Current Approaches to the Use of Artificial Intelligence for Injury Risk Assessment and Performance Prediction in Team Sports: a Systematic Review

Artificial intelligence (AI) techniques and methods have attracted considerable attention in the information industry and in society as a whole, due to the large amount of data and the imminent need to transform this data into useful knowledge and practical solutions [1‐3]. However, the effective use of data in some areas is still under development, as is the case in sports. As in most other areas of society, increasing volume of data has been gathered in all kinds of sports, and automated data analysis became an important and fast developing field. Careful analyses of these large data sets can enhance our knowledge in sport sciences while at the same time assist in the decision-making of the practitioners who work on the optimization of training and competition strategies [4, 5].

Data science has emerged as a strategical area to exploit knowledge in sports science aiming to fill some gaps left by traditional statistical methods. As a hybrid knowledge area, data science is more than the combination of statistics and computer science as it requires training in how to weave statistical and computational techniques into a larger framework, problem by problem, and to address discipline-specific questions [6]. A holistic view of data science requires an understanding of the context of data, appreciating the responsibilities involved in using private and public data, and clearly communicating what a dataset can and cannot tell us about the real world [6], in our case, in the sports world. Based on learning models, the algorithms can be tuned and be optimized in order to produce better results for supporting decisions and provide applied knowledge to athletes and sport professionals. These algorithms are applied as supervised learning (e.g., classification and regression) and unsupervised learning (e.g., clustering). The supervised learning requires input and output data to develop a predictive model whereas the unsupervised learning is based on input data only [7]. Data science is an emerging area in both industry and academic ways, leading to more evidence-based decision-making across many walks of life, including social networking services, streaming services, health care, manufacturing, education, financial modeling, policing, and marketing [8‐11]. Also, in our area, linking science and technology to increase sports efficiency has been touted as a path with a promising future. For this to happen, the work directed toward the innovation, introduction, and improvement of processes performed by research and development (R&D) departments in the world’s largest technology companies has been also suggested in the sport field. Because of its fast-moving environment, sport professionals combine data (e.g., physical, technical and tactical) with their expert opinion to inform decisions on the players [12].

However, success in team sport, from the sport science and medicine staff perspective, is to use evidence-based knowledge in an effective manner to develop the decision-making process for injury risk reduction and athletes’ performance optimization [13]. How can the adoption of the research findings and innovations be improved hence improving the compliance of players and coaches/managers with injury prevention and performance enhancement programs? To answer this question, we must accept that players’ needs come first and that players and coaches are the main actors. Therefore, we need to add value to them, offering solutions that really affect their daily lives [13].

One key issue in the sports industry is the possibility to predict injury risk and performance. Historically, the ability of the coaching staff to prescribe training to achieve optimal athletic performance with low risk of injury can be attributed to many years of personal experience. However, modern approaches aiming in adopting scientific methods for the effective development of optimal training programs are warranted [14]. The application of contemporary statistical approaches from AI open an interesting perspective for dealing with injury prevention and for improving the performance models [15, 16].

Therefore, the understanding of the state-of-the-art of AI techniques or methods applied to team sport warrants investigation. With this in mind, the purpose of this study is to give an overview of the current state of the application of AI in team sports. In particular, our review is aiming in answering the following questions: (1) Which AI approaches have been applied to studies investigating injury risk and sport performance in team sports? (2) What are the team sports that have been using AI techniques for predicting injury risk and athletes' performance?

The initial search returned 3086 articles (for details see Fig. 1). After the removal of duplicated articles (n = 614), a total of 2472 studies were retained for full-text screening. Following eligibility assessment, 5 studies with a risk of bias score of 4 or less were considered poor and were excluded (for details see Additional file 1: Table S1). The remaining studies were evaluated between 6 and 7 points in terms of quality. We were unable to find the full manuscript for 3 research studies which were excluded from further analysis. During the revision of the reference lists, 1 study found meeting all the criteria and was included in the analysis. Thus, 58 studies in total were included in this systematic review [15, 20‐76].

The main purpose of this study is to provide an overview of the current state of the application of AI in assessing the injury risk and predicting performance in team sports athletes. The AI techniques or methods with the greatest potential were artificial neural networks, decision tree classifier, Markov process, and support vector machine.

Lapham and Bartlett [77]’s study was one of the first studies using AI in the analysis of sports performance. In 1995, they demonstrated that the increasing role of computers in the decision-making process, through the use of AI techniques, would be a potentially rewarding future direction for the discipline and already pointed out the artificial neural networks successes in experiments, along with their potential benefits. Another study [78] published over 20 years ago reported the use of AI, with decision tree classifier and Bayesian classification, in the diagnosing of sport injuries. Naive Bayesian classifier with the fuzzy discretization of numerical attributes was superior to other methods and was estimated as the most appropriate for practical use [78]. However, this method is not so popular nowadays, whereas the decision tree classifier has been mostly used so far.

McCall et al. [12] recommended that professional soccer teams included R&D into their daily activities to reduce injuries and to optimize performance. Most of the participants in the studies included in our paper were soccer players (26%), and this could be due to the fact that most of the studies were in soccer. In this particular sport, AI was applied to “training load” [71, 73], “training process/knee injury causes” [63], “heart defect detection” [64], “ground reaction force pattern” [65], “psychosocial stress factors” [69], and “screening” [15]. In summary, the AI techniques or methods applied to predict the injury risk (n = 7 studies) in the sample are composed of collegiate (14%) and professional (86%) soccer players and were distributed as following: 43% were on artificial neural network [63‐65], 29% on decision tree classifier [15, 71], and 14% on each of Bayesian logistic regression [69] and least absolute shrinkage and selection operator [73] (for details see Additional file 2: Table S2). Regarding the soccer performance domain, AI was applied to “technical and tactical analysis” [30, 36, 44, 45, 48, 50], “match attendance” [31], as well as on “psychological dynamics of cooperative teamwork” [35]. The AI techniques or methods used to predict sporting performance (n = 8 studies) in youth academy (13%), semi-professional (13%), and professional (75%) soccer players were as follows: 25% each of artificial neural network [30, 31], Bayesian network [35, 36] and decision tree classifier [44, 45], and 13% each of fuzzy clustering [48] and K-means clustering [50]. A great challenge for the future is the application of AI models to integrate all related variables for injury risk assessment and performance prediction.

The second sport in AI application was basketball. Most of the studies on performance were in the “technical and tactical analysis” area [20, 22‐25, 41‐43, 55, 62], whereas one study was on the “physical, technical, and tactical analysis” [21]. Their focus justifies the recognized importance of physical, technical, and tactical analysis for basketball success [79, 80]. On a separate note, one study described a system for the automatic detection and tracking of the ball trajectory during a free throw based on algorithms [81]. According to the authors of that paper, coaches using this system will be able to monitor the trajectory of the ball and the parameters of the free throw, a fact that will help them correct the technique and hence improve the athletic performance of basketball players [81]. Our analysis showed that the AI techniques or methods used to predict sporting performance (n = 11 studies) were with youth academy (9%), semi-professional (9%), and professional (82%) basketball players. The methods reported were as follows: on each artificial neural network (55%) [20‐25], on decision tree classifier (27%) [41‐43], and 9% each of Markov process [55] and support vector machine + decision tree classifier [62]. For injury risk assessment, the studies were based on “training process/knee injury causes” [66] and “heart defect detection” [64]. The prediction of knee injury risk is essential due to the high costs of medical treatment [82] and to the high rates of medically disqualifying injuries among young athletes [83]. Heart defect detection for basketball players is also highly desired because sudden cardiac death risk among males was estimated at more than 10 times higher than that in the overall athletic population of the National Collegiate Athletic Association (NCAA) Division 1 (i.e., 1 in 5,200 vs. 1 in 53,703 athletes per year) [84, 85]. For injury risk prediction, the most frequently used AI technique or method (n = 2 studies) applied in collegiate and professional basketball players was artificial neural network [64, 66].

AI application into handball and volleyball were well reported. In terms of injury risk assessment, ground reaction force patterns is a standard method of investigation in sports medicine and biomechanics [65] that is also related to knee injuries in multidirectional sports, such as handball and volleyball [86, 87]. These injuries may occur during activities, such as jump landing, cutting, and pivoting [86, 87]. Furthermore, injury prediction with AI technique or method based on “screening” [15] was successfully applied to handball players. This approach may assist in identifying the injury risk with higher probability compared to the use of isolated screening tests [88, 89]. The AI techniques or methods (n = 2 studies) used to predict injury risk in professional handball (100%) and volleyball (50%) players were artificial neural network [65] and decision tree classifier [15]. For handball performance, all the studies were on the “technical and tactical analysis” area with youth academy based studies using artificial neural network (n = 4 studies) [26‐29]. For volleyball performance, most of the studies were on the “technical and tactical analysis” area too [33, 52, 53, 61] and only one, with professional players, was on “physical, technical, and tactical analysis” [51]. AI techniques or methods mostly used for performance prediction (n = 5 studies) were the Markov process (40%) [52, 53], artificial neural network (20%) [33], K-nearest neighbor (20%) [51], and support vector machine (20%) [61]. The areas of physical, technical, and tactical analysis are of high importance to athletic performance in both sports [90, 91], and coaches/players may now be able to use the tools of AI for better decision-making.

Training load management is a concern in Australian football [92‐94]. Along this line, AI methods and techniques have been used to predict the risk of injury through the “training load” [67] and “screening” [70] in that sport. Interestingly, there is only one study where the authors did not recommend the supervised learning techniques based on “screening” [70]. In that study, the use of eccentric hamstring strength, age, and previous hamstring strain injury [70] was not able to predict the risk of injury. In this latter study, the Bayesian network was applied in professional Australian football players with the area under the curve (AUC) of 54% being classified as a poor performance metric [70]. On the other hand, when the artificial neural network based on “training load” was used with a sample of professional athletes of the same sport, it was recommended by the authors [67]. The prediction of performance in Australian football was based on heart rate [38] and “technical and tactical analysis” [39, 40] data with the use of decision tree classifier (n = 3 studies in total). The authors concluded that AI could add benefits and, hence, was recommended in Australian football [38‐40].

Injuries cost over 1 billion dollars to Major League Baseball teams, and thus, preventing them is of high priority in that sport too [95]. There is a predominance of upper extremity injuries in baseball players, the surgical procedure most closely associated with baseball being the reconstruction of the ulnar collateral ligament of the elbow [96]. So far, AI has been used to predict the injury risk after ulnar collateral ligament reconstruction in professional baseball players [76]. With regard to performance improvement in baseball [95], we only found AI approaches in the “technical and tactical analysis” area [32, 37, 57, 58]. The AI techniques or methods used to predict sporting performance (n = 4 studies) in professional baseball players were support vector machine [57, 58], artificial neural network [32], and Bayesian network [37].

Data show that American football players who sustained concussions may experience significant salary reductions and perform worse after concussion [97]. According to Navarro et al. [97], the year-over-year change in contract value for the concussion group resulted in a mean overall salary reduction of USD $300 k ± $1,300 k/year. Moreover, the performance score reduction for all offensive scoring players sustaining concussions was statistically significant (i.e., pre, 9 ± 6 fantasy football point “FFP”s/game; post, 7 ± 4 FFPs/game). These aforementioned factors supported the search for AI techniques and methods to assist in predicting the risk of a concussion in a more accurate way [74, 75]. In addition, AI techniques or methods on “technical and tactical analysis” area have been applied in American football players [46, 54]. This sample was composed by collegiate [54, 75], professional, [46] and retired professional [74]. The most frequently used methods in that sport were the support vector machine (50%) [74, 75] and decision tree classifier (25%) [46] with Markov process (25%) [54].

With regard to the rest of team sports, AI applications were reported in professional ice hockey (Markov process [56] and support vector machine [59, 60]), cricket (K-means clustering [49]), field hockey (decision tree classifier [47]), and rugby (artificial neural network [34]). All these AI applications were on the “technical and tactical analysis” area, confirming the importance of these aspects for success in the team sports [98]. With regard to injury risk prediction, the decision tree classifier has been used in professional rugby players using the “training load” as the predictor [72]. In beach volleyball players, the monitoring data from wearable sensors has been used to model the injury risk via artificial neural network [68]. These studies highlight the importance of adequate management of training load in these sports too [99‐101] along with the use of technology and innovative approaches in data management in order to protect the athletes’ health [102, 103].

Our analysis showed that the AI techniques or methods for predicting injury risk and sporting performance mostly used in team sports were artificial neural networks, decision tree classifier, support vector machine, and Markov process. The team sports with the most AI applications were soccer, basketball, handball, and volleyball. The current state of development in the area proposes a promising future with regard to AI use in team sports. Further evaluation research based on prospective methods is warranted to establish the predictive performance of specific AI techniques and methods.