Chaos, Solitons and Fractals 167 (2023) 113079

Contents lists available at ScienceDirect

Chaos, Solitons and Fractals

journal homepage: www.elsevier.com/locate/chaos

Predicting injury risk using machine learning in male youth soccer players
Francisco Javier Robles-Palazón a, José M. Puerta-Callejón b, José A. Gámez b,
Mark De Ste Croix c, Antonio Cejudo a, Fernando Santonja d, Pilar Sainz de Baranda a,
Francisco Ayala a, c, *
a
Department of Physical Activity and Sport, Faculty of Sport Sciences, Campus of Excellence Mare Nostrum, University of Murcia, Murcia, Spain
b
Department of Computer Systems, University of Castilla-La Mancha, Albacete, Spain
c
School of Sport and Exercise, University of Gloucestershire, Gloucester, United Kingdom
d
Virgin of the Arrixaca University Hospital, Faculty of Medicine, Campus of Excellence Mare Nostrum, University of Murcia, Murcia, Spain

A R T I C L E I N F O A B S T R A C T

Keywords: The aim of this study was twofold: a) to build models using machine learning techniques on data from an
Screen extensive screening battery to prospectively predict lower extremity soft tissue (LE-ST) injuries in non-elite male
Associated football youth soccer players, and b) to compare models' performance scores (i.e., predictive accuracy) to select the best
Prediction model
fit. A sample of 260 male youth soccer players from the academies of five different Spanish non-professional
Adolescent
Prevention
clubs completed the follow-up. Players were engaged in a pre-season assessment that covered several personal
characteristics (e.g., anthropometric measures), psychological constructs (e.g., trait-anxiety), and physical fitness
and neuromuscular measures (e.g., range of motion [ROM], landing kinematics). Afterwards, all LE-ST injuries
were monitored over one competitive season. The predictive ability (i.e., area under the receiver operating
characteristic curve [AUC] and F-score) of several screening models was analysed and compared to select the one
with the highest scores. A total of 45 LE-ST injuries were recorded over the season. The best fit screening model
developed (AUC = 0.700, F-score = 0.380) allowed to successfully identify one in two (True Positive rate = 53.7
%) and three in four (True Negative rate = 73.9 %) players at high or low risk of suffering a LE-ST injury
throughout the in-season phase, respectively, using a subset of six field-based measures (knee medial displace­
ment in the drop jump, asymmetry in the peak vertical ground reaction force during landing, body mass index,
asymmetry in the frontal plane projection angle assessed through the tuck jump, asymmetry in the passive hip
internal rotation ROM, and ankle dorsiflexion with the knee extended ROM). Given that these measures require
little equipment to be recorded and can be employed quickly (approximately 5–10 min) and easily by trained
staff in a single player, the model developed might be included in the injury management strategy for youth
soccer.

1. Introduction cruciate ligament [ACL] of the knee, anterior inferior tibiofibular liga­
ment of the ankle) are the most commonly diagnosed types of injury in
Despite the numerous health-related benefits, the participation in a youth soccer players [1,6]. These lower extremity soft tissue (LE-ST)
very physically demanding team sport such as soccer (i.e., associated injuries frequently result in players missing sport participation for an
football) results in a notable increase in injury risk [1]. Epidemiological extensive period of time [6]. In addition, young players who sustain LE-
studies have reported that the frequency and severity of injuries among ST injuries during soccer participation may experience important re­
youth soccer players accelerate and peak during adolescence [2,3], sidual symptoms that can have major negative consequences in their
when periods of rapid and non-uniform growth in skeletal structures are long-term athlete development and limit their ability to engage in ex­
experienced, leading to alterations in both physical performance and ercise and athlete activities later in life [7]. Consequently, soccer-related
motor control/function [4,5]. Thigh muscle/tendon strains (hamstring LE-ST injuries can counter the beneficial health related effects of sport
and quadriceps) and knee and ankle ligament sprains and tears (anterior participation at a young age if a child or adolescent is unable to continue

* Corresponding author at: Department of Physical Activity and Sport, University of Murcia, Faculty of Sport Sciences, C/Argentina s/n, 30720 Santiago de la
Ribera-San Javier, Murcia, Spain.
E-mail address: [email protected] (F. Ayala).

https://doi.org/10.1016/j.chaos.2022.113079
Received 4 October 2022; Received in revised form 23 December 2022; Accepted 26 December 2022
Available online 14 January 2023
0960-0779/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
F.J. Robles-Palazón et al. Chaos, Solitons and Fractals: the interdisciplinary journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena 167 (2023) 113079

participating because of the effects of injury [8]. specificity, respectively. Furthermore, these studies [14,21] have also
Most of the LE-ST injuries documented in youth soccer have shown a identified interactions of asymmetry, knee valgus angle and body size as
non-contact mechanism [1] and hence, they might be considered as contributing factors to an injurious profile in elite youth soccer players.
preventable [9]. Thus, the implementation of multicomponent strategies However, it should be acknowledged that a limitation of any pre­
aimed at mitigating the risk of injury in such cohorts is a big challenge diction model developed through the use of classification algorithms is
that coaches and physical trainers need to consider. It has been sug­ that its generalisation to individuals with different characteristics (e.g.,
gested that for an injury prevention measure to be highly effective, its sport background, exposure to causal factors of injury, physical perfor­
design must be targeted on each player's individual needs [10]. There­ mance) to those who were employed in its building and validation
fore, the use of a valid field-based screening method that allows coaches process may be sub-optimal. In this sense, the well-documented differ­
and physical trainers to profile injury risk and identify those factors that ences in several physical performance measurements [22] between elite
impact most on the likelihood of sustaining a LE-ST injury in each of and non-elite (i.e., sub-elite or amateur) youth soccer players may lead
their youth soccer players may be a valuable tool to help design tailored to a dramatic reduction in the ability of these two currently available
preventive measures. screening models to predict LE-ST injuries in the latter cohort. Given
There is a general agreement that LE-ST injury is a multifactorial that a large proportion of the young participants play for non-
phenomenon in which several factors of different nature (e.g., personal professional clubs, engaged in local and regional leagues, and that the
characteristics, psychological constructs, neuromechanical parameters) injury incidence and severity is still high in this cohort [1], studies
might interact among them in a non-linear fashion (complex relation­ aimed at building injury risk factor models to identify non-elite youth
ships) and have an impact on the likelihood (i.e., risk) of this one ap­ soccer players at high risk of LE-ST injury are urgently warranted.
pears (or not) in an athlete (i.e., soccer players) [11–14]. Likewise, Therefore, the aim of this study was twofold: a) to build models using
epidemiological studies in soccer have documented that the LE-ST injury machine learning techniques on data from an extensive screening bat­
is an imbalanced phenomenon so that in a typical team the number of tery to prospectively predict LE-ST injuries in non-elite male youth
players who sustain a LE-ST every competitive season (minority class) is soccer players, and b) to compare models' performance scores (i.e.,
much lower than the non-injured players (majority class) [15]. Most of predictive accuracy) to select the best fit.
the screening models available currently to make prospective pre­
dictions on new cases of LE-ST injuries in soccer have been built using 2. Materials and methods
traditional statistical techniques (mainly binary logistic regression) that
were not originally conceived to manage complex (non-linear) and This study was carried out following the Transparent Reporting of a
imbalanced phenomena (as the LE-ST is) [16–18]. Furthermore, these multivariable prediction model for Individual Prognosis or Diagnosis
models have been designed using information coming from one in (TRIPOD) guidelines [23]. The TRIPOD checklist is provided in online
isolation or a few factors (no more than six) assessed in a limited sample supplementary file 1.
of soccer players. Consequently, it is not surprising that these traditional
models present inadequate performance scores (i.e., predictive accu­ 2.1. Participants
racy) so that in most of them a clear bias (for many reasons) toward the
majority class (known as the negative class) is shown, and therefore, A total sample of 301 male youth soccer players from the academies
there is a higher misclassification rate for the minority class instances of five different Spanish non-professional soccer clubs were recruited for
(called the positive examples), which represent the most important this study. All players were engaged in regional (non-national) youth
concept [19]. In other words, these models usually report high speci­ soccer leagues of the south-east of Spain. Participants routinely
ficity (also called true negative rate [non-injured players who were well- completed from two (most of the weeks in the U11–12 age group) to
classified]) but very low sensitivity (also called true positive rate three (most of the weeks in U13–14, U15–16, and U17–19 age groups)
[injured players who were well-classified]). Therefore, it has been training sessions (90 min each) per week on non-consecutive days and
argued that the complexity of injury means a broader statistical and played one competitive match (match duration: U11–12 = 60 min,
conceptual approach is needed to make more accurate prospective U13–14 = 70 min, U15–16 = 80 min, U17–19 = 90 min) per week
predictions of new cases of injuries and better understand relationships (usually at the weekend) during the season. Participants were included
between risk factors [14,20]. in this study if they met the following criteria: 1) they were free from
In the last five years, a growing number of studies have used pain, illness and/or injury during the whole data collection phase and 2)
contemporary Machine Learning algorithms (mainly classification [e.g., they were regularly involved in soccer training and competition. Players
Random Forest and ADTree] and regression algorithms [e.g., Naïve who conveyed the presence of orthopaedic problems that did not allow
Bayes and Neural Networks]) which have been specifically designed to them to carry out one or more of the field-based tests, or who were
deal with imbalanced problems where a large number of factors are transferred to a different club and were not available for follow-up
involved and resampling methods (e.g., K-fold cross validation, leave- testing at the end of 9-months were excluded. Coaches, parents and
one-out, bootstrapping) to build screening models to profile athletes' children were informed in both oral and written forms, and parental
injury risk in team sport showing, in most of the cases, promising pre­ consent to participate in the study was obtained together with assent
dictive accuracy [11–13]. Only two recent studies [14,21] have devel­ from participants. Ethical approval was granted by the Ethics and Sci­
oped screening models using field-based tests to predict injuries through entific Committee of the University of Murcia (ID: 1551/2017) in
the use of decision tree based classifiers (XGBoost [21] and bagging accordance with the Declaration of Helsinki.
ensemble method with a J48con decision tree as base classifier [14]) in Finally, a sample of 260 male youth soccer players of four different
youth soccer players. In particular, these two studies have built models age categories (age-based categories [n]: U11–12 [78], U13–14 [69],
to classify youth players into two groups, positives (high risk of injury) U15–16 [50], U17–19 [63]) completed this study (Table 1). Forty-one
and negatives (low risk of injury), based on anthropometric (e.g., age, players were removed from the initial sample of 301 young based on
standing and sitting height, body mass), physical fitness (e.g., sprint and the exclusion criteria (n = 11 players reported a presence of pain and
jump [vertical and horizontal] performance, agility, lower back and orthopaedic problems, n = 14 players did not provide the required
posterior chain flexibility) and neuromuscular (e.g., tuck jump knee signed informed consent before the start of the study, and n = 16 players
valgus angle, unilateral landing peak vertical ground reaction force and were transferred to another club or left their club before the end of the
asymmetry) measures in elite young male players from the youth follow up period).
academies of six English [14] and seven Belgium [21] premier league
soccer clubs, reporting moderate to high levels of sensitivity and

2
F.J. Robles-Palazón et al. Chaos, Solitons and Fractals: the interdisciplinary journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena 167 (2023) 113079

Table 1
Descriptive anthropometric values (mean ± standard deviation) by age group.
Group N Age (years) Body mass (kg) Stature (cm) Leg length (cm) Maturity offset

U11–12 78 11.1 ± 0.5 39.8 ± 7.4 148.1 ± 6.6 72.8 ± 4.2 − 2.4 ± 0.6
U13–14 69 13.3 ± 0.4 51.9 ± 8.6 162.3 ± 7.8 80.8 ± 5.4 − 0.7 ± 0.6
U15–16 50 15.0 ± 0.5 62.6 ± 8.5 173.2 ± 6.3 84.9 ± 3.9 1.1 ± 0.6
U17–19 63 17.3 ± 0.8 68.7 ± 8.4 176.6 ± 7.3 86.2 ± 5.5 2.6 ± 0.7

U: under.

2.2. Study design specific moment when the questionnaire is completed, whereas the trait
items describe the athletes' general anxiety level. For the purposes of this
This study used a prospective cohort design. Particularly, all LE-ST research, only the trait anxiety was analysed.
injuries sustained in training and competition during a period of 9 Mood states were evaluated using the Spanish adapted version for
months following the initial assessment session (in-season phase) were adolescent athletes of the Profile of Mood States (POMS) scale [33]. This
tracked for all players. Participants were required to attend their version comprises seven different psychological factors (tension,
respective club's training facilities during the pre-season phase depression, anger, vigour, fatigue, confusion, and friendliness) in a 33-
(September) of the years 2017 (n = 175 players) and 2018 (n = 85 item scale.
players) to undergo an assessment of several personal characteristics, The Spanish version of the Psychological Characteristics Related to
psychological constructs, and physical fitness/neuromuscular measures. Sport Performance questionnaire (CPRD) was used to measure the
following psychological characteristics: stress control, influence of per­
formance evaluation, motivation, team cohesion and mental skills [34].
2.3. Procedure The questionnaire consists of 55 items graded in a 5-option Likert scale
(from totally disagree to totally agree) (for more information on psy­
The assessment session was split into three different parts. The first chological risk factors recorded, please see online supplementary file 3).
part was designed to get data concerning the participants' personal or
individual characteristics. Secondly, a number of psychological con­ 2.3.3. Physical fitness, neuromuscular capability and biomechanical
structs related to anxiety and mood state were evaluated. Finally, in the measures
third part several physical performance, neuromuscular capability and Players completed a standardised dynamic warm-up, which included
biomechanical measures were assessed through 10 field-based tests. All whole body exercises emphasising dynamic mobilisation and gradually
measures were taken by six trained and experienced testers (one master progressing in intensity [35], before the physical performance, neuro­
and two PhD students and three senior researchers with three, six and muscular capability and biomechanical measures were taken. In
more than ten years of experience, respectively), coordinated by the particular, these measures were concurrently recorded using a rando­
principal investigator (FJR-P) to guarantee standardisation of protocols. mised circuit style approach (due to time constraints) (Fig. 1) from six
All measurements have demonstrated moderate to good reliability jump tests, a linear 30 m sprint test, the ROM-Sport battery, Y-Balance
(intraclass correlation coefficients [ICCs] > 0.80 and standard error of test and Illinois agility test.
measurements expressed as percentage [%SEM] < 10 %) as it has been
described elsewhere [24–29].
2.3.3.1. Jump tests. Four vertical and two horizontal jump tests were
performed and several measures of performance, kinematic and kinetic
2.3.1. Personal or individual characteristics
variables and neuromuscular parameters were extracted from them.
Personal or individual measures (player position [goalkeeper, de­
Three to five attempts of each jump test were performed. For each
fender, midfielder or forward], years of playing soccer, training fre­
variable, the best absolute score recorded in the attempts carried out was
quency, dominant leg [determined by the player's preferred kicking leg],
selected for the subsequent analysis (for more information on measures
self-reported 12 months LE-ST time loss injury history [yes or no], and
obtained from the Jump tests, please see online supplementary file 4).
chronological age) were recorded using an ad hoc questionnaire.
2.3.3.1.1. Vertical jump tests. Tuck jump assessment (TJA). Tuck
Anthropometric measures (body mass, stature [i.e., standing height],
jumps were performed in place for 10 consecutive seconds following the
sitting height, body mass index [BMI], and leg and tibia lengths) and
procedure previously suggested by Myer et al. [36]. Each participant's
maturity status were also measured. Body mass was measured on a
technique was assessed at frontal and sagittal planes. A 2-dimensional
calibrated physician scale (SECA 799, Hamburg, Germany). Standing
video cameras (model: Lumix DMC-FZ200; Panasonic, Japan) were
and sitting height were recorded to the nearest 0.1 cm on a measurement
positioned in both planes at a height of 0.70 m and a distance of 5 m
platform (SECA 799, Hamburg, Germany) with seated height measured
from the landing area to capture the test and grade each player's tech­
using a box. Leg length was calculated as the length measured in cen­
nique retrospectively. Afterwards, frontal plane projection angles
timetres from the anterior superior iliac spine to the most distal portion
(FPPA) at the point of maximum knee flexion were analysed, and the
of the medial tibial malleolus [25]. Tibia length was defined as the
presence of knee valgus was subjectively classified as minor (< 10◦ ),
distance between the lateral knee joint line and the lateral malleolus
moderate (10◦ –20◦ ) or severe (> 20◦ ) following the methodology
[30]. Stage of maturation was calculated in a non-invasive manner using
described by Read et al. [27]. Additionally, hip flexion (HF), knee
a regression equation comprising measures of age, body mass, standing
flexion (KF), and ankle flexion (AF) was assessed at initial contact and
height and sitting height [31]. Using this method, maturity offset
peak maximum flexion in the sagittal plane [29]. All scores were marked
(calculation of years from peak height velocity [PHV]) was determined
by two experienced testers in 2-D landing kinematic assessments.
(for more information on the personal or individual risk factors recor­
Drop vertical jump (DVJ). A double leg drop vertical jump from a box
ded, please see online supplementary file 2).
height of 40 cm and without arm swing was performed on a contact
platform connected to the Ergo tester (Ergo Jump Bosco System, Italia)
2.3.2. Psychological constructs
unit [37]. Both jump height and reactive strength index (RSI = jump
The Spanish version of the State-Trait Anxiety Inventory (STAI)
height/contact time) were considered to assess stretch-shortening cycle
questionnaire was used to measure the current state and trait anxiety of
(SSC) function and hence, recorded. A 2-dimensional landing kinematic
the players [32]. This questionnaire consists of 40 items (20 for state and
analysis following the methodology described for the TJA was also
20 for trait). The state items describe how the athletes feel just at the

3
F.J. Robles-Palazón et al. Chaos, Solitons and Fractals: the interdisciplinary journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena 167 (2023) 113079

Fig. 1. Circuit style approach.

carried out. In addition to the FPPA, the knee medial displacement was measured by means of three pairs of Microgate Witty photocells
(KMD) (expressed as the displacement measure [d2–d1] between the (Microgate, Italy) placed 1.0 m above the ground level. Each sprint was
initial contact [d1] and the maximal peak knee flexion [d2]) [30] the initiated from an individually chosen standing position, 50 cm behind
knee-to-ankle separation ratio (KASR) (defined as the ratio of distance the photocell gate, which started a digital timer. The theoretical
between knees and ankles during peak knee flexion [KASR = knee/ maximal force (F0), velocity (V0), maximal power output (Pmax) and
ankle]) [28] and the knee separation distance (KSD) (expressed as the mechanical effectiveness of ground force application (ratio of force [RF]
difference [d2-d1] between knee separation distance at the initial con­ and decrease in the RF over acceleration [DRF]) during a 30 m-sprint
tact [d1] and the peak knee flexion [d2]) [28] were also used to assess were also analysed. For this purpose, all sprint trials were recorded
knee valgus during DVJ tests. All trials were retrospectively analysed by through an iPad Air (Apple Inc., USA) and retrospectively analysed by a
the same two experienced testers in 2-D landing kinematics assessments. single tester using the MySprint app [26]. The analysis of sprint force-
Countermovement jump (CMJ). A double leg countermovement velocity profile in youth athletes has proven to be reliable in previous
jump without arm swing was performed on a contact platform con­ research [40] (for more information on measures obtained from the
nected to the Ergo tester (Ergo Jump Bosco System, Italia) unit. Jump Sprint, please see online supplementary file 5).
height was recorded for subsequent analyses.
Single leg countermovement jump (SLCMJ). A single leg (dominant 2.3.3.3. ROM-Sport battery. The passive hip extension (PHE), hip
and non-dominant) countermovement jump was also performed on a adduction with hip flexed 90◦ (PHADHF90◦ ), hip flexion with knee flexed
force platform (9286AA, Kistler, Switzerland). Height, peak vertical (PHFKF) and extended (PHFKE), hip abduction with hip neutral (PHABD)
ground reaction force (pVGRF) during take-off and landing, and peak and hip flexed 90◦ (PHABDHF90◦ ), hip external (PHER) and internal
landing force timing (pLFT) were captured at a sampling rate of 1000 (PHIR) rotation, knee flexion (PKF), ankle dorsiflexion with knee flexed
Hz. A threshold of > 10 N to determine contact and < 10 N to determine (ADFKF) and extended (ADFKE) ROM measures of the dominant and non-
flight moments was used, and no filter was applied to the data obtained dominant legs were evaluated according to the methodology suggested
for subsequent analyses [38]. The pVGRF at take-off and landing were by Cejudo et al. [24]. For each joint ROM measure, side-to-side differ­
normalised to body weight (BW), and side-to-side differences for each of ences were also calculated. When a side-to-side difference ≥ 8◦ was
these variables were calculated. Asymmetries in all SLCMJ variables found, players were categorised as showing bilateral asymmetries [41]
were determined when bilateral differences were ≥ 10 %. (for more information on data collected with the ROM-Sport battery,
2.3.3.1.2. Horizontal jump tests. Standing long jump (SLJ). Jump please see online supplementary file 6).
distance in a SLJ was measured to the nearest centimetre from the
starting line to the player's heel with a standard tape measure. Free 2.3.3.4. Y-Balance test. Dynamic postural control was evaluated using
movement of the arms was allowed during the test. the Y-Balance test [25]. The distance obtained in each direction (ante­
Single hop for distance (SHD). Jump performance in a SHD was also rior, posteromedial, and posterolateral) was normalised by dividing by
measured for dominant and non-dominant legs [39]. The jump distance the previously measured leg length to standardise the reach distance
in cm was then normalised and presented as percentage of leg length ([excursion distance/leg length] x 100 = % leg length) [25]. Bilateral
(SHD/leg length*100 = % leg length). Bilateral differences were differences between dominant and non-dominant legs were also calcu­
calculated and asymmetry was considered when differences ≥ 10 %. lated for each distance, and differences ≥ 10 % for anterior, poster­
omedial, and posterolateral directions were considered as asymmetries.
2.3.3.2. Sprint. Time during a 10–20 and 30 m sprint in a straight line Finally, to obtain a global measure of the balance test for each leg, data

4
F.J. Robles-Palazón et al. Chaos, Solitons and Fractals: the interdisciplinary journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena 167 (2023) 113079

from each direction were averaged to calculate a composite score (for characteristic (ROC) curve. Thus, the area under the ROC curve (AUC)
more information on measures obtained from the Y-Balance test, please was employed as a measure of a classifier's performance for evaluating
see online supplementary file 7). which models showed high (0.90–1.00), moderate (0.70–0.90), low
(0.50–0.70) and fail (< 0.50) scores [45]. For the purpose of this study,
2.3.3.5. Illinois agility test. Players' agility was assessed using the Illinois only algorithms with performance scores (AUC) above 0.70 were
agility test, which has been commonly used in measuring agility in considered acceptable. Also, two extra measures from the confusion
soccer [42]. The length of the zone was 10 m, while the width (distance matrix were selected as evaluation criteria: true positive (TP) rate and
between the start and finish points) was 5 m. Four cones were placed in true negative (TN) rate. In imbalanced domains, when the AUC has
the centre of the testing area at a distance of 3.3 m from one another. reached a high score (> 0.70), the classification performance may not be
Four cones were used to mark the start, finish, and two turning points. as good as the AUC value reflects because plenty of “clear” negative
The participants started the test lying face down, with their hands at samples (instances that can be clearly classified into the negative label of
shoulder level. The trial started on the “go” command, and the partici­ the class variable) exist in the dataset. These clear negative samples may
pants began to run as fast as possible. The trial was completed when the increase the AUC score, but a few other “border line” negative samples
players crossed the finish line without having knocked any cones over. remain mixed with the positive samples (i.e., class overlapping and/or
Time was measured using a photocell system (Microgate Witty photo­ small disjuncts), which are difficult to distinguish and classify by the
cells; Microgate, Italy). algorithms. These few remaining border line negative samples may
decrease performance (when some of them are wrongly classified [i.e.,
2.3.4. Injury surveillance false positive]), including precision and recall, while very slightly
The procedures for data collection and reporting injury occurrences influencing the AUC score. In consequence, Zou et al. [46] recommend
described in the International Consensus Statement were followed in the using the F-score together with the AUC as a classification measurement
current research [43]. For the purpose of this research, an injury was for imbalanced problems.
defined as any non-contact, soft tissue (muscle, tendon, and/or liga­ Similar to previously published studies aimed at building prediction
ment) injury sustained by a player during a training session or compe­ models to identify elite soccer [11,13] and futsal [12] players at high (or
tition which resulted in a player being unable to take a full part in future low) risk of injury based on a supervised learning perspective (i.e., it is
soccer training or match play (time loss injury definition). Injuries were defined by its use of labeled datasets according to the class variable
classified as non-contact where no clear contact or collision with another [injury yes vs. injury no] to train algorithms that classify data or predict
player, object or ball occurred. Only lower extremity injuries were outcomes), the taxonomy for external (resampling techniques), internal
considered for the analysis as these incidents are the most common at (ensemble techniques) and cost-sensitive methods for learning with
youth soccer practice [1]. All injuries were recorded by team doctors imbalanced data sets suggested by López et al. [19] and Elkarami et al.
and physiotherapists of each club, and players were considered injured [47] was applied. A brief description of each of the techniques employed
until the medical staff allowed them to fully participate in training and is provided in online supplementary file 8 as well as in previous studies
competition. Injury severity was defined as slight/minimal (1–3 days), [12,13]. According to Robertson [48] four different subsets or categories
minor/mild (4–7 days), moderate (8–28 days), and severe (> 28 days) of base learning algorithms can be defined according to their internal
based on lay-off time from soccer. functioning to help sports practitioners improve their decision-making
The club medical staff documented LE-ST injuries on an injury report processes on training prescription to optimise sports performance and
form described elsewhere [43]. As some inconsistencies in the diagnosis mitigate injury risk: a) regression algorithms (estimating relationships
of minimal LE-ST injuries by medical staff teams were found at the end between variables on a continuous scale [e.g., linear regression, neural
of the 9-month follow-up period, only LE-ST injuries showing a time loss networks]), clustering algorithms (grouping sets of items based on their
of ≥ 4 days were chosen for the subsequent statistical analysis. Due to levels of similarity to one another [e.g., K-means and hierarchical]),
the confounding effects of previous injuries, only the first occurring rule-based algorithms (extracting rules from data based on frequency
injury for each player during the season was considered in the analyses and predictability [e.g., support vector machines and decision rules])
[14,21]. and classification algorithms (identifying which category an instance
belongs to and base on a training set of data [e.g., decision trees and
Random Forest]). Therefore, six well-known learning algorithms (C4.5,
2.4. Statistical analysis ADTree, SMO, KNN, and Random Forest [RF]) from the categories
established by Robertson [48] were selected as base classifiers to be used
Data from questionnaires and field-based tests were collected in in the resampling, ensemble, and cost sensitive methods. With all al­
paper format and transferred into a spreadsheet using a double manual gorithms applied to all base classifiers, a total of 72 models were
data entry processing technique [44]. Identified discrepancies were generated. To allow comparison of the constructed models to a baseline
corrected upon agreement to reach an error level of 0 %. After having model, a ZeroR classifier was also used.
performed a rigorous data cleaning process (identified anomalies or Some specific pre-processing tasks (missing data imputation and
errors were corrected [32 cases]) we had an imbalanced (displaying an feature selection) were exclusively carried out in the training folds so
imbalance ratio of 0.21) and a high-dimensional data set comprising of that the classification task could be performed appropriately. In
260 male youth soccer players and 135 potential risk factors. In this particular, missing data were substituted by the mean value of the cor­
research, an anomaly or error was defined as a value or score that could responding variable according to the age category of the players.
not be classified as true or real because of the consequence of a human Due to the high dimensionality of the data set, before running the
error or a machine failure. An example of an error was a jumping height algorithms included in the taxonomy described in online supplementary
value of 256 cm since it is impossible for an adolescent to jump as such file 8, a feature selection process was conducted with the aim of helping
height. base classifiers to reduce the feature space and eliminate irrelevant,
To assess the performance of the algorithms selected, the fivefold weakly relevant and/or redundant features. Particularly, the meta­
stratified cross-validation technique was applied. The fivefold stratified classifier “attribute selected classifier” available in Weka's repository
cross validation was repeated a hundred times and results were averaged was employed. We used as attribute evaluator the classify subset eval­
over the runs to obtain a more reliable estimate for the predictive ability. uator filter [49] and the GreedyStepwise as search technique. To inter­
A wide variety of classification performance measures may be obtained pret and visualise the behaviour and relevance of the variables selected,
from the stratified cross-validation technique. A well-known approach the Shapley Additive exPlanations (SHAP) approach (SHAP summary
to produce an evaluation criterion is to use the receiver operating plot) was used [50]. This approach visualises every single player or

5
F.J. Robles-Palazón et al. Chaos, Solitons and Fractals: the interdisciplinary journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena 167 (2023) 113079

injury case and gives an overview of the variables in the model by order SMO as base classifier technique was the one that showed the highest F-
of importance (vertically listed features), with the top ones having a score (0.380 ± 0.105) and hence, it was considered as the “best fit
higher global impact on the model than bottom ones. The SHAP-values model”. Therefore, the final screening model to prospectively classify
represent the impact of a variable in the decision-making process. Dots male youth soccer players as having a high or low risk of suffering a LE-
representing the SHAP-values for each feature value of a player in the ST injury in the following 9 months of competitive season comprised
dataset are plotted horizontally next to the feature. Negative SHAP- 100 different SMO (rule-based) classifiers (an example of one of these
values represent a higher probability of a positive prediction (i.e., SMO classifiers can be found in Fig. 2, and the rest may be obtained upon
being injured). Each dot is colored by the value (i.e., measured value) of request to the authors). In terms of practical applications, each classifier
the feature for an individual. has a vote (yes [high risk of LE-ST injury] or no [lower risk of LE-ST
injury]), and the final decision regarding whether or not a player
3. Results might sustain an injury is determined by the combination of the votes of
each individual classifier to each class (yes or no).
3.1. Lower extremity soft-tissue injuries epidemiology For the model finally selected (UBAG with SMO as base classifier), an
analysis of the average influence that each of its six variables has in the
There were 61 LE-ST injuries over the 9-month follow-up period. Of decision-making process regarding whether or not a player might suffer
them, 36 were classified as thigh muscle (18 hamstrings, 8 quadriceps, an injury was carried out by the SHAP approach and can be visualised in
and 10 adductors) injuries, 9 as knee (5 ligament sprains) injuries, and 7 Fig. 3. The variable that demonstrated the biggest impact was knee
as ankle (all ligament sprains) injuries. The distribution of injuries be­ medial displacement (dominant leg) in the DVJ, followed by asymmetry
tween legs was 43 dominant leg and 18 non-dominant leg. A total of 26 in the peak vertical ground reaction force during landing in the SLCMJ,
injuries happened during training sessions and 35 during matches. With body mass index, asymmetry in the frontal plane projection angle
regard to severity, most injuries were categorised as moderate (n = 40), assessed through the TJA, asymmetry in the passive hip internal rotation
while only 6 incidents were classified as severe injuries and 15 as minor/ ROM, and ankle dorsiflexion with the knee extended (dominant leg)
mild injuries. Thirteen players sustained multiple LE-ST injuries during ROM. In Fig. 4, the SHAP values for each feature value of an individual
the observation period (10 players were injured twice and three players in the dataset are displayed.
three times) and thus, only their first incident (i.e., the index injury) was
used for the analyses. Consequently, 45 LE-ST injuries were finally used 4. Discussion
to build the prediction models.
The aim of this study was twofold: a) to build models using machine
3.2. Prediction models for lower extremity soft tissue injuries learning techniques on data from an extensive screening battery to
prospectively predict LE-ST injuries in non-elite male youth soccer
As all the algorithms employed in this study can be found in the players, and b) to compare their performance scores (i.e., accuracy) to
Weka experimenter, only the scheme (and not the full code) of the al­ select the best fit prediction model. In this sense, the present study has
gorithm finally selected is displayed in online supplementary file 9 and built a screening model (AUC = 0.700) based on six pre-season field-
the model is publicly available on https://data.mendeley.com/datasets/ based measures to predict LE-ST injuries in male youth soccer players. In
2mw6w556yg/1 in order to allow practitioners to use it with their male particular, the model developed successfully identifies one out of every
youth soccer players. two (TP rate = 53.7 %) and three out of every four (TN rate = 73.9 %)
The feature selection process conducted in the data set identified a male youth soccer players at high or low risk of suffering a LE-ST injury
subset of six measures as the most relevant (considering the individual throughout the in-season phase, respectively.
predictive ability of each feature as well as the degree of redundancy The ability of the derived model in the current study to predict LE-ST
among them) (Table 2) on which was subsequently applied the taxon­ injuries is similar to the model developed by Oliver et al. [14] (AUC =
omy of learning algorithms explained in the “Materials and methods” 0.663, TP rate = 55.6 %, TN rate = 74.2 %) but lower than the model
section. reported by Rommers et al. [21] (AUC = 0.850, TP rate = 85 %), albeit
The baseline ZeroR classifier achieved an AUC of 0.5 ± 0, specificity both using elite-level male youth soccer players. Three different argu­
of 100 % and sensitivity of 0 %. Table 3 shows the average AUC results ments may explain the higher performance scores reported by Rommers
for all resampling, ensemble and cost-sensitive learning methods sepa­ et al.'s [21] model compared to those shown in the current prediction
rately for each decision base classifiers, nearly all of which have greater models and that built by Oliver et al. [14]:
accuracy and sensitivity than the baseline model. As a result, a total of 3 The first argument that may be used to explain these differences in
algorithms built (using this subset of features) prediction models with the models' performance is the larger number of players that were
AUC scores ≥ 0.7 (Table 4). Among these 3 algorithms, the UBAG with enrolled in the study conducted by Rommers et al. [21] (n = 734) in
comparison with Oliver et al.'s [14] study (n = 355) and the current
Table 2 research (n = 260). In studies dealing with class imbalance problems,
Features selected after having applied the classify subset evaluator filter to the such as the LE-ST injury phenomenon, in which the number of injured
data set. players (minority class) prospectively reported is always much lower
Name Labels than the non-injured participants (majority class) [19,51], large sample
sizes may be required to ensure having enough instances in the minority
KMD (dominant leg) [DVJ] 0 (varus), 1 (slight valgus), 2 (moderate valgus) or 3
(severe valgus)
class to avoid them being considered as noise by the learning algorithms
BMI Numeric during the process of building models. In this sense, Japkowicz & Ste­
ROM-ADFKE (dominant leg) Numeric phen [52] demonstrated that the error rate caused by imbalanced class
Landing BIL-pVGRF [SLCMJ] 0 (Asymmetry) or 1 (No Asymmetry) distribution decreases when the number of examples of the minority
ROM-BIL-PHIR 0 (Asymmetry) or 1 (No Asymmetry)
class is representative. While Rommers et al. [21] identified 368 injured
BIL-FPPA [TJA] 0 (Asymmetry) or 1 (No Asymmetry)
players throughout the follow up, Oliver et al. [14] and the current study
DVJ: drop vertical jump; KMD: knee medial displacement; BMI: body mass used 99 and 45 injuries respectively to develop the prediction models.
index; ROM: range of motion; ADFKE: ankle dorsiflexion with the knee extended; Therefore, in the model built by Rommers et al. [21], patterns that were
pVGRF: peak vertical ground reaction force; SLCMJ: single-leg countermove­
defined by injury players could have been better learned and this may
ment jump; PHIR: passive hip internal rotation; FPPA: frontal plane projection
have positively impacted on its predictive ability.
angle; TJA: tuck jump assessment; BIL: bilateral ratio.
The second argument is linked to the fact that the imbalance ratios

6
F.J. Robles-Palazón et al. Chaos, Solitons and Fractals: the interdisciplinary journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena 167 (2023) 113079

Table 3
AUC results (mean ± standard deviation) for the five base classifiers in isolation and after applying in them the resampling, ensemble (Classic, Boosting-based,
Bagging-based, and Class-balanced ensembles), and cost-sensitive learning techniques selected.
Technique Base classifiers

C4.5 ADTree SMO KNN RF

AUC AUC AUC AUC AUC

None 0.600 ± 0.105 0.619 ± 0.100 0.499 ± 0.005 0.613 ± 0.097 0.605 ± 0.101

Resampling techniques
SMOTE 0.606 ± 0.098 0.620 ± 0.098 0.631 ± 0.088 0.615 ± 0.099 0.613 ± 0.099
ROS 0.603 ± 0.100 0.617 ± 0.098 0.625 ± 0.088 0.613 ± 0.100 0.608 ± 0.099
RUS 0.603 ± 0.100 0.623 ± 0.097 0.619 ± 0.088 0.631 ± 0.096 0.624 ± 0.096
ENN 0.599 ± 0.097 0.619 ± 0.098 0.499 ± 0.007 0.609 ± 0.097 0.618 ± 0.011

Classic ensembles
ADB1 0.636 ± 0.091 0.614 ± 0.098 0.610 ± 0.076 0.575 ± 0.106 – –
M1 0.636 ± 0.092 0.610 ± 0.100 0.682 ± 0.085 0.598 ± 0.095 – –
BAG 0.636 ± 0.096 0.628 ± 0.096 0.568 ± 0.094 0.640 ± 0.098 – –

Boosting-based ensembles
SBO 0.614 ± 0.097 0.611 ± 0.100 0.671 ± 0.091 0.609 ± 0.095 – –
RUSB 0.623 ± 0.098 0.610 ± 0.101 0.677 ± 0.088 0.634 ± 0.092 – –

Bagging-based ensembles
OBAG 0.685 ± 0.079 0.637 ± 0.095 0.697 ± 0.089 0.649 ± 0.096 – –
UBAG 0.653 ± 0.089 0.631 ± 0.096 0.700 ± 0.088 0.667 ± 0.091 – –
SBAG 0.632 ± 0.094 0.638 ± 0.095 0.695 ± 0.089 0.650 ± 0.094 – –

Cost-sensitive classification
MetaCost 0.577 ± 0.099 0.623 ± 0.103 0.500 ± 0.011 0.604 ± 0.097 – –
CS-Classifier 0.597 ± 0.101 0.618 ± 0.098 0.539 ± 0.066 0.621 ± 0.096 – –

Class-balanced ensembles with a cost-sensitive classifier

CS-OBAG 0.631 ± 0.096 0.640 ± 0.095 0.704 ± 0.085 0.648 ± 0.097 – –
CS-UBAG 0.648 ± 0.092 0.637 ± 0.095 0.703 ± 0.084 0.662 ± 0.092 – –
CS-SBAG 0.639 ± 0.092 0.640 ± 0.095 0.699 ± 0.087 0.658 ± 0.094 – –

Highlighted in bold are the algorithms that built prediction models with AUC scores ≥0.7.

Table 4
Sub-set of algorithms that allowed building predictive models with AUC scores ≥0.7.
Technique Performance measures

AUC TP rate (%) TN rate (%) F-score

UBAG [SMO] 0.700 ± 0.088 53.7 ± 17.0 73.9 ± 7.7 0.380 ± 0.105
CS-UBAG [SMO] 0.703 ± 0.084 75.2 ± 14.9 51.0 ± 9.4 0.368 ± 0.060
CS-OBAG [SMO] 0.704 ± 0.085 72.8 ± 15.2 55.1 ± 9.3 0.379 ± 0.066

Highlighted in bold is the algorithm with the highest F-score. AUC: area under the receiver operating characteristic curve; TP: true positive; TN: true negative.

∑ ∑
(IR = injured players / non-injured players) of the dichotomic class Rommers et al. [21] and Oliver et al. [14] might be one of the main
variable (injury yes or no) in Oliver et al.'s [14] study (IR = 0.39) and the reasons for the lower injury rates and consequently the higher imbal­
current study (IR = 0.21) were much higher than the one observed in ance ratio found in the current research. The participants of our study
Rommers et al.'s [21] study (IR = 1.00). In fact, the data set used by routinely completed a total of two (U11–12 players) and three (U13–14,
Rommers et al. [21] to build their injury prediction model did not show U15–16, and U17–19 players) 90-min training sessions per week on non-
an imbalanced distribution in the class variable as the number of injured consecutive days and played one competitive match usually at the
(n = 368) and non-injured (n = 366) players was almost the same. Class weekend. In addition, in all age groups, the competitive season was
distribution (i.e., the proportion of instances [e.g., soccer players] divided into three blocks of 9–12 weeks separated by a 2–3-week break
belonging to each class [injured vs non-injured] in a data-set) plays a key (coinciding with Christmas and Easter festivities). On the contrary, it is
role in classification problems. Highly imbalance data sets usually tend plausible that the elite youth soccer players (mainly those belonging to
to suffer from class overlapping and/or small disjuncts, which difficult the more advanced age groups) who took part in both Oliver et al.'s [14]
classifier learning [51]. Thus, although Oliver et al. [14] and the current and Rommers et al.'s [21] studies could have shown larger (i.e., number
study have used learning algorithms specially designed to deal with class of training session per week) and higher physically demanding weekly
imbalance problems and acceptable predictive accuracy results were exposures to the game of soccer than our participants. This higher fre­
reported, the lower IR in the study of Rommers et al. [21] may have quency and intensity in the exposure to soccer that usually elite
allowed lower misclassification rates and hence, better accuracy scores. adolescent (> 14 years old) players have in comparison with their
In this sense, the different weekly exposure (in terms of frequency and counterpart non-elite players might be attributed to the early sport
physical demands) to the soccer play that could have occurred between specialisation process that usually is observed in the youth academies of
our sample of amateur youth soccer players and the elite ones used by professional soccer clubs. Furthermore, it is also possible that the

7
F.J. Robles-Palazón et al. Chaos, Solitons and Fractals: the interdisciplinary journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena 167 (2023) 113079

Fig. 2. Description of the first UBAG [SMO] classifier. BMI: body mass index; ROM: range of motion; ADFKE: ankle dorsiflexion with the knee extended; BIL: bilateral
ratio; PHIR: passive hip internal rotation; pVGRF: peak vertical ground reaction force; SLCMJ: single-leg countermovement jump; FPPA: frontal plane projection
angle; TJA: tuck jump assessment; KMD: knee medial displacement; DVJ: drop vertical jump.

Fig. 3. SHAP values for each feature. KMD: knee medial displacement; DVJ: drop vertical jump; BIL: bilateral ratio; pVGRF: peak vertical ground reaction force;
SLCMJ: single-leg countermovement jump; BMI: body mass index; FPPA: frontal plane projection angle; TJA: tuck jump assessment; ROM: range of motion; PHIR:
passive hip internal rotation; ADFKE: ankle dorsiflexion with the knee extended.

participants in Oliver et al.'s [14] and Rommers et al.'s [21] studies may ability to predict injuries. In particular, Rommers et al. [21] used a hold
have had shorter Christmas and Easter breaks (in case they had any of out with 20 % of the same as test data to assess the predictive ability of
them) than our amateur youth soccer players. Therefore, the larger and its model whereas Oliver et al. [14] employed a five-fold cross validation
higher physically demanding weekly exposure alongside the shorter technique and the present study repeated 100 times this five-fold cross
resting periods that may have had the elite youth soccer players that validation procedure in an attempt to achieve a more accurate estima­
took part in these two studies may have led them to a progressive and tion of the models' performance. It has been suggested that the k-fold
chronic accumulation of fatigue that could have dramatically increased cross validation is a more powerful preventive technique against model
their risk of injury. However, as neither Oliver et al. [14] nor Rommers performance overfitting than the hold out because the validation metrics
et al. [21] reported the weekly exposure to the game of soccer in their calculated for each fold are combined to give an overall estimate of the
participants, this hypothesis should be considered with a degree of model's performance, reducing the risk of accidentally obtaining a really
caution. optimistic test data [53]. Unlike the current study, neither Rommers
Finally, the last aspect that might have also played a key role in the et al. [21] nor Oliver et al. [14] uploaded their respective data sets into a
higher predictive ability observed in the model published by Rommers public repository. Consequently, we were not able to apply the resam­
et al. [21] is the less exigent resampling method applied to determine its pling technique used in the current study to assess the prediction ability

8
F.J. Robles-Palazón et al. Chaos, Solitons and Fractals: the interdisciplinary journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena 167 (2023) 113079

Fig. 4. SHAP summary plot. The features

in the model are listed from the most (top)
to least (bottom) important by their global
impact on the model. Dots representing
the SHAP values for each feature value of
an individual in the dataset are plotted
horizontally next to the feature. Over­
lapping points are jittered in y-axis direc­
tion, so a sense of the distribution of the
Shapley values per variable is achieved.
The higher the absolute value (either pos­
itive or negative), the higher the impor­
tance in the classification decision-making
process. Positive SHAP values represent a
higher probability of a negative prediction
(i.e., No injured). Each dot is colored by
the value (i.e., measured value) of the
feature for an individual, where blue rep­
resents the lower values (e.g., lower BMI
score) and red the higher values (e.g.,
higher BMI scores). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

of their models in order to inform whether (or not) their performance players [14] and it is deemed to place additional stress on the weaker leg
scores could have suffered from overfitting (and specify to some extent) predisposing it to increased injury risk. Importantly, these six measures
due to a less exigent validation technique. are considered modifiable risk factors and hence, some strategies can be
Another main finding of the current study is that of the 135 potential implemented to optimise these factors in each player to lower the
risk factors obtained from the several questionnaires and field-based probability of suffering a LE-ST injury. In this regard, previous studies
tests carried out during the pre-season testing session conducted in have demonstrated that the regular application of short (not > 20–25
each soccer team, only six (Table 2) were finally selected as the most min) bouts of multi-component exercises during training sessions can
important features related to LE-ST injuries. This subset comprised of an significantly improve, among other aspects, neuromuscular control and
anthropometric parameter (BMI), three neuromuscular measures (KMD performance and help to control body weight in team sport athletes
in the dominant leg [DVJ], landing BIL-pVGRF [SLCMJ] and BIL-FPPA (including young soccer players) [42,57]. Therefore, these multi-
[TJA]) and two joint ROMs (ROM-BIL-PHIR and ROM-ADFKE in the component programs may be powerful tools to be used by practi­
dominant leg) allowed us to build a model to predict LE-ST injuries in tioners as preventive measures in those soccer players categorised at
male youth soccer players. Therefore, one of the main advantages of the high risk of LE-ST injury.
model presented in this study is that it only needs five to ten minutes to Finally, it should be highlighted that simulations ran in our labora­
run the screen in a single player, unlike Rommers et al.'s [21] model tory showed that giving the four basic algorithms used in this study
where 20 measures recorded from a questionnaire and five different (C4.5, ADTree, SMO and KNN) the opportunity to select by themselves
field-based tests are required, which can take longer than 45 min to (according to their own criteria) the most relevant variables did not
collect all data in a single athlete. The six measures selected have been improve the predictive performance of the models but increased its
consistently proposed as primary injury risk factors for LE-ST injuries in complexity. Furthermore, simulations were also run with other attribute
several prospective and biomechanical studies conducted in paediatric evaluators (such as InfoGain and Correlations) to select relevant features
athlete population [14,54,55]. As it is shown in Fig. 3, a higher knee and none of them improved the performance scores presented in this
medial displacement (i.e., dynamic knee valgus) of the dominant leg in study.
DVJ (SHAP score = 0.32) and the presence of asymmetries in pVGRF at
landing from SLCMJ (SHAP score = 0.17) were identified as the two
most important predictors for LE-ST injury. A higher body mass index 4.1. Limitations
(SHAP score = 0.05), bilateral differences ≥ 10 % in FPPA measured
through the TJA manoeuvre (SHAP score = 0.03) and ≥ 8◦ in PHIR ROM This study has also some limitations that should be acknowledged.
(SHAP score = 0.03), and lower ADFKE ROM of the dominant leg (SHAP Even though all the variables collected during the screening session are
score = 0.02) had a smaller effect on the prediction model. It is beyond considered as risk factors for LE-ST injuries, there are additional mea­
the scope of this study to describe into detail the potential mechanisms sures from various questionnaires and field-based tests that were not
that justify the reasons why each of these six measures themselves might assessed in this research (due to time restrictions) and that may have
increase the vulnerability to LE-ST injury in this cohort of soccer players. enhanced the ability to predict LE-ST injuries in this cohort of young
However, the proposed mechanisms might include altered frontal (i.e., athletes (e.g., trunk stability measures, relative leg stiffness, and change
the adoption of an excessive dynamic valgus motion at the knee [high of direction kinematics). Likewise, the complex interaction of growth,
KMD and FPPA scores]), sagittal (ankle ROM) and transverse (hip in­ maturity timing and tempo across players of varying age and maturity
ternal rotation ROM) planes during the execution of high intensity along with the fact that a non-single type of injury (e.g., hamstring
weight-bearing dynamic tasks (e.g., landing from a jump, side-stepping, strains, ACL tears) was analysed may have reduced the ability of the
pressing and tackling) that may produce increased loading of the knee feature selection algorithm applied to the data set to reduce its dimen­
and ankle [54,56]. Likewise, it has been suggested that increased BMI sionality (through removing redundant and not relevant measures), and
scores may imply changes in moments of inertia, forces and de­ thus could have penalised the performance of the model. Future studies
formations experienced by various soft tissues during high intensity should assess whether (or not) the use of more homogeneous samples, in
movements (e.g., high speed running, change of direction) [4], which terms of maturity status, and focusing the attention on single types of
may be associated with injury risk, particularly muscle strains [55]. injury may increase the predictive ability of the screening models.
Asymmetries in pVGRF at landing from SLCMJ have been also identified Another limitation of the current study is that only the first occurring
by previous studies as a primary injury risk factor in male youth soccer injury of every player was considered in the analysis. Consequently,
because players can sustain multiple injuries over one season, the

9
F.J. Robles-Palazón et al. Chaos, Solitons and Fractals: the interdisciplinary journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena 167 (2023) 113079

analysis does not reflect the complete picture. Furthermore, players order of presentation of the authors.
were only tested at the end of the pre-season with subsequent injuries
monitored over the entire season. Anthropometric, physical fitness,
neuromuscular capability and biomechanical measures change over the Declaration of competing interest
course of the season due to training and natural development [21,55],
which may have negatively impacted on the models ability to predict The authors declare that they have no known competing financial
injuries. Therefore, future studies should conduct screening session interests or personal relationships that could have appeared to influence
every few months in order to obtain accurate screening data that is the work reported in this paper.
closer to the time of injury, mitigating the effects of training, growth and
maturation. Data availability

5. Conclusions The scheme of the algorithm finally selected is displayed in online

supplementary file 9 and the model is publicly available on https://data.
Due to the application of machine learning techniques, the current mendeley.com/datasets/2mw6w556yg/1.
study has developed a screening model based on six field-based mea­
sures that showed moderate validity (AUC score = 0.700, TPrate = 53.7 Acknowledgements
% and TNrate = 73.9 % determined through the exigent repeated cross-
validation resampling technique) for identifying youth soccer players at The authors would like to thank the participating players and
risk of LE-ST injury. Furthermore, and thanks to the SHAP approach, it is coaches for their collaboration in this study, and the assessment team for
possible to determine the influence of each risk factor selected (i.e., KMD their support in the data collection.
[dominant leg] in the DVJ, landing BIL-pVGRF [SLCMJ], BMI, BIL-FPPA
[TJA], ROM-BIL-PHIR and ROM-ADFKE [dominant leg]) in the predic­
References
tion model (injury yes vs. injury no). Given that these measures require
little equipment to be obtained and can be employed quickly (approxi­ [1] Robles-Palazón FJ, López-Valenciano A, De Ste Croix MB, Oliver JL, García-
mately 5–10 min) and easily by trained staff in a single player, the model Gómez A, Sainz de Baranda P, et al. Epidemiology of injuries in male and female
youth football players: a systematic review and meta-analysis. J Sport Health Sci
developed in this study might be included as an essential component of
2022;11:681–95.
the injury management strategy in youth soccer. [2] Bult HJ, Barendrecht M, Tak IJR. Injury risk and injury burden are related to age
Supplementary data to this article can be found online at https://doi. group and peak height velocity among talented male youth soccer players. Orthop
org/10.1016/j.chaos.2022.113079. J Sports Med 2018;6:1–10. https://doi.org/10.1177/2325967118811042.
[3] van der Sluis A, Elferink-Gemser MT, Coelho-e-Silva MJ, Nijboer JA, Brink MS,
Visscher C. Sport injuries aligned to peak height velocity in talented pubertal
Funding soccer players. Int J Sports Med 2014;35:351–5. https://doi.org/10.1055/s-0033-
1349874.
[4] Hawkins D, Metheny J. Overuse injuries in youth sports: biomechanical
Francisco Javier Robles-Palazón was initially supported by the Pro­ considerations. Med Sci Sport Exerc 2001;33:1701–7.
gram of Human Resources Formation for Science and Technology [5] Philippaerts RM, Vaeyens R, Janssens M, Van Renterghem B, Matthys D, Craen R,
(20326/FPI/2017) from the Seneca Foundation: Agency for Science and et al. The relationship between peak height velocity and physical performance in
youth soccer players. J Sports Sci 2006;24:221–30. https://doi.org/10.1080/
Technology in the Region of Murcia (Spain) and subsequently by a 02640410500189371.
Margarita Salas postdoctoral fellowship (UMU/R-1500/2021) given by [6] Wik EH, Lolli L, Chamari K, Materne O, Di Salvo V, Gregson W, et al. Injury
the Spanish Ministry of Universities and funded by the European Union- patterns differ with age in male youth football: a four-season prospective study of
1111 time-loss injuries in an elite national academy. Br J Sports Med 2021;55:
NextGenerationEU. Francisco Ayala was supported by a Ramón y Cajal 794–800. https://doi.org/10.1136/bjsports-2020-103430.
postdoctoral fellowship given by the Spanish Ministry of Science and [7] Verhagen E, Bolling C, Finch CF. Caution this drug may cause serious harm! Why
Innovation (RYC2019-028383-I). This study is part of the projects we must report adverse effects of physical activity promotion. Br J Sports Med
2015;49:1–2. https://doi.org/10.1136/bjsports-2014-093604.
entitled “Estudio del riesgo de lesión en jóvenes deportistas a través de
[8] Maffulli N, Longo UG, Gougoulias N, Loppini M, Denaro V. Long-term health
redes de inteligencia artificial” (DEP2017-88775-P) and “El fútbol outcomes of youth sports injuries. Br J Sports Med 2010;44:21–5. https://doi.org/
femenino importa: identificación del riesgo de lesión a través de la 10.1136/bjsm.2009.069526.
inteligencia artificial” (PID2020-115886RB-I00), funded by the Spanish [9] Rössler R, Junge A, Bizzini M, Verhagen E, Chomiak J, aus der Fünten K, et al.
A multinational cluster randomised controlled trial to assess the efficacy of “11 +
Ministry of Science and Innovation, the State Research Agency (AEI) and Kids”: a warm-up programme to prevent injuries in children’s football. Sports Med
the European Regional Development Fund (ERDF). The funders had no 2018;48:1493–504. https://doi.org/10.1007/s40279-017-0834-8.
role in study design, data analysis, interpretation, or the decision to [10] Bahr R. Why screening tests to predict injury do not work—and probably never
will...: a critical review. Br J Sports Med 2016;50:776–80. https://doi.org/
submit the work for publication. 10.1136/bjsports-2016-096256.
[11] Ayala F, López-Valenciano A, Gámez Martín JA, De Ste Croix M, Vera-Garcia FJ,
Ethics approval and consent to participate del García-Vaquero M, P, et al. A preventive model for hamstring injuries in
professional soccer: learning algorithms. Int J Sports Med 2019;40:344–53.
[12] Ruiz-Pérez I, López-Valenciano A, Hernández-Sánchez S, Puerta-Callejón JM, De
The experimental procedures used in this study were in accordance Ste Croix M, Sainz de Baranda P, et al. A field-based approach to determine soft
with the Declaration of Helsinki and were approved by the Ethics and tissue injury risk in elite futsal using novel machine learning techniques. Front
Psychol 2021;12:1–15. https://doi.org/10.3389/fpsyg.2021.610210.
Scientific Committee of the University of Murcia, Spain (ID: 1551/ [13] López-Valenciano A, Ayala F, Puerta JM, De Ste Croix M, Vera-García F,
2017). All participants provided written informed consent prior to the Hernández-Sánchez S, et al. A preventive model for muscle injuries: a novel
study. approach based on learning algorithms. Med Sci Sports Exerc 2018;50:915–27.
https://doi.org/10.1016/j.physbeh.2017.03.040.
[14] Oliver JL, Ayala F, De Ste Croix MBA, Lloyd RS, Myer GD, Read PJ. Using machine
CRediT authorship contribution statement learning to improve our understanding of injury risk and prediction in elite male
youth football players. J Sci Med Sport 2020;23:1044–8. https://doi.org/10.1016/
j.jsams.2020.04.021.
FJR-P, MDSC, PSdB, and FA conceived and designed the research;
[15] Robles-Palazón FJ, Ruiz-Pérez I, Aparicio-Sarmiento A, Cejudo A, Ayala F, Sainz de
FJR-P, AC, FS, and FA obtained the data; FJR-P, JMP-C, JAG, and FA Baranda P. Incidence, burden, and pattern of injuries in spanish male youth soccer
analysed and interpreted the data; FJR-P, PSdB, and FA led the drafting players: a prospective cohort study. Phys Ther Sport 2022;56:48–59. https://doi.
of the manuscript; JMP-C, JAG, MDSC, AC, and FS revised the manu­ org/10.1016/j.ptsp.2022.06.005.
[16] De Ridder R, Witvrouw E, Dolphens M, Roosen P, Van Ginckel A. Hip strength as an
script critically for important intellectual content. All authors have read intrinsic risk factor for lateral ankle sprains in youth soccer players. Am J Sports
and approved the final version of the manuscript, and agree with the Med 2017;45:410–6.

10
F.J. Robles-Palazón et al. Chaos, Solitons and Fractals: the interdisciplinary journal of Nonlinear Science, and Nonequilibrium and Complex Phenomena 167 (2023) 113079

[17] Ko J, Rosen AB, Brown CN. Functional performance tests identify lateral ankle [36] Myer GD, Ford KR, Hewett TE. Tuck jump assessment for reducing anterior
sprain risk: a prospective pilot study in adolescent soccer players. Scand J Med Sci cruciate ligament injury risk. Athl Ther Today 2008;13:39–44.
Sport 2018;28:2611–6. https://doi.org/10.1111/sms.13279. [37] Thomas K, French D, Hayes PR. The effect of two plyometric training techniques on
[18] Padua DA, DiStefano LJ, Beutler AI, de la Motte SJ, DiStefano MJ, Marshall SW. muscular power and agility in youth soccer players. J Strength Cond Res 2009;23:
The landing error scoring system as a screening tool for an anterior cruciate 332–5.
ligament injury-prevention program in elite-youth soccer athletes. J Athl Train [38] Street G, McMillan S, Board W, Rasmussen M, Heneghan JM. Sources of error in
2015;50:589–95. https://doi.org/10.4085/1062-6050-50.1.10. determining countermovement jump height with the impulse method. J Appl
[19] López V, Fernández A, García S, Palade V, Herrera F. An insight into classification Biomech 2001;17:43–54. https://doi.org/10.1123/jab.17.1.43.
with imbalanced data: empirical results and current trends on using data intrinsic [39] Grindem H, Logerstedt D, Eitzen I, Moksnes H, Axe MJ, Snyder-Mackler L, et al.
characteristics. Inf Sci (Ny) 2013;250:113–41. https://doi.org/10.1016/j. Single-legged hop tests as predictors of self-reported knee function in
ins.2013.07.007. nonoperatively treated individuals with anterior cruciate ligament injury. Am J
[20] Bittencourt NFN, Meeuwisse WH, Mendonça LD, Nettel-Aguirre A, Ocarino JM, Sports Med 2011;39:2347–54. https://doi.org/10.1177/0363546511417085.
Fonseca ST. Complex systems approach for sports injuries: moving from risk factor [40] Runacres A, Bezodis NE, Mackintosh KA, McNarry MA. The reliability of force-
identification to injury pattern recognition - narrative review and new concept. Br velocity-power profiling during over-ground sprinting in children and adolescents.
J Sports Med 2016;50:1309–14. https://doi.org/10.1136/bjsports-2015-095850. J Sports Sci 2019;37:2131–7. https://doi.org/10.1080/02640414.2019.1622316.
[21] Rommers N, Rössler R, Verhagen E, Vandecasteele F, Verstockt S, Vaeyens R, et al. [41] Robles-Palazón FJ, Ayala F, Cejudo A, De Ste Croix M, Sainz de Baranda P,
A machine learning approach to assess injury risk in elite youth football players. Santonja F. Effects of age and maturation on lower extremity range of motion in
Med Sci Sport Exerc 2020;52:1745–51. https://doi.org/10.1249/ male youth soccer players. J Strength Cond Res 2022;36:1417–25. https://doi.org/
mss.0000000000002305. 10.1519/JSC.0000000000003642.
[22] Vaeyens R, Malina RM, Janssens M, Van Renterghem B, Bourgois J, Vrijens J, et al. [42] Faude O, Rössler R, Petushek EJ, Roth R, Zahner L, Donath L. Neuromuscular
A multidisciplinary selection model for youth soccer: the Ghent youth soccer adaptations to multimodal injury prevention programs in youth sports: a
project. Br J Sports Med 2006;40:928–34. https://doi.org/10.1136/ systematic review with meta-analysis of randomized controlled trials. Front
bjsm.2006.029652. Physiol 2017;8:791. https://doi.org/10.3389/fphys.2017.00791.
[23] Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a [43] Fuller CW, Ekstrand J, Junge A, Andersen TE, Bahr R, Dvorak J, et al. Consensus
multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the statement on injury definitions and data collection procedures in studies of football
TRIPOD statement. Ann Intern Med 2015;162:55–63. https://doi.org/10.7326/ (soccer) injuries. Br J Sports Med 2006;40:193–201. https://doi.org/10.1136/
M14-0697. bjsm.2005.025270.
[24] Cejudo A, Sainz de Baranda P, Ayala F, De Ste Croix M, Santonja-Medina F. [44] Paulsen A, Overgaard S, Lauritsen JM. Quality of data entry using single entry,
Assessment of the range of movement of the lower limb in sport: advantages of the double entry and automated forms processing-an example based on a study of
rom-sport I battery. Int J Environ Res Public Health 2020;17:7606. https://doi. patient-reported outcomes. PLoS One 2012;7:1–6. https://doi.org/10.1371/
org/10.3390/ijerph17207606. journal.pone.0035087.
[25] Shaffer SW, Teyhen DS, Lorenson CL, Warren RL, Koreerat CM, Straseske CA, et al. [45] Altman DG, Bland JM. Diagnostic tests 3: receiver operating characteristic plots.
Y-balance test: a reliability study involving multiple raters. Mil Med 2013;178: BMJ 1994;309:188. https://doi.org/10.1136/bmj.309.6947.102.
1264–70. https://doi.org/10.7205/MILMED-D-13-00222. [46] Zou Q, Xie S, Lin Z, Wu M, Ju Y. Finding the best classification threshold in
[26] Romero-Franco N, Jiménez-Reyes P, Castaño-Zambudio A, Capelo-Ramírez F, imbalanced classification. Big Data Res 2016;5:2–8. https://doi.org/10.1016/j.
Rodríguez-Juan JJ, González-Hernández J, et al. Sprint performance and bdr.2015.12.001.
mechanical outputs computed with an iPhone app: comparison with existing [47] Elkarami B, Alkhateeb A, Rueda L. Cost-sensitive classification on class-balanced
reference methods. Eur J Sport Sci 2017;17:386–92. https://doi.org/10.1080/ ensembles for imbalanced non-coding RNA data. In: 2016 IEEE EMBS Int. Student
17461391.2016.1249031. Conf. ISC. IEEE; 2016. p. 1–4. https://doi.org/10.1109/EMBSISC.2016.7508607.
[27] Read PJ, Oliver JL, De Ste Croix MBA, Myer GD, Lloyd RS. Landing kinematics in [48] Robertson PS. Man & machine: adaptive tools for the contemporary performance
elite male youth soccer players of different chronologic ages and stages of analyst. J Sports Sci 2020;38:2118–26. https://doi.org/10.1080/
maturation. J Athl Train 2018;53:372–8. https://doi.org/10.4085/1062-6050- 02640414.2020.1774143.
493-16. [49] Hall MA, Smith LA. Practical feature subset selection for machine learning. In:
[28] Ortiz A, Rosario-Canales M, Rodríguez A, Seda A, Figueroa C, Venegas-Ríos H. Comput Sci ’98 Proc. 21st Australas. Comput. Sci. Conf. ACSC’98; 1998. p. 181–91.
Reliability and concurrent validity between two-dimensional and three- [50] Lundberg SM, Erion GG, Lee SI. ConsistentIndividualized Feature Attribution for
dimensional evaluations of knee valgus during drop jumps. Open Access J Sports Tree Ensembles. ArXiv; 2018. 1802.03888.
Med 2016;7:65–73. https://doi.org/10.2147/oajsm.s100242. [51] Galar M, Fernandez A, Barrenechea E, Bustince H, Herrera F. A review on
[29] Robles-Palazón FJ, Ruiz-Pérez I, Oliver JL, Ayala F, Sainz de Baranda P. Reliability, ensembles for the class imbalance problem: bagging-, boosting-, and hybrid-based
validity, and maturation-related differences of frontal and sagittal plane landing approaches. IEEE Trans Syst Man, Cybern Part C (Applications Rev 2012;42:
kinematic measures during drop jump and tuck jump screening tests in male youth 463–84. https://doi.org/10.1109/TSMCC.2011.2161285.
soccer players. Phys Ther Sport 2021;50:206–16. https://doi.org/10.1016/j. [52] Japkowicz N, Stephen S. The class imbalance problem: a systematic study. Intell
ptsp.2021.05.009. Data Anal 2002;6:429–49. https://doi.org/10.3233/IDA-2002-6504.
[30] Myer GD, Ford KR, Hewett TE. New method to identify athletes at high risk of ACL [53] Yadav S, Shukla S. Analysis of k-fold cross-validation over hold-out validation on
injury using clinic-based measurements and freeware computer analysis. Br J colossal datasets for quality classification. In: 2016 IEEE 6th Int. Conf. Adv.
Sports Med 2011;45:238–44. https://doi.org/10.1136/bjsm.2010.072843. Comput. IEEE; 2016. p. 78–83. https://doi.org/10.1109/IACC.2016.25.
[31] Mirwald RL, Baxter-Jones ADG, Bailey DA, Beunen GP. An assessment of maturity [54] Hewett TE, Myer GD, Ford KR, Heidt RS, Colosimo AJ, McLean SG, et al.
from anthropometric measurements. Med Sci Sports Exerc 2002;34:689–94. Biomechanical measures of neuromuscular control and valgus loading of the knee
https://doi.org/10.1249/00005768-200204000-00020. predict anterior cruciate ligament injury risk in female athletes: a prospective
[32] Buela-Casal G, Guillén-Riquelme A, Seisdedos Cubero N. Cuestionario de ansiedad study. Am J Sports Med 2005;33:492–501. https://doi.org/10.1177/
estado-rasgo: adaptación española. Madrid: TEA Ediciones; 2011. 0363546504269591.
[33] Andrade E, Arce C, Armental J, Rodríguez M, de Francisco C. Indicadores del [55] Kemper GLJ, van der Sluis A, Brink MS, Visscher C, Frencken WGP, Elferink-
estado de ánimo en deportistas adolescentes según el Modelo multidimensional del Gemser MT. Anthropometric injury risk factors in elite-standard youth soccer. Int J
POMS. Psicothema 2008;20:630–5. Sports Med 2015;36:1112–7. https://doi.org/10.1055/s-0035-1555778.
[34] Gimeno F, Buceta JM, Pérez-Llanta M. El cuestionario «Características psicológicas [56] Koga H, Nakamae A, Shima Y, Bahr R, Krosshaug T. Hip and ankle kinematics in
relacionadas con el rendimiento deportivo»(CPRD): características psicométricas. noncontact anterior cruciate ligament injury situations: video analysis using
Análise Psicológica 2001;19:93–113. model-based image matching. Am J Sports Med 2018;46:333–40. https://doi.org/
[35] Taylor KL, Sheppard JM, Lee H, Plummer N. Negative effect of static stretching 10.1177/0363546517732750.
restored when combined with a sport specific warm-up component. J Sci Med Sport [57] McHugh MP. Oversized young athletes: a weighty concern. Br J Sports Med 2010;
2009;12:657–61. https://doi.org/10.1016/j.jsams.2008.04.004. 44:45–9. https://doi.org/10.1136/bjsm.2009.068106.

11