Machine learning application in soccer: a systematic review

Goldfit Soccer: A Multidimensional Model for Talent Identification of Young Soccer Players

Studies have provided empirical evidence on the prognostic relevance of test batteries and the "coach's eye" for talent identification. The aims were 1) to compare objective and subjective assessments as well as a combined soccer-specific potential index between future selected and non-selected players and 2) to evaluate the prognostic validity of a multidimensional model as a tool for talent identification in soccer. The sample was composed by 132 male players (14,5 ± 1,4 years; regional competitive level) from U13 to U17 age groups of a Brazilian soccer talent development program. Athletes completed a multidimensional test battery and were subjectively rated by their coaches for sporting potential. Players' success was evaluated five years later and was operationalized by achieving under-20 category of the Brazilian Championship or adult professional level (success rate, 15.9%). Confirming univariate prognostic validity, future selected outperformed non-selected players regarding 20-m sprint (p = .009), agility (p = .04), countermovement jump (p = .04), sit-and-reach (p = .001), Yo-Yo IR1 (p = .001), dribbling (p < .001), perceived competence (p = .007), peaking under pressure (p = .01), confidence/motivation (p = .03), coping skills (p = .03), intangibles (p < .001) and player potential (p < .001). A combined index (objective tests, athlete's assessments and coach's eye) named Gold Score Soccer (GSS) showed high prognostic validity (p < .001). A binary logistic regression estimated the probability of success (yes/not) with GSS, ambidextrous and predicted age at peak height velocity as predictors. This multidimensional model named GoldFit Soccer showed high prognostic validity (sensitivity = 85.7%; specificity = 83.8%; accuracy = 84.1%; area under the ROC curve = .93 [.87-.98]). Thus, GoldFit Soccer is a valid multidimensional scientific model for talent identification in soccer.

Who are the best passing players in professional soccer? A machine learning approach for classifying passes with different levels of difficulty and discriminating the best passing players

The present study aimed to assess the use of technical-tactical variables and machine learning (ML) classifiers in the automatic classification of the passing difficulty (DP) level in soccer matches and to illustrate the use of the model with the best performance to distinguish the best passing players. We compared eight ML classifiers according to their accuracy performance in classifying passing events using 35 technical-tactical variables based on spatiotemporal data. The Support Vector Machine (SVM) algorithm achieved a balanced accuracy of 0.70 ± 0.04%, considering a multi-class classification. Next, we illustrate the use of the best-performing classifier in the assessment of players. In our study, 2,522 pass actions were classified by the SVM algorithm as low (53.9%), medium (23.6%), and high difficulty passes (22.5%). Furthermore, we used successful rates in low-DP, medium-DP, and high-DP as inputs for principal component analysis (PCA). The first principal component (PC1) showed a higher correlation with high-DP (0.80), followed by medium-DP (0.73), and low-DP accuracy (0.24). The PC1 scores were used to rank the best passing players. This information can be a very rich performance indication by ranking the best passing players and teams and can be applied in offensive sequences analysis and talent identification.

Predicting Soccer Players' Fitness Status Through a Machine-Learning Approach

PURPOSE

The study had 3 purposes: (1) to develop an index using machine-learning techniques to predict the fitness status of soccer players, (2) to explore the index's validity and its relationship with a submaximal run test (SMFT), and (3) to analyze the impact of weekly training load on the index and SMFT outcomes.

METHODS

The study involved 50 players from an Italian professional soccer club. External and internal loads were collected during training sessions. Various machine-learning algorithms were assessed for their ability to predict heart-rate responses during the training drills based on external load data. The fitness index, calculated as the difference between actual and predicted heart rates, was correlated with SMFT outcomes.

RESULTS

Random forest regression (mean absolute error = 3.8 [0.05]) outperformed the other machine-learning algorithms (extreme gradient boosting and linear regression). Average speed, minutes from the start of the training session, and the work:rest ratio were identified as the most important features. The fitness index displayed a very large correlation (r = .70) with SMFT outcomes, with the highest result observed during possession games and physical conditioning exercises. The study revealed that heart-rate responses from SMFT and the fitness index could diverge throughout the season, suggesting different aspects of fitness.

CONCLUSIONS

This study introduces an "invisible monitoring" approach to assess soccer player fitness in the training environment. The developed fitness index, in conjunction with traditional fitness tests, provides a comprehensive understanding of player readiness. This research paves the way for practical applications in soccer, enabling personalized training adjustments and injury prevention.

Relationship between external and internal load indicators and injury using machine learning in professional soccer: a systematic review and meta-analysis

This study verified the relationship between internal load (IL) and external load (EL) and their association on injury risk (IR) prediction considering machine learning (ML) approaches. Studies were included if: (1) participants were male professional soccer players; (2) carried out for at least 2 sessions, exercises, or competitions; (3) correlated training load (TL) with non-contact injuries; (4) applied ML approaches to predict TL and non-contact injuries. TL included: IL indicators (Rating of Perceived Exertion, RPE; Session-RPE, Heart Rate, HR) and EL indicators (Global Positioning System, GPS variables); the relationship between EL and IL through index, ratio, formula; ML indicators included performance measures, predictive performance of ML methods, measure of feature importance, relevant predictors, outcome variable, predictor variable, data pre-processing, features selection, ML methods. Twenty-five studies were included. Eleven addressed the relationship between EL and IL. Five used EL/IL indexes. Five studies predicted IL indicators. Three studies investigated the association between EL and IL with IR. One study predicted IR using ML. Significant positive correlations were found between S-RPE and total distance (TD) (r = 0.73; 95% CI (0.64 to 0.82)) as well as between S-RPE and player load (PL) (r = 0.76; 95% CI (0.68 to 0.84)). Association between IL and EL and their relationship with injuries were found. RPE, S-RPE, and HR were associated with different EL indicators. A positive relationship between EL and IL indicators and IR was also observed. Moreover, new indexes or ratios (integrating EL and IL) to improve knowledge regarding TL and fitness status were also applied. ML can predict IL indicators (HR and RPE), and IR. The present systematic review was registered in PROSPERO (CRD42021245312).

Exploring the interplay between metabolic power and equivalent distance in training games and official matches in soccer: a machine learning approach

Introduction: This study aimed to explore the interplay between metabolic power (MP) and equivalent distance (ED) and their respective roles in training games (TGs) and official soccer matches. Furthermore, the secondary objective was to investigate the connection between external training load (ETL), determined by the interplay of metabolic power and equivalent distance, and internal training load (ITL) assessed through HR-based methods, serving as a measure of criterion validity. Methods: Twenty-one elite professional male soccer players participated in the study. Players were monitored during 11 months of full training and overall official matches. The study used a dataset of 4269 training games and 380 official matches split into training and test sets. In terms of machine learning methods, the study applied several techniques, including K-Nearest Neighbors, Decision Tree, Random Forest, and Support-Vector Machine classifiers. The dataset was divided into two subsets: a training set used for model training and a test set used for evaluation. Results: Based on metabolic power and equivalent distance, the study successfully employed four machine learning methods to accurately distinguish between the two types of soccer activities: TGs and official matches. The area under the curve (AUC) values ranged from 0.90 to 0.96, demonstrating high discriminatory power, with accuracy levels ranging from 0.89 to 0.98. Furthermore, the significant correlations observed between Edwards' training load (TL) and TL calculated from metabolic power metrics confirm the validity of these variables in assessing external training load in soccer. The correlation coefficients (r values) ranged from 0.59 to 0.87, all reaching statistical significance at p < 0.001. Discussion: These results underscore the critical importance of investigating the interaction between metabolic power and equivalent distance in soccer. While the overall intensity may appear similar between TGs and official matches, it is evident that underlying factors contributing to this intensity differ significantly. This highlights the necessity for more comprehensive analyses of the specific elements influencing physical effort during these activities. By addressing this fundamental aspect, this study contributes valuable insights to the field of sports science, aiding in the development of tailored training programs and strategies that can optimize player performance and reduce the risk of injuries in elite soccer.

An Overview of Machine Learning Applications in Sports Injury Prediction

Use injuries, i.e., injuries caused by repetitive strain on the body, represent a serious problem in athletics that has traditionally relied on historic datasets and human experience for prevention. Existing methodologies have been frustratingly slow at developing higher precision prevention practices. Technological advancements have permitted the emergence of artificial intelligence and machine learning (ML) as promising toolsets to enhance both injury mitigation and rehabilitation protocols. This article provides a comprehensive overview of recent advances in ML techniques as they have been applied to sports injury prediction and prevention. A comprehensive literature review was conducted searching PubMed/Medline, Institute of Electrical and Electronics Engineers (IEEE)/Institute of Engineering and Technology (IET), and ScienceDirect. Ovid Discovery and Google Scholar were used to provide additional aggregate results and a grey literature search. A focus was placed on papers published from 2017 to 2022. Algorithms of interest were limited to K-Nearest Neighbor (KNN), K-means, decision tree, random forest, gradient boosting and AdaBoost, and neural networks. A total of 42 original research papers were included, and their results were summarized. We conclude that given the current lack of open source, uniform data sets, as well as a reliance on dated regression models, no strong conclusions about the real-world efficacy of ML as it applies to sports injury prediction can be made. However, it is suggested that addressing these two issues will allow powerful, novel ML architectures to be deployed, thus rapidly advancing the state of this field, and providing validated clinical tools.

SoccerNet: A Gated Recurrent Unit-based model to predict soccer match winners

Winning football matches is the major goal of all football clubs in the world. Football being the most popular game in the world, many studies have been conducted to analyze and predict match winners based on players' physical and technical performance. In this study, we analyzed the matches from the professional football league of Qatar Stars League (QSL) covering the matches held in the last ten seasons. We incorporated the highest number of professional matches from the last ten seasons covering from 2011 up to 2022 and proposed SoccerNet, a Gated Recurrent Unit (GRU)-based deep learning-based model to predict match winners with over 80% accuracy. We considered match- and player-related information captured by STATS platform in a time slot of 15 minutes. Then we analyzed players' performance at different positions on the field at different stages of the match. Our results indicated that in QSL, the defenders' role in matches is more dominant than midfielders and forwarders. Moreover, our analysis suggests that the last 15-30 minutes of match segments of the matches from QSL have a more significant impact on the match result than other match segments. To the best of our knowledge, the proposed model is the first DL-based model in predicting match winners from any professional football leagues in the Middle East and North Africa (MENA) region. We believe the results will support the coaching staff and team management for QSL in designing game strategies and improve the overall quality of performance of the players.