et al. [20] and Ehrmann et al. [11] find that injured players, in both rugby and
soccer, show significantly higher physical activity in the week preceding the injury with respect
to their seasonal averages.
In skating, Foster et al. [21] measure training workload by the session load, i.e., the product
of the perceived exertion and the duration of the training session. When the session load out-
weighs a skater’s ability to fully recover before the next session, the skater suffers from the so-
called “overtraining syndrome”, a condition that can cause injury [21]. In basketball, Anderson
et al. [18] find a strong correlation between injury risk and the so-called monotony, i.e., the
ratio between the mean and the standard deviation of the session load recorded in the past 7
days. Moreover, Brink et al. [8] observe that injured young soccer players (age <18) recorded
higher values of monotony in the week preceding the injury than non-injured players.
Injury forecasting in soccer
PLOS ONE | July 25, 2018 2 / 15
to guarantee that the experiments performed on
the transformed data produce the same results as
the ones shown in the paper. We specify that the
researchers from FC Barcelona participated to the
study only as collaborators, and that FC Barcelona
is not the owner of the data. For data requests
please contact us at the following email addresses: or
Funding: This work is partially supported by the
European Community’s H2020 Program under the
funding scheme “INFRAIA-1-2014-2015: Research
Infrastructures” grant agreement 654024, www., “SoBigData”. The funders had no
role in study design, data collection and analysis,
decision to publish, or preparation of the
manuscript. There was no additional external
funding received for this study.
Competing interests: The authors have declared
that no competing interests exist.
Venturelli et al. [12] perform several periodic physical tests on young soccer players
(age <18) and find that jump height, body size and the presence of previous injuries are signif-
icantly correlated with the probability of thigh strain injury. Talukder et al. [22] create a classi-
fier to predict 19% of the injuries that occurred in NBA. They also show that the most
important features for predicting injuries are the average speed, the number of past competi-
tions played, the average distance covered, the number of minutes played to date and the aver-
age field goals attempted. An attempt to injury forecasting in soccer has been made by
Kampakis [23], although it considers a reduced set of features obtaining an accuracy that is, in
the best scenario, not significantly better than random classifiers.
Material and method
Data collection and feature extraction
We set up a study on twenty-six Italian professional male players (age = 26±4 years;
height = 179±5 cm; body mass = 78±8 kg) during season 2013/2014. Six central backs, three
fullbacks, seven midfielders, eight wingers and two forwards were recruited. Participants gave
their written informed consent to participate in the study.
We monitored the physical activity of players during 23 weeks–from January 1st to May
31st, 2014 –using portable 10 Hz GPS devices integrated with a 100Hz 3-D accelerometer, a
3D gyroscope, a 3D digital compass (STATSports Viper). The devices were placed between the
players’ scapulae through a tight vest. We recorded a total of 931 individual training sessions
during the 23 weeks. From the data collected by the devices, we extracted a set of training
workload indicators through the software package Viper Version 2.1 provided by STATSports
The club’s medical staff recorded 23 non-contact injuries during the study. According to
the UEFA regulations [24], a non-contact injury is defined as any tissue damage sustained by a
player that causes absence in physical activities for at least the day after the day of the onset.
We observed that 19 out of 23 injuries are associated with players who got injured at least once
in the past. In particular, half of the players never get injured during the study, while the others
get injured once (seven players), twice (five players) or four times (one player). For every
player, we collected information about age, body mass index, height and role on the field.
Moreover, for every single training session of a player, we collected information about the play
time in the official game before the training session and the number of official games played
before the training session.
From the players’ GPS data we extract 12 features describing different aspects of the work-
load in a training session [25]. Two features–Total Distance (d
) and High Speed Running
Distance (d
)–are kinematic, i.e., they quantify a player’s overall movement during a train-
ing session. Three features–Metabolic Distance (d
), High Metabolic Load Distance (d
and High Metabolic Load Distance per minute (d
)–are metabolic, i.e., they quantify the
energy expenditure of a player’s overall movement during a training session. The remaining
seven features–Explosive Distance (d
), number of accelerations above 2m/s2 (Acc
), num-
ber of accelerations above 3m/s2 (Acc
), number of decelerations above 2m/s2 (Dec
), number
of decelerations above 3m/s2 (Dec
), Dynamic Stress Load (DSL) and Fatigue Index (FI)–are
mechanical features describing a player’s overall muscular-scheletrical load during a training
session. In addition, we associated a player’s training session with feature PI, indicating the
number of the player’s previous injuries up to that session. Table 1 and S1 Appendix provides
the description and some statistics of the workload features extracted from the GPS data,
Injury forecasting in soccer
PLOS ONE | July 25, 2018 3 / 15
Multi-dimensional and data-driven injury forecaster
We construct a multi-dimensional model to forecast whether or not a player will get injured
based on his recent training workload. The construction of the injury forecaster consists of
two phases. In the first phase (training dataset construction), given a set of features S, a training
dataset Tis created where each example refers to a single player’s training session and consists
of: (i) a vector of features describing both the player’s personal features and his recent work-
load, including the current training session; (ii) an injury label, indicating whether (1) or not
(0) the player gets injured in the next game or training session. In the second phase (model
construction and validation), a decision tree learner is used to train an injury classifier on the
training dataset T.
Phase 1: Training dataset construction. From the features extracted from GPS data,
which are described in Table 1, we construct a training dataset Tconsisting of 55 features and
952 examples (i.e., individual training sessions). S4 Appendix provides an example of the con-
struction of T. These 55 features are:
18 daily features:the 12 workload features extracted from the GPS data and the 6 personal
features described in Table 1.
12 EWMA features: 12 features computed as the Exponential Weighted Moving Average
(EWMA) of the 12 workload features in Table 1. The EWMA decreases exponentially the
weights of the values according to their recency, i.e., the more recent a value, the more it is
weighted in an exponential function according to a decay α= 2/(span+1). In our experi-
ments we consider a span equal to six (see S5 Appendix).
12 ACWR features:12 features consisting of the ACWR of the 12 workload features in
Table 1. Given a feature, the ACWR of a player is the ratio between (i) the player’s acute
Table 1. Training workload features used in our study. Description of the training workload features extracted from
GPS data and the players’ personal features collected during the study. We defined four categories of features: kine-
matic features (blue), metabolic features (red), mechanical features (green) and personal features (white).
Distance in meters covered during the training session
Distance in meters covered above 5.5m/s
Distance in meters covered at metabolic power
Distance in meters covered by a player with a Metabolic Power is above 25.5W/Kg
Distance in meters covered by a player with a Metabolic Power is above 25.5W/Kg per minute
Distance in meters covered above 25.5W/Kg and below 19.8Km/h
Number of accelerations above 2m/s2
Number of accelerations above 3m/s2
Number of decelerations above 2m/s2
Number of decelerations above 3m/s2
DSL Total of the weighted impacts of magnitude above 2g. Impacts are collisions and step impacts during
FI Ratio between DSL and speed intensity
Age age of players
BMI Body Mass Index: ratio between weight (in kg) and the square of height (in meters)
Role Role of the player
PI Number of injuries of the players before each training session
Minutes of play in previous games
Games Number of games played before each training session
Injury forecasting in soccer
PLOS ONE | July 25, 2018 4 / 15
workload, computed as the average of the values of the feature in the last 6 days; (ii) the play-
er’s chronic workload, computed as the average of the values of the feature in the last 27 days
12 MSWR features:12 features consisting of the monotony of the 12 workload features in
Table 1. Given a feature, the monotony of a player is the ratio between the mean and the
standard deviation of the values of the feature in the last week [3,10,18].
1 previous injury feature:to take into account both the number of a player’s previous inju-
ries and their distance to the current training session we compute feature PI
, the EWMA
of feature PI computed with a span equal to six. PI
reflects the distance between the cur-
rent training session and the training session when the player returned to regular training
after an injury. PI
= 0 indicates that the player never got injured in the past; PI
indicates that he got injured at least once in the past; PI
>1 indicate that he got
injured more than once in the past (see S6 Appendix).
We select 30% of Tand obtain T
(step 1 and 2 in Fig 1) to perform a feature selection
process to determine the most relevant features for classification using Recursive Feature Elim-
ination with Cross-Validation (RFECV; we use the publicly available Python package scikit-
learn to perform RFECV and to train and validate the decision tree–
[27]. In RFECV, the subset of features producing the maximum score on the validation data is
considered to be the best feature subset [27]. The feature selection process is aimed at reducing
the dimensionality of the feature space and hence the risk of overfitting, and allowing for an
easier interpretation of the resulting machine learning model, due to the lower number of fea-
tures [28].
The class distribution in training dataset T
is highly unbalanced since we have 279
non-injury examples and just 7 injury examples. To adjust this imbalance we oversample the
minority class in T
by using the adaptive synthetic sampling approach (ADASYN; We
use the ADASYN function provided by the publicly available Python package imblearn– [29]. The ADASYN algorithm generates examples of
the minority class to equalize the distribution of classes, hence reducing the learning bias (See
S7 Appendix). Finally, we use T
to detect the best hyper parameters of a decision tree
classifier DT (Step 2 in Fig 1).
Phase 2: Model construction and validation. We then split T
into two folds, f
, in order to perform a stratified cross validation (step 3 in Fig 1; we use only two folds in
order to not excessively reduce the minority class size). In this step, we oversample fold f
using ADASYN and test DT on the other fold f
(which is not oversampled). For cross valida-
tion purposes, we perform again step 3 inverting f
and f
. The goodness of the forecasting
model is evaluated by four metrics (i.e., precision, recall, F1-score and AUC) described in S8
Appendix. Note that, for injury forecasting purposes, we are interested in achieving high values
of precision and recall on class 1 (injury). Let us assume that a coach makes a decision about
whether or not to “stop” a player based on the suggestion of the injury forecaster, i.e., the
player skips next training session or game every time the forecaster’s prediction associated
with the player’s current training session is 1 (injury). In this scenario, the forecaster’s preci-
sion indicates how much we can trust the predictions: the higher the precision, the more a clas-
sifier’s predictions are reliable, i.e., the probability that the player will actually get injured is
high. Trusting an injury forecaster with low precision is risky as it means producing many
false positives (i.e., false alarms) and frequently stopping players unnecessarily, a condition
clubs want to avoid especially for the key players. The recall indicates the fraction of injuries
the forecaster detects over the total number of injuries: the higher the recall the more injuries
Injury forecasting in soccer
PLOS ONE | July 25, 2018 5 / 15
the forecaster can detect. An injury forecaster with low recall detects just a small fraction of the
injuries, meaning that many players will attend next training session or game and actually get
injured. Trusting a forecaster with a low recall is risky as it would misclassify many actual inju-
ries as non-injuries.
We repeated the entire injury prediction approach (i.e., all the three steps in Fig 1) 10,000
times in order to assess its stability with respect to the choice of the injury examples in the two
folds. For the sake of comparison, we implemented other injury forecasters based on the
ACWR and the monotony (or MSWR) techniques, which are among the two most used tech-
niques for injury risk estimation and prediction in professional soccer (see S2 Appendix and
S3 Appendix for details). Moreover, we compare our injury forecaster with four baselines.
Baseline B
randomly assigns a class to an example by respecting the distribution of classes.
Baseline B
always assigns the non-injury class, while baseline B
always assigns the injury
class. Baseline B
is a classifier which assigns class 1 (injury) if PI
>0, and 0 (no injury)
otherwise. We also compare DT with a Random Forest classifier (RF) and a Logit classifier
Fig 1. Construction of the training dataset and the forecasting model. In step 1 we split the dataset into two parts: T
(30% of T) and T
(70% of T). We
then oversample the minority class in T
by using ADASYN, select the most important features and fit the hyper parameters (Step 2). We then split T
two folds in order to perform a stratified cross validation (step 3).
Injury forecasting in soccer
PLOS ONE | July 25, 2018 6 / 15
Table 2 compares the performance of DT with the performance of RF, LR, the ACWR and
MSWR forecasters, and the four baselines. The results in Table 2 refer to the mean and the stan-
dard deviation of the evaluation metrics over 10,000 cross validation tasks. We find that DT has
recall = 0.80±0.07 and precision = 0.50±0.11 on the injury class, meaning that the decision tree
can predict almost all the injuries (80%) and that it correctly labels a training session as an injury
in 50% of the cases. This is a significant improvement with respect to both the baselines B
,. . .,
for which the maximum precision is about 6%, and the ACWR- and MSWR-based injury
forecasters, for which the maximum precision is lower than 4%. RF has better recall but worse
precision (recall = 0.87±0.05, precision = 0.41±0.08) that DT, while LR has much lower perfor-
mance than the decision tree (Table 2). These results show that, typically, DT drastically reduces
false alarms and hence scenarios where players are “stopped” unnecessarily before next game or
training session. On the one hand, the distributions of the forecasters’ performances over the
10,000 tests indicate that the quality of the injury forecasting strongly depends on the type of
injuries in the training set, which in turn depends on the different training and test split made
in each trial (Fig 2). On the other hand, the higher performance detected by DT, compared to
several baselines and the ACWR- and MSWR-based injury forecasters, shows that our approach
outperforms state-of-the-art approaches and achieve good results in forecasting injuries. The
results for DT without ADASYN and the oversampling process are presented in S9 Appendix.
As a further test of the forecasting potential of our approach we investigate the benefit of
using our multi-dimensional injury forecaster in a real-world injury prevention scenario,
where we assume that a club equips with appropriate GPS sensor technologies and starts
recording training workload data since the first training session of the season (in other words,
no data are available to the club before the beginning of the season). Assuming that we train
the injury forecaster with new data every week, how many injuries the club can actually pre-
vent throughout the season?
Table 2. Performance of DT compared to RF, LR, the four baselines and the ACWR- and MSWR-based forecasters. For each forecaster we report precision, recall
and F1 on the two classes and the overall AUC.
precision recall F1 AUC
DT NI 0.96±0.05 0.87±0.09 0.91±0.04 0.76±0.12
I0.50±0.11 0.80±0.07 0.64±0.10
RF NI 0.94±0.06 0.90±0.08 0.93±0.07 0.78±0.15
I0.41±0.08 0.87±0.05 0.65±0.08
LR NI 0.69±0.11 0.61±0.15 0.65±0.13 0.60±0.03
I0.18±0.03 0.60±0.08 0.31±0.06
B4 NI 0.98 0.77 0.86 0.60
I0.04 0.43 0.07
B1 NI 0.98 0.98 0.98 0.51
I0.06 0.05 0.05
B2 NI 0.98 1.00 0.99 0.50
I0.00 0.00 0.00
B3 NI 0.00 0.00 0.00 0.50
I0.02 1.00 0.04
NI 1.00 0.43 0.60 0.67
I0.04 0.91 0.07
NI 0.98 0.80 0.88 0.57
I0.04 0.33 0.07
Injury forecasting in soccer
PLOS ONE | July 25, 2018 7 / 15
To answer this question we group the training sessions by week and proceed from the least
recent to the most recent week. At training week w
we first construct the dataset T
of all the training examples collected up to week i, oversampling the injury examples through
ADASYN and reducing the feature space through RFECV. Then, we use T
to train DT
, RF
, B
1,i,. . .,
, the ACWR- and MSWR-based forecasters and try to predict the injuries in
week w
. At week i, we evaluate the accuracy of our approach by the cumulative F1-score,
i.e., the F1-score computed by considering all the predictions made up to week iby the models
,. . ., DT
. Due to the initial scarcity of data, we start the forecasting task from week w
Fig 3 and S7 Table show the evolution of the cumulative F1-score and the feature extracted
by RFECV as the season goes by, respectively. We find that in the first weeks DT has a poor
predictive performance and misses many injuries (the black crosses in Fig 3). The predictive
ability of DT improves significantly throughout the season: as more and more training and
injury examples are collected, the forecasting model predicts most of the injuries in the second
half of the season (the red crosses in Fig 3). We find that DT is the one performing the best,
outperforming all the other models from week w
. In particular, DT detects 9 injuries out of
14 from w
to the end of the season, resulting in F1-score = 0.60 and precision = 0.56. After an
initial period of data collection, the injury forecaster becomes a useful tool to prevent the inju-
ries of players and, by extracting the rules from the decision tree as we show in the next section,
to understand the reasons behind the forecasted injuries as well as the injuries that are not
detected by the model.
Interpretation of the injury forecaster
A set of simple rules can be extracted from DT build on w
, allowing for the investigation of
the reasons behind the observed injuries. These rules can be seen as a short handbook for
coaches and athletic trainers, which can consult it to modify the training schedule and improve
the players’ fitness.
Fig 4B visualizes DT highlighting two types of node: decision nodes (black boxes) and leaf
nodes (green or red boxes). Each decision node has two branches each indicating the next
node to select depending on the range of values of the feature associated with the decision
node. A leaf node represents the final prediction based on a player’s individual training ses-
sion. There are two possible final decisions: Injury (red boxes) indicates that the player will get
injured in next game or training session; or No-Injury (green boxes) otherwise. Given a feature
Fig 2. Classifiers performances. Distributions of the classifiers—DT, LR and RF—performances obtained testing the algorithms 10,000 times. This figure shows the
performance of the baselines and the ACWR- and MSWR-based injury forecasters as well.
Injury forecasting in soccer
PLOS ONE | July 25, 2018 8 / 15
vector describing a player’s training session, the prediction associated with it is obtained by fol-
lowing the path from the root of the tree down to a leaf node, through the decision nodes. Fig
4shows the rules and the tree extracted from the DT built until w
. At the end of the season,
the RFECV process selects just 3 features out of 55: PI
, d
and d
The importances of these features in DT, computed as the mean decrease in Gini coefficient,
are 0.71, 0.23 and 0.06, respectively [30].
As a practical example of application of these rules, let us consider a player’s training session
with PI
= 0.28, d
=126.58 and d
= 1.66, associated with an injury.
This example is associated with rule 2 (Fig 4A), corresponding to the following decision path:
dHSRðEWMAÞ>112:35 !dTOTðMSWRÞ1:78 !PIðEWMAÞ>0:03 !PIðEWMAÞ0:68
From the rules in Fig 4A we summarize three main injury scenarios in DT:
1. a previous injury can lead to a new injury when a player has a HSR
(high speed run-
ning distance) lower than 112.35 (rule 1 in Fig 4A). This rule describes 42% of the injuries
in the dataset and it is correct in 100% of the cases.
2. a previous injury can lead to a new injury when a player has a HSR
higher than
112.35 and a D
(total distance Monotony) three times lower than 1.78 (rule 2 in
Fig 4A). This rule describes 30% of the injuries and has an accuracy of 100%.
Fig 3. Performance of forecasters in the evolutive scenario. As the season goes by, we plot week by week the cumulative F1-score of the forecasters DT, RF, LR, B
trained on the data collected up to that week. Black crosses indicate injuries that not detect by DT, red crosses indicate injures correctly predicted by DT. For every
week iwe highlight in red the number of injuries detected by DT up to week i.
Injury forecasting in soccer
PLOS ONE | July 25, 2018 9 / 15
3. a previous injury can lead a new injury when a player has a HSR
higher than 112.35
and a D
two and half times higher than the player’s average (rules 3 and 4 in Fig
4A). These rules have a cumulative frequency of 28% and a mean accuracy of 75±5%.
These scenarios suggest that coaches and athletic trainers must take care of the total dis-
tance and the distance at high speed running performed by the players who recently returned
to play after an injury.
Our experiments produce three remarkable results. First, DT can detect around 80% of the
injuries with about 50% precision, far better than the baselines and state-of-the-art injury risk
estimation techniques (see Table 2). The decision tree’s false positive rate is small, indicating
that it reduces the “false alarms”, i.e., situations where the classifier is wrong in predicting that
an injury will happen. In professional soccer, false alarms are deprecable because the scarcity
Fig 4. Interpretation of the multi-dimensional injury forecaster. (a) The six injury rules extracted from DT. For each rule we show the range of values of every
feature, its frequency (Freq) and accuracy (Acc). (b) A schematic visualization of decision tree. Black boxes are decision nodes, green boxes are leaf nodes for class No-
Injury, red boxes are leaf nodes for class Injury.
Injury forecasting in soccer
PLOS ONE | July 25, 2018 10 / 15
of players can negatively affect the performance of a team [2]. Our model also produces a mod-
erate false negative rate, meaning that situations where a player that will get injured is classified
as out of risk are infrequent.
Second remarkable results is that, in a real-world scenario of injury prevention where a
club starts collecting the data for the first time and re-train the injury forecaster as the season
goes by, the injury forecaster results in a cumulative F1-score = 0.60 on the injury class (Fig 3),
much better than the baselines, RF and LR (Table 2). Throughout the season, the usage of the
forecasting model allows for the prevention of more than half of the injuries. The forecasting
ability of DT is affected by the initial period where data are scarce. This suggests that an initial
period of data collection is needed in order to gather the adequate amount of data, and only
then a reliable forecasting model can be trained on the collected data. The length of the data
collection period depends on the club’s needs and strategy, including the frequency of training
sessions and games, the frequency of injuries, the number of available players and the tolerated
level of false alarms. Regarding this aspect, in our dataset, we observe that the performance of
the classifiers stabilizes after 14 weeks of data collection (see Fig 3).
Third, in the evolutive scenario the features selected change as the season goes by (see S7
Table). This is probably due to the initial scarcity of data and to the type of injuries that have
occurred up that a given moment. We observed that the just 3 out of 55 features are selected by
the feature selection (PI
, d
and d
) after 14 weeks of data collection,
and that these set of features remains stable for all subsequent weeks. Feature PI
, the
most important among the three and the only feature that is always selected as the season goes
by (see S7 Table), reflects the temporal distance between a player’s current training session and
his coming back to regular training after a previous injury. Less than half of the injuries
detected by DT in the evolutive scenario happened immediately after the coming back to regu-
lar training of injured player. Furthermore, 60% of the injuries detected by DT happened long
after a previous injury and are characterized by specific values of d
and d
which indicate that the a player’s kinematic variability affects his injury risk. It is worth to
notice that the single feature PI
alone does not provide a significant predictive power,
as the baseline B
, which is based on it, has a much lower accuracy than DT. It is hence the
combination of the three features which allows us to predict when a player will get injured.
Our results suggest that the club should take particular care of the first training sessions of
players who come back to regular training after a previous injury, as in this conditions they are
more likely to get injured again. In these first days and in the days long after the players return
to regular physical activity, the club should control kinematic workloads, which can lead to
injuries at specific values as well.
Injuries involve a great economic cost to the club, due to the expensive process of recovery
and rehabilitation for the players. Injury prevention can reduce these costs by avoiding the
injuries of players, which means improving the team’s performance and the player’s mental
state as well as reducing the seasonal costs of medical care. We estimate that 139 days of
absence during the seasons are due to injuries, corresponding to 6% of the working days. We
observe that a player returned to regular physical activity within 5 days (i.e., 15 times out of 23
injuries), while only 6 times a player needed more than 5 days to recover. We use a method
proposed in the literature [4] to estimate that the minimum total cost related to injuries that in
this soccer club is 11,583 euros (139x83 euros = days of absence x minimal legal salary per day)
corresponding to 3.81% of the salary cost of the club. If our model was used as the season goes
by to stop the players for which an injury is predicted, the club could had been able to prevent
9 injuries out of 14 and save 8,881 euros (107x83 euros = day of absence x minimal legal salary
per day), that represents a 77% decrease of injury costs.
Injury forecasting in soccer
PLOS ONE | July 25, 2018 11 / 15
In this paper we proposed a multi-dimensional approach to injury forecasting in soccer, fully
based on automatically collected GPS data and machine learning. As we showed, our injury
forecaster provides a good trade-off between accuracy and interpretability, reducing the num-
ber of false alarms with respect to state-of-the-art approaches and at the same time providing a
simple handbook of rules to understand the reasons behind the observed injuries. We showed
that the forecaster can be profitably used early in the season, and that it allows the club to save
a considerable part of the seasonal injury-related costs. Our approach opens a novel perspec-
tive on injury prevention, providing a methodology for evaluating and interpreting the com-
plex relations between injury risk and training performance in professional soccer.
Our work can be extended in many directions. First, we can include performance features
extracted from official games, where the player is exposed to the highest physical and psycho-
logical stress. Second, we can investigate the “transferability” of our approach from a club to
another, i.e., if a forecaster trained on a set of players can be successfully applied to a distinct
set of players, not used during the training process. In this case, it would be possible to exploit
collective information to train a more powerful forecaster which includes training examples
from different players, clubs, and leagues. Third, if data covering several seasons of a player’s
activity are available, a distinct forecaster can be trained for each player by combining GPS
data with other types of health data, such as heart rate, ventilation, and lactate.
Supporting information
S1 Appendix. Descriptive statistics of the workload features.
S2 Appendix. The ACWR method.
S3 Appendix. The MSWR method.
S4 Appendix. Example of the training dataset construction.
S5 Appendix. Exponential Weighted Moving Average (EWMA).
S6 Appendix. Computation of PI
S7 Appendix. Adaptive synthetic sampling approach.
S8 Appendix. Classifiers metrics assessment.
S9 Appendix. Predictions results.
S1 Table. Descriptive statistics of the 12 training workload features. We provide three cate-
gories of training workload features: kinematic features (blue), metabolic features (red) and
mechanical features (green).
Injury forecasting in soccer
PLOS ONE | July 25, 2018 12 / 15
S2 Table. Performance of ACWR predictor. We report precision (prec), recall (rec), F1-score
(F1) and Area Under the Curve (AUC) for the injury class and the non- injury class for all the
predictors based on ACWR and MSWR. We also provide predictive performance of four base-
line predictors B
and B
S3 Table. Injury prediction report of ACWRq. We report precision (prec), recall (rec),
F1-score (F1) and Area Under the Curve (AUC) for the injury class and the non-injury class
for all the predictors defined on ACWR and monotony methodologies. We also provide pre-
dictive performance of four baseline predictors B
and B
S4 Table. Performance of MSWR predictor. We report precision (prec), recall (rec), F1-score
(F1) and Area Under the Curve (AUC) for the injury class and the non- injury class for all the
predictors based on ACWR and MSWR. We also provide predictive performance of four base-
line predictors B
and B
S5 Table. PI
values after n training days (i.e., n= 1, . . ., 6) since the return of a player
to regular training. We report the values for different n of previous injuries (i.e., n= 1, . . ., 4).
is the number of training days long after players return to regular physical activity. 6+ indi-
cates values for 6 and more than 6 days.
S6 Table. Performance of the classifiers on T
, T and T
.For each classifier, we
report the precision (prec), recall (rec) and F1-score (F1) on the two classes and the overall
S7 Table. Feature selection real-world scenario. Features extracted by RFECV in each T
built as the season went by.
S1 Fig. Distribution of workload features. We provide three categories of training workload
features: kinematic features (blue), metabolic features (red) and mechanical features (green).
S2 Fig. Injury risk in ACWR groups. The plots show Injury Likelihood (IL) for pre- defined
ACWR groups [29], for every of the 12 training workload features considered in our study.
Bars are colored according to feature categorization defined in Table 1.
S3 Fig. Injury likelihood in ACWR groups. The plots show IL for the ACWR groups defined
the quantiles of the distribution, for every of the 12 training workload features considered in
our study. We provide three categories of training workload features: kinematic features
(blue), metabolic features (red) and mechanical features (green).
S4 Fig. Injury risk in MSWR groups. The plots show the Injury Likelihood (IL) for the
MSWR groups for every of the 12 training workload features considered in our study. Bars are
colored according to feature categorization defined in Table 1.
Injury forecasting in soccer
PLOS ONE | July 25, 2018 13 / 15
S5 Fig. We plot the AUC and F1-score of EWMA with span = 1, . . ., 10 in CALL. The red
line reflects the best span to injury prediction.
This work is partially supported by the European Community’s H2020 Program under the
funding scheme “INFRAIA-1-2014-2015: Research Infrastructures” grant agreement 654024,, “SoBigData”.
Author Contributions
Conceptualization: Alessio Rossi, Luca Pappalardo, Paolo Cintia.
Data curation: F. Marcello Iaia.
Formal analysis: Alessio Rossi.
Investigation: Alessio Rossi.
Methodology: Alessio Rossi.
Validation: Alessio Rossi, Luca Pappalardo, Paolo Cintia.
Writing – original draft: Alessio Rossi, Luca Pappalardo, Paolo Cintia.
Writing – review & editing: Alessio Rossi, Luca Pappalardo, Paolo Cintia, Javier Fernàndez,
Daniel Medina.
1. Ha
¨gglund M, Walde
´n M, Magnusson H, Kristenson H, Bengtsson H, Exstrand J. Injuries affect team
performance negatively in professional football: an 11-year follow-up of the UEFA Champions League
injury study. British Journal of Sports Medicine,, 2013.
PMID: 23645832
2. Hurley OA. Impact of Player Injuries on Teams’ Mental States, and Subsequent Performances, at the
Rugby World Cup 2015. Frontiers in Psychology 7:807,
PMID: 27375511
3. Lehmann EE, Schulze GG. What Does it Take to be a Star?–The Role of Performance and the Media
for German Soccer Players. Applied Economics Quarterly 54:1, pp. 59–70,
aeq.54.1.59, 2008.
4. Ferna
´ndez-Cuevas I., Gomez-Carmona P, Sillero-Quintana M, Noya-Salces J, Arnaiz-Lastras J, Pas-
´n A. Economic costs estimation of soccer injuries in first and second Spanish division profes-
sional teams. 15th Annual Congress of the European College of Sport Sciences ECSS, 23th 26th june.
5. Gudmundsson H, Horton M. Spatio-Temporal Analysis of Team Sports, ACM Computing Surveys 50,
2, Article 22 (April 2017), 34 pages.
6. Stein M, Janetzko H, Seebacher D, Ja
¨ger A, Nagel M, Ho
¨lsch J, et al. How to Make Sense of Team
Sport Data: From Acquisition to Data Modeling and Research Aspects. Data, 2:1, 2,
3390/data2010002, 2017.
7. Rossi A, Savino M, Perri E, Aliberti G, Trecroci A, Iaia M. Characterization of in-season elite football
trainings by GPS features: The Identity Card of a Short-Term Football Training Cycle. 16th IEEE Inter-
national Conference on Data Mining Workshops, pp. 160–166,
0030, 2016
8. Pappalardo L, Cintia P. Quantifying the relation between performance and success in soccer, Advances
in Complex Systems, 20 (4),, 2017.
9. Cintia P, Pappalardo L, Pedreschi D, Giannotti F, Malvaldi M. The harsh rule of the goals: data-driven
performance indicators for football teams, In Proceedings of the 2015 IEEE International Conference
on Data Science and Advanced Analytics (DSAA’2015),,
Injury forecasting in soccer
PLOS ONE | July 25, 2018 14 / 15
10. Brink MS, Visscher C, Arends S, Zwerver J, Post WJ, Lemmink KA. Monitoring stress and recovery:
new insights for the prevention of injuries and illnesses in elite youth soccer players. Br J Sports Med.
2010; 44: 809–15. PMID: 20511621
11. Ehrmann FE, Duncan CS, Sindhusake D, Franzsen WN, Greene DA. GPS and injury prevention in pro-
fessional soccer. J Strength Cond Res. 2015; 30:306–307.
0000000000001093 PMID: 26200191
12. Venturelli M, Schena F, Zanolla L, Bishop D. Injury risk factors in young soccer players detected by a
multivariate survival model. Journal of Science and Medicine in Sport. 2011; 14:293–298. https://doi.
org/10.1016/j.jsams.2011.02.013 PMID: 21474378
13. Kirkendall DT, Dvorak J. Effective Injury Prevention in Soccer. The physician and sports medicine,
38:1,, 2010.
14. Gabbett TJ. The development and application of an injury prediction model for noncontact, soft-tissue
injuries in elite collision sport athletes. The Journal of Strength & Conditioning Research. 2010; 24
(10):2593–2603. PMID: 20847703
15. Gabbett TJ, Ullah S. Relationship between running loads and soft-tissue injury in elite team sport ath-
letes. J Strength Cond Res. 2012; 26: 953–960.
PMID: 22323001
16. Rogalski B, Dawson B, Heasman J, Gabbett TJ. Training and game loads and injury risk in elite Austra-
lian footballers. J Sci Med Sport. 2013; 16: 499–503.
PMID: 23333045
17. Gabbett TJ. The training-injury prevention paradox: should athletes be training smarter and harder? Br
J Sports Med. 2016. PMID: 26758673
18. Anderson L, Triplett-McBride T, Foster C, Doberstein S, Brice G. Impact of training patterns on inci-
dence of illness and injury during a women’s collegiate basketball season. The Journal of Strength &
Conditioning Research. 2003; 17: 734–738.
19. Gabbett TJ. Reductions in pre-season training loads reduce training injury rates in rugby leagueplayers.
British Journal of Sports Medicine. 2004; 38: 74–749.
20. Hulin BT, Gabbett TJ, Blanch P, Chapman P, Bailey D, Orchard JV. Spikes in acute workload are asso-
ciated with increased injury risk in elite cricket fast bowlers. Br J Sports Med. 2014; 48:708–712. https:// PMID: 23962877
21. Foster C. Monitoring training in athletes with reference to overtraining syndrome. Med Sci Sports Exerc.
1998; 30:1164–1168. PMID: 9662690
22. Talukder H, Vincent T, Foster G, Hu C, Huerta J, Kumar A, et al. Preventing in-game injuries for NBA
players. MIT Sloan Analytics Conference. Boston; 2016.
23. Kampakis S. Predictive modeling of football injuries, Phd Thesis, University College London, 2016
24. Hagglund M, Walden M, Bahr R, Ekstrand J. Methods for epidemiological studyof injuries to profes-
sional football players: developing the UEFA model. British Journal of Sports Medicine, 39:6, 340–346,, 2005. PMID: 15911603
25. Duncan MJ, Badland HM, Mummery WK. Applying GPS to enhance understanding of transport-related
physical activity. Journal of Science and Medicine in Sport. 2009; 12: 549–556.
j.jsams.2008.10.010 PMID: 19237315
26. Murray NB, Gabbett TJ, Townshend AD, Blanch P. Calculation acute:chronic workloadratios using
exponential weighted moving averages provides a more sensitive indicator of injury likelihood than roll-
ing averages. Br J Sports Med. 2016. PMID: 28003238
27. Guyon I, Weston J, Barnhill S, Vapnik V. Gene Selection for Cancer Classification Using Support Vector
Machines. Machine Learning 46, 2002.
28. James G, Witten D, Hastie T, Tibshirani R. An Introduction to Statistical Learning. New York, NY:
Springer New York; 2013.
29. He H, Bai Y, Garcia EA, Li S. ADASYN: Adaptive synthetic sampling approach for imbalanced learning.
2008 IEEE International Joint Conference on Neural Networks.
30. Kazemitabar J, Amini A, Bloniarz A, Talwalkar A. Variable Importance using Decision Trees. 31st Con-
ference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
Injury forecasting in soccer
PLOS ONE | July 25, 2018 15 / 15

Supplementary resources (21)

... The 30 studies collectively included 9,142 participants, with 6,549 males (median: 96; range 25-2322), and 860 females (median: 105; range 26-187). Ten (33%) studies included soccer athletes [34,39,44,45,47,48,[50][51][52]62], five (17%) athletes from undefined sports [33,42,43,49,55], five (17%) basketball athletes [41,44,54,56,58], two (7%) Australian Rules Football athletes [36,53], two (7%) volleyball athletes [44,58], and 12 (40%) studies included athletes from baseball, floorball, football, gymnastics, handball, korfball, ice hockey, rugby, or swimming [35,37,38,41,45,46,49,[57][58][59][60][61]. Fourteen (47%) studies included professional or elite athletes [33, 34, 36, 38, 39, 45, 46, 52-54, 56-59, 62], eight (27%) college athletes [35,44,47,49,50,55,57,60,61], five (17%) youth athletes [37,40,41,48,51], two (7%) recreational adult athletes [43,54], and two (7%) high school athletes [42,55]. ...
... Twenty-one (70%) studies used a prospective cohort design [33, 34, 37, 38, 40-42, 44, 45, 47, 48, 50-54, 57, 58, 60, 61], five (17%) studies were retrospective cohort studies [35,36,39,46,56], and four (13%) studies used a case control study design [43,49,55,59]. Eighteen (60%) studies developed models using regression [33, 35-40, 43, 44, 47, 49, 50, 54-56, 58, 60, 61], four (13%) studies used machine learning methods [34,45,51,52], seven (23%) studies investigated both regression and Content courtesy of Springer Nature, terms of use apply. Rights reserved. ...
... Ten studies (33%) dichotomized continuous predictors prior to developing the model [36, 42-45, 47-49, 60, 61], and an additional three studies (10%) were unclear about whether continuous predictors were dichotomized prior to developing the model [33,34,50]. Twelve studies (40%) selected predictors through univariable analyses (i.e., based on their unadjusted association with the outcome) before model development [36,37,43,48,49,54,56,57,59,61], seven studies (23%) were unclear on how predictors were selected before or during model development [38,40,45,46,50,58], one study (3%) used recursive cross-validation methods (sic) [52], four (13%) included all predictors [44,47,51,53], three (10%) used backwards selection [33,35,39], three (10%) used least absolute shrinkage and selection operator (LASSO) [41,42,62], and two (7%) included both univariable (before model development) and backward step-wise selection (during model development) [55,60]. Three machine learning studies (10%) [45,52,53] over-sampled to reportedly control for class imbalance, and two (7%) [45,62] under-sampled to control for class imbalance. ...
Full-text available
Background An increasing number of musculoskeletal injury prediction models are being developed and implemented in sports medicine. Prediction model quality needs to be evaluated so clinicians can be informed of their potential usefulness. Objective To evaluate the methodological conduct and completeness of reporting of musculoskeletal injury prediction models in sport. Methods A systematic review was performed from inception to June 2021. Studies were included if they: (1) predicted sport injury; (2) used regression, machine learning, or deep learning models; (3) were written in English; (4) were peer reviewed. Results Thirty studies (204 models) were included; 60% of studies utilized only regression methods, 13% only machine learning, and 27% both regression and machine learning approaches. All studies developed a prediction model and no studies externally validated a prediction model. Two percent of models (7% of studies) were low risk of bias and 98% of models (93% of studies) were high or unclear risk of bias. Three studies (10%) performed an a priori sample size calculation; 14 (47%) performed internal validation. Nineteen studies (63%) reported discrimination and two (7%) reported calibration. Four studies (13%) reported model equations for statistical predictions and no machine learning studies reported code or hyperparameters. Conclusion Existing sport musculoskeletal injury prediction models were poorly developed and have a high risk of bias. No models could be recommended for use in practice. The majority of models were developed with small sample sizes, had inadequate assessment of model performance, and were poorly reported. To create clinically useful sports musculoskeletal injury prediction models, considerable improvements in methodology and reporting are urgently required.
... Rossi and colleagues [33] examined non-contact injuries. The authors collected 954 data recordings (each data record held information about players' daily training load) from 80 training sessions, using 18 training load variables. ...
... In sum, Rossi et al. [33], López-Valenciano et al. [36], Ayala et al. [37], Oliver et al. [41] and Vallance et al. [42] implemented various white-box, tree-based machine learning algorithms in their models. Naglah et al. [35], Vallance et al. [42], Venturelli et al. [43], and Kampakis [44] applied black-box machine learning algorithms (support vector machine, artificial neural networks, Cox regression). ...
... The majority of articles used techniques such as SMOTE, random undersampling, and random oversampling to counter data imbalance. Further, all articles used crossvalidation, although note that Rossi et al. [33] used a prequential evaluation approach (common in stream data classification-also noted below), whereby their model was repeatedly tested on incoming (in their case, weekly) small data batches, which were then added to the training data-this approach of evaluation and updating with new data may more closely mirror the real-world experience of practitioners using all available data to predict injuries in real time. Table 1 gives basic descriptive information about each study, including players' ages, types of injury, and time frame-each of these factors could be important in determining which features are selected during machine learning as the most prominent injury predictors. ...
Full-text available
Attempts to better understand the relationship between training and competition load and injury in football are essential for helping to understand adaptation to training programmes, assessing fatigue and recovery, and minimising the risk of injury and illness. To this end, technological advancements have enabled the collection of multiple points of data for use in analysis and injury prediction. The full breadth of available data has, however, only recently begun to be explored using suitable statistical methods. Advances in automatic and interactive data analysis with the help of machine learning are now being used to better establish the intricacies of the player load and injury relationship. In this article, we examine this recent research, describing the analyses and algorithms used, reporting the key findings, and comparing model fit. To date, the vast array of variables used in analysis as proxy indicators of player load, alongside differences in approach to key aspects of data treatment—such as response to data imbalance, model fitting, and a lack of multi-season data—limit a systematic evaluation of findings and the drawing of a unified conclusion. If, however, the limitations of current studies can be addressed, machine learning has much to offer the field and could in future provide solutions to the training load and injury paradox through enhanced and systematic analysis of athlete data.
... Moreover, computational intelligence methods such as Fuzzy systems in [4,16] are applied for the sports training session. The evolution of machine learning approaches, such as Support Vector Machine (SVM), K-nearest Neighbour (K-NN), Random Forest (RF), Logistic Regression (LR), Decision Tree (DT), etc. opens a path for planning, realization, control, and evaluation in sports training [2,11,12,21]. However, Artificial Neural Network (ANN) in [12,3,20] are also performed to evaluate the human actions in sports sessions. ...
... The evolution of machine learning approaches, such as Support Vector Machine (SVM), K-nearest Neighbour (K-NN), Random Forest, Logistic Regression (LR), Decision Tree (DT), etc. open a path planning, realization, control, and evaluation in sports training [2,11,12,21]. Acikmese et al. [2] proposed a system for recognizing basketball training tactics automatically using motion sensors. It was helpful for sport gesture recognition and achieved 99.5% accuracy for SVM. ...
... However, no model performed accurately enough to express. Rossi et al. [21] applied a decision tree for estimating injuries by using training load and physical ability. It also generated a set of rules for professional soccer players. ...
Full-text available
Recognizing tennis actions during a practice session can be widely used as the coaching assistant. Accurate classification of actions is tricky, with a massive similarity among different actions. Hybrid deep neural networks composed of Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) such as Long Short-term Memory (LSTM) are widely employed while dealing with spatial and temporal features. We leverage transfer learning as the spatial feature extractor, allowing weights extracted by training over a massive dataset. Transfer learning from three pre–trained models such as InceptionResNetV2, ResNet152V2, and Xception are utilized in this work. Two approaches are exercised to determine the best one: (a) developing a single hybrid CNN–LSTM model and (b) passing extracted CNN features to a single LSTM-based model. Experimental results prove the effectiveness of the latter approach for recognizing indoor tennis actions. The publicly available THETIS dataset is employed for all the evaluations and approaches, which contains 12 different tennis actions performed in indoor practice sessions. The Xception–based model outperforms other models by attaining 75% accuracy
... Actually, the use of multidimensional models is fundamental to injury prediction, because sports injuries are a consequence of complex interactions of multiple risk factors [5,6]. Moreover, in addition to prediction task, predictive modeling should provide injury risk factors to implement interventions to minimize the level of risk maximizing the training effect [5][6][7]. ...
... Differently, performing tests periodically (monitoring approach) showed a moderate-high injury prediction accuracy [3]. Monitoring training workloads along the season permits having an overview of the players' fitness status and consequently of their injury risk [7,[13][14][15][16][17][18]. As a matter of fact, it was demonstrated that the greater the training exposure is, the greater the injury risk is [13,19,20]. ...
... These values were used as input in the machine learning model. This metric was found to be an important index for injury prediction in the previous paper [7,22,27,28]. ...
Full-text available
Purpose By analyzing external workloads with machine learning models (ML), it is now possible to predict injuries, but with a moderate accuracy. The increment of the prediction ability is nowadays mandatory to reduce the high number of false positives. The aim of this study was to investigate if players’ blood sample profiles could increase the predictive ability of the models trained only on external training workloads. Method Eighteen elite soccer players competing in Italian league (Serie B) during the seasons 2017/2018 and 2018/2019 took part in this study. Players’ blood samples parameters (i.e., Hematocrit, Hemoglobin, number of red blood cells, ferritin, and sideremia) were recorded through the two soccer seasons to group them into two main groups using a non-supervised ML algorithm (k-means). Additionally to external workloads data recorded every training or match day using a GPS device (K-GPS 10 Hz, K-Sport International, Italy), this grouping was used as a predictor for injury risk. The goodness of ML models trained were tested to assess the influence of blood sample profile to injury prediction. Results Hematocrit, Hemoglobin, number of red blood cells, testosterone, and ferritin were the most important features that allowed to profile players and to analyze the response to external workloads for each type of player profile. Players’ blood samples’ characteristics permitted to personalize the decision-making rules of the ML models based on external workloads reaching an accuracy of 63%. This approach increased the injury prediction ability of about 15% compared to models that take into consideration only training workloads’ features. The influence of each external workload varied in accordance with the players’ blood sample characteristics and the physiological demands of a specific period of the season. Conclusion Field experts should hence not only monitor the external workloads to assess the status of the players, but additional information derived from individuals’ characteristics permits to have a more complete overview of the players well-being. In this way, coaches could better personalize the training program maximizing the training effect and minimizing the injury risk.
... Predicting a future overuse injury enables the performance of an intervention on modifiable factors in a timely manner to help avoid the actual development of an overuse injury. Hence, overuse injury prediction has attracted widespread interest because of its potential to help prevent an injury from occurring [15]. ...
... In recent years, multiple machine-learning-based injury prediction models have been proposed to tackle the challenging injury prediction task [5,15,21,[24][25][26][27][28][29][30][31]. In most machinelearning-based injury prediction studies, the input data are collected by measuring a set of tests performed by the participants or monitoring the participants on regular moments. ...
... In most machinelearning-based injury prediction studies, the input data are collected by measuring a set of tests performed by the participants or monitoring the participants on regular moments. Furthermore, the current literature on machine-learning injury prediction models largely focuses on elite athletes that excel in one particular sport, e.g., running [24], soccer [15,31], or football [25][26][27]. As a result, the machine learning models and findings from these studies might not be transferable to athletes practising different sports. ...
Full-text available
Even though practicing sports has great health benefits, it also entails a risk of developing overuse injuries, which can elicit a negative impact on physical, mental, and financial health. Being able to predict the risk of an overuse injury arising is of widespread interest because this may play a vital role in preventing its occurrence. In this paper, we present a machine learning model trained to predict the occurrence of a lower-limb overuse injury (LLOI). This model was trained and evaluated using data from a three-dimensional accelerometer on the lower back, collected during a Cooper test performed by 161 first-year undergraduate students of a movement science program. In this study, gender-specific models performed better than mixed-gender models. The estimated area under the receiving operating characteristic curve of the best-performing male- and female-specific models, trained according to the presented approach, was, respectively, 0.615 and 0.645. In addition, the best-performing models were achieved by combining statistical and sports-specific features. Overall, the results demonstrated that a machine learning injury prediction model is a promising, yet challenging approach.
... If not the first study on the topic, one of the first was that by Rossi et al. [17]. The authors collected GPS data that described 26 professional male players' workload over one season and constructed injury prediction models. ...
... The best machine learning algorithm could detect around 80% of the injuries with 50% precision and an AUC of 0.76. As the authors argued, the algorithm was "far better than the baseline and state-of-the-art injury risk estimation techniques" [17]. It should be mentioned at this point that risk estimation is not equal to risk prediction. ...
Full-text available
This narrative review paper aimed to discuss the literature on machine learning applications in soccer with an emphasis on injury risk assessment. A secondary aim was to provide practical tips for the health and performance staff in soccer clubs on how machine learning can provide a competitive advantage. Performance analysis is the area with the majority of research so far. Other domains of soccer science and medicine with machine learning use are injury risk assessment, players’ workload and wellness monitoring, movement analysis, players’ career trajectory, club performance, and match attendance. Regarding injuries, which is a hot topic, machine learning does not seem to have a high predictive ability at the moment (models specificity ranged from 74.2%-97.7%. sensitivity from 15.2%-55.6% with area under the curve of 0.66–0.83). It seems, though, that machine learning can help to identify the early signs of elevated risk for a musculoskeletal injury. Future research should account for musculoskeletal injuries’ dynamic nature for machine learning to provide more meaningful results for practitioners in soccer.
... For this reason, the literature about data mining and machine learning approaches is quickly growing since the last decade. These approaches could help to have a complete overview of players' wellness, permitting the development of mathematical models able to provide accurate prediction and consequently useful insights about injury risks (Rossi et al., 2018;Ayala et al., 2019;Pappalardo et al., 2019;Seow et al., 2020;Vallance et al., 2020;Van Eetvelde et al., 2021;Rossi et al., 2022) and internal training load (Rossi et al., 2016;Rossi et al., 2017;Rossi et al., 2019). ...
Full-text available
Training for success has increasingly become a balance between maintaining high performance standards and avoiding the negative consequences of accumulated fatigue. The aim of this study is to develop a big data analytics framework to predict players' wellness according to the external and internal workloads performed in previous days. Such a framework is useful for coaches and staff to simulate the players' response to scheduled training in order to adapt the training stimulus to the players' fatigue response. 17 players competing in the Italian championship (Serie A) were recruited for this study. Players' Global Position System (GPS) data was recorded during each training and match. Moreover, every morning each player has filled in a questionnaire about their perceived wellness (WI) that consists of a 7-point Likert scale for 4 items (fatigue, sleep, stress, and muscle soreness). Finally, the rate of perceived exertion (RPE) was used to assess the effort performed by the players after each training or match. The main findings of this study are that it is possible to accurately estimate players' WI considering their workload history as input. The machine learning framework proposed in this study is useful for sports scientists, athletic trainers, and coaches to maximise the periodization of the training based on the physiological requests of a specific period of the season.
... Widely used for athlete monitoring purposes [1][2][3][4][5][6][7][8][9][10][11] , GNSS permit discriminating the physical demand at exercise through objective mechanical parameters, computed from GNSS and Inertial Measurement Units (IMU) signals 12 . Data collected from these wearable devices provide useful insights for understanding a player's activity and its relationship with performance outcomes or injury occurrences during practice 6,[13][14][15][16] . ...
Full-text available
This study aims to predict individual Acceleration-Velocity profiles (A-V) from Global Navigation Satellite System (GNSS) measurements in real-world situations. Data were collected from professional players in the Superleague division during a 1.5 season period (2019-2021). A baseline modeling performance was provided by time-series forecasting methods and compared with two multivariate modeling approaches using ridge regularisation and long short term memory neural networks. The multivariate models considered commercial features and new features extracted from GNSS raw data as predictor variables. A control condition in which profiles were predicted from predictors of the same session outlined the predictability of A-V profiles. Multivariate models were fitted either per player or over the group of players. Predictor variables were pooled according to the mean or an exponential weighting function. As expected, the control condition provided lower error rates than other models on average (p = 0.001). Reference and multivariate models did not show significant differences in error rates (p = 0.124), regardless of the nature of predictors (commercial features or extracted from signal processing methods) or the pooling method used. In addition, models built over a larger population did not provide significantly more accurate predictions. In conclusion, GNSS features seemed to be of limited relevance for predicting individual A-V profiles. However, new signal processing features open up new perspectives in athletic performance or injury occurrence modeling, mainly if higher sampling rate tracking systems are considered.
... For example, when Germany won the football World Cup in 2014, the national team was supported by the German software company SAP to optimize performance and competitor analysis (SAP, 2014). Further digital technologies have since been used to track the injuries of athletes and at times even predict their probability (Bittencourt et al., 2016;Rossi et al., 2018). The second benefit of digital technologies has been their improvement of the organization and management of sports clubs and the stakeholders involved, such as athletes, managers, investors, or fans. ...
Conference Paper
Blockchain innovations such as digital fan tokens and non-fungible tokens (NFTs) have garnered notable attention in the sports industry, yet the wider industry is struggling to keep up with the pace of digitalization. To harness the potential of blockchain technology, sports management practitioners and information systems (IS) researchers need to gain a much better understanding. Hence, the purpose of this study is to advance the theoretical understanding of blockchain in the sports sector. Thereby, we identify, consolidate, and classify blockchain use cases in the domain through a thorough review of the literature published to date. In addition, we (1) provide an overview and classification of blockchain use cases, (2) identify various opportunities for internal and external stakeholders to benefit from blockchain technology, and (3) derive a theoretical concept for blockchain technology regarding its properties, applications, and stakeholders. Furthermore, we (4) propose beneficial directions for future research in this emerging field.
Full-text available
The availability of massive data about sports activities offers nowadays the opportunity to quantify the relation between performance and success. In this study, we analyze more than 6,000 games and 10 million events in six European leagues and investigate this relation in soccer competitions. We discover that a team's position in a competition's final ranking is significantly related to its typical performance, as described by a set of technical features extracted from the soccer data. Moreover we find that, while victory and defeats can be explained by the team's performance during a game, it is difficult to detect draws by using a machine learning approach. We then simulate the outcomes of an entire season of each league only relying on technical data, i.e. excluding the goals scored, exploiting a machine learning model trained on data from past seasons. The simulation produces a team ranking (the PC ranking) which is close to the actual ranking, suggesting that a complex systems' view on soccer has the potential of revealing hidden patterns regarding the relation between performance and success.
Full-text available
Automatic and interactive data analysis is instrumental in making use of increasing amounts of complex data. Owing to novel sensor modalities, analysis of data generated in professional team sport leagues such as soccer, baseball, and basketball has recently become of concern, with potentially high commercial and research interest. The analysis of team ball games can serve many goals, e.g., in coaching to understand effects of strategies and tactics, or to derive insights improving performance. Also, it is often decisive to trainers and analysts to understand why a certain movement of a player or groups of players happened, and what the respective influencing factors are. We consider team sport as group movement including collaboration and competition of individuals following specific rule sets. Analyzing team sports is a challenging problem as it involves joint understanding of heterogeneous data perspectives, including high-dimensional, video, and movement data, as well as considering team behavior and rules (constraints) given in the particular team sport. We identify important components of team sport data, exemplified by the soccer case, and explain how to analyze team sport data in general. We identify challenges arising when facing these data sets and we propose a multi-facet view and analysis including pattern detection, context-aware analysis, and visual explanation. We also present applicable methods and technologies covering the heterogeneous aspects in team sport data.
Full-text available
Objective: To determine if any differences exist between the rolling averages and exponentially weighted moving averages (EWMA) models of acute:chronic workload ratio (ACWR) calculation and subsequent injury risk. Methods: A cohort of 59 elite Australian football players from 1 club participated in this 2-year study. Global positioning system (GPS) technology was used to quantify external workloads of players, and non-contact 'time-loss' injuries were recorded. The ACWR were calculated for a range of variables using 2 models: (1) rolling averages, and (2) EWMA. Logistic regression models were used to assess both the likelihood of sustaining an injury and the difference in injury likelihood between models. Results: There were significant differences in the ACWR values between models for moderate (ACWR 1.0-1.49; p=0.021), high (ACWR 1.50-1.99; p=0.012) and very high (ACWR >2.0; p=0.001) ACWR ranges. Although both models demonstrated significant (p<0.05) associations between a very high ACWR (ie, >2.0) and an increase in injury risk for total distance ((relative risk, RR)=6.52-21.28) and high-speed distance (RR=5.87-13.43), the EWMA model was more sensitive for detecting this increased risk. The variance (R(2)) in injury explained by each ACWR model was significantly (p<0.05) greater using the EWMA model. Conclusions: These findings demonstrate that large spikes in workload are associated with an increased injury risk using both models, although the EWMA model is more sensitive to detect increases in injury risk with higher ACWR.
Conference Paper
Full-text available
Football training periodization is widely acknowledged as crucial to obtain the best performance throughout matches and to reduce the risk of injuries. Thus, the aim of this study is to detect the in-season short-term training cycles in an Italian elite football team. 80 trainings of 26 elite football players were monitored during 23 in-season weeks by a global position system (GPS). Machine learning process and autocorrelation analyses were performed in order to detect pattern inside in-season football trainings. Extra tree random forest classifier (ETRFC) was used to create a supervised machine learning process able to describe the football trainings cycle. This analytical model allows us to produce reliable decisions and results learning from historical relationships and trends in the data. In addition, the autocorrelation analysis allows us to detect similarity between observations between the data. Based on these analysis, it was found that the in-season football trainings are characterized by a series of short-term cycles. This kind of periodization follows a sinusoidal model because the short-term cycle detected in the in-season trainings is composed of two parts with different training loads. In particular, in the days long before the match football players perform higher training loads than in the close ones. To enhance performance and reduce risk of injuries, it would be essential to provide correct stimuli in each short-term cycle per day. Thus, developing a valid method able to define the correct training loads in each training day may be central for coaches and athletic trainers to periodize correctly the football trainings.
Full-text available
Injuries are considered an inevitable by-product of participation in elite sport (Hurley et al., 2007), and over the past 30 years, psychologists have proposed a number of models to explain athletes' psychological reactions to their injuries. These models have included the grief model (Kübler-Ross, 1969; Rotella, 1988) and the cognitive appraisal models (Brewer, 1998; Wiese-Bjornstal et al., 1998). When examining the value of these models, researchers have identified a number of factors mediating athletes' responses to their injuries. Factors such as social support (Clement and Shannon, 2011) and perceived consequences of injuries for the athletes (Hurley et al., 2007) have been popular lines of investigation. A less popular avenue of research has been the impact of tournament-ending injuries on the mental states of remaining squad members called upon to perform without their injured colleagues. Given the lack of research in this area, it could be fruitful to discuss the potential impact of player attrition rates on their teams' mental states, and subsequent performances, using the Rugby World Cup 2015 (RWC 2015) as a case study. This case study is considered appropriate because many countries suffered serious injuries to some of their best players just before, or during, this RWC (Lewis, 2015). For example, the Irish team lost one of its most experienced players, Paul O'Connell, due to injury early in the tournament, as did South Africa, who was forced to play without captain, Jean De Villiers, from game three of the tournament. Similarly, the New Zealand "All Blacks" lost their experienced player, Tony Woodcock, due to injury before their quarter final match. Despite the loss of Woodcock, the New Zealand team was successful in defending its RWC title from 2011. Indeed in 2011, New Zealand played without four of its out-half players. Is it possible the New Zealand team did something different, compared to the other nations, to prepare successfully for both the RWC 2011 and 2015 tournaments? This paper will explore possible reasons why some of the other nations failed to perform to their potential at the RWC 2015. This paper is organized as follows: First, the mental challenges team players may have faced following the loss of influential players through injury will be discussed. The potential for emotional contagion to have permeated throughout such teams, and its possible impact on those teams' subsequent performances will dominate this discussion (Moll et al., 2010). Second, some possible areas for future research will be suggested, focusing on the role support staff (including coaches, medical staff, and sport psychologists) working with such teams might play in helping their teams to thrive in such circumstances in the future.
Full-text available
Background: There is dogma that higher training load causes higher injury rates. However, there is also evidence that training has a protective effect against injury. For example, team sport athletes who performed more than 18 weeks of training before sustaining their initial injuries were at reduced risk of sustaining a subsequent injury, while high chronic workloads have been shown to decrease the risk of injury. Second, across a wide range of sports, well-developed physical qualities are associated with a reduced risk of injury. Clearly, for athletes to develop the physical capacities required to provide a protective effect against injury, they must be prepared to train hard. Finally, there is also evidence that under-training may increase injury risk. Collectively, these results emphasise that reductions in workloads may not always be the best approach to protect against injury. Main thesis: This paper describes the 'Training-Injury Prevention Paradox' model; a phenomenon whereby athletes accustomed to high training loads have fewer injuries than athletes training at lower workloads. The Model is based on evidence that non-contact injuries are not caused by training per se, but more likely by an inappropriate training programme. Excessive and rapid increases in training loads are likely responsible for a large proportion of non-contact, soft-tissue injuries. If training load is an important determinant of injury, it must be accurately measured up to twice daily and over periods of weeks and months (a season). This paper outlines ways of monitoring training load ('internal' and 'external' loads) and suggests capturing both recent ('acute') training loads and more medium-term ('chronic') training loads to best capture the player's training burden. I describe the critical variable-acute:chronic workload ratio)-as a best practice predictor of training-related injuries. This provides the foundation for interventions to reduce players risk, and thus, time-loss injuries. Summary: The appropriately graded prescription of high training loads should improve players' fitness, which in turn may protect against injury, ultimately leading to (1) greater physical outputs and resilience in competition, and (2) a greater proportion of the squad available for selection each week.
Conference Paper
Full-text available
Sports analytics in general, and football (soccer in USA) analytics in particular, have evolved in recent years in an amazing way, thanks to automated or semi-automated sensing technologies that provide high-fidelity data streams extracted from every game. In this paper we propose a data-driven approach and show that there is a large potential to boost the understanding of football team performance. From observational data of football games we extract a set of pass-based performance indicators and summarize them in the H indicator. We observe a strong correlation among the proposed indicator and the success of a team, and therefore perform a simulation on the four major European championships (78 teams, almost 1500 games). The outcome of each game in the championship was replaced by a synthetic outcome (win, loss or draw) based on the performance indicators computed for each team. We found that the final rankings in the simulated championships are very close to the actual rankings in the real championships, and show that teams with high ranking error show extreme values of a defense/attack efficiency measure, the Pezzali score. Our results are surprising given the simplicity of the proposed indicators, suggesting that a complex systems' view on football data has the potential of revealing hidden patterns and behavior of superior quality.
Team-based invasion sports such as football, basketball, and hockey are similar in the sense that the players are able to move freely around the playing area and that player and team performance cannot be fully analysed without considering the movements and interactions of all players as a group. State-of-the-art object tracking systems now produce spatio-temporal traces of player trajectories with high definition and high frequency, and this, in turn, has facilitated a variety of research efforts, across many disciplines, to extract insight from the trajectories. We survey recent research efforts that use spatio-temporal data from team sports as input and involve non-trivial computation. This article categorises the research efforts in a coherent framework and identifies a number of open research questions.
Conference Paper
The goal of this thesis is to investigate the potential of predictive modelling for football injuries. This work was conducted in close collaboration with Tottenham Hotspurs FC (THFC), the PGA European tour and the participation of Wolverhampton Wanderers (WW). Three investigations were conducted: 1. Predicting the recovery time of football injuries using the UEFA injury recordings: The UEFA recordings is a common standard for recording injuries in professional football. For this investigation, three datasets of UEFA injury recordings were available: one from THFC, one from WW and one that was constructed by merging both. Poisson, negative binomial and ordinal regression were used to model the recovery time after an injury and assess the significance of various injury-related covariates. Then, different machine learning algorithms (support vector machines, Gaussian processes, neural networks, random forests, naïve Bayes and k-nearest neighbours) were used in order to build a predictive model. The performance of the machine learning models is then improved by using feature selection conducted through correlation-based subset feature selection and random forests. 2. Predicting injuries in professional football using exposure records: The relationship between exposure (in training hours and match hours) in professional football athletes and injury incidence was studied. A common problem in football is understanding how the training schedule of an athlete can affect the chance of him getting injured. The task was to predict the number of days a player can train before he gets injured. The dataset consisted of the exposure records of professional footballers in Tottenham Hotspur Football Club from the season 2012-2013. The problem was approached by a Gaussian process model equipped with a dynamic time warping kernel that allowed the calculation of the similarity of exposure records of different lengths. 3. Predicting intrinsic injury incidence using in-training GPS measurements: A significant percentage of football injuries can be attributed to overtraining and fatigue. GPS data collected during training sessions might provide indicators of fatigue, or might be used to detect very intense training sessions which can lead to overtraining. This research used GPS data gathered during training sessions of the first team of THFC, in order to predict whether an injury would take place during a week. The data consisted of 69 variables in total. Two different binary classification approaches were followed and a variety of algorithms were applied (supervised principal component analysis, random forests, naïve Bayes, support vector machines, Gaussian process, neural networks, ridge logistic regression and k-nearest neighbours). Supervised principal component analysis shows the best results, while it also allows the extraction of components that reduce the total number of variables to 3 or 4 components which correlate with injury incidence. The first investigation contributes the following to the field: • It provides models based on the UEFA injury recordings, a standard used by many clubs, which makes it easier to replicate and apply the results. • It investigates which variables seem to be more highly related to the prediction of recovery after an injury. • It provides a comparison of models for predicting the time to return to play after injury. The second investigation contributes the following to the field: • It provides a model that can be used to predict the time when the first injury of the season will take place. • It provides a kernel that can be utilized by a Gaussian process in order to measure the similarity of training and match schedules, even if the time series involved are of different lengths. The third investigation contributes the following to the field: • It provides a model to predict injury on a given week based on GPS data gathered from training sessions. • It provides components, extracted through supervised principal component analysis, that correlate with injury incidence and can be used to summarize the large number of GPS variables in a parsimonious way.