Conceptual Structure and Current Trends in Artificial Intelligence, Machine Learning, and Deep Learning Research in Sports: A Bibliometric Review

Abstract

Artificial intelligence and its subcategories of machine learning and deep learning are gaining increasing importance and attention in the context of sports research. This has also meant that the number of corresponding publications has become complex and unmanageably large in human terms. In the current state of the research field, there is a lack of bibliometric analysis, which would prove useful for obtaining insights into the large amounts of available literature. Therefore, the present work aims to identify important research issues, elucidate the conceptual structure of the research field, and unpack the evolutionary trends and the direction of hot topics regarding key themes in the research field of artificial intelligence in sports. Using the Scopus database, 1215 documents (reviews and articles) were selected. Bibliometric analysis was performed using VOSviewer and bibliometrix R package. The main findings are as follows: (a) the literature and research interest concerning AI and its subcategories is growing exponentially; (b) the top 20 most cited works comprise 32.52% of the total citations; (c) the top 10 journals are responsible for 28.64% of all published documents; (d) strong collaborative relationships are present, along with small, isolated collaboration networks of individual institutions; (e) the three most productive countries are China, the USA, and Germany; (f) different research themes can be characterized using author keywords with current trend topics, e.g., in the fields of biomechanics, injury prevention or prediction, new algorithms, and learning approaches. AI research activities in the fields of sports pedagogy, sports sociology, and sports economics seem to have played a subordinate role thus far. Overall, the findings of this study expand knowledge on the research situation as well as the development of research topics regarding the use of artificial intelligence in sports, and may guide researchers to identify currently relevant topics and gaps in the research.

1. Introduction

In general, both artificial intelligence (AI) (hereafter, the term artificial intelligence (AI) is used even when individual publications refer specifically to machine learning (ML) or deep learning) and the subcategories machine learning (ML) and deep learning have experienced a boom in recent years. AI is generally understood to be a type of human intelligence that is executed by machines. Machine learning is an approach to achieving AI, describing computer systems that are able to learn from examples or experiences without being explicitly programmed. Deep learning has also emerged as an important technique for implementing ML as method based on artificial neural networks in the context of automatic feature discovery through representation learning [1].

AI is established in many situations and is now indispensable. For example, AI has experienced great success in facial recognition on smartphones [2], as well as speech processing, synthesis, and translation (e.g., Google Translate, Alexa, Google Assistant, ChatGPT [3]), and it has long since become routine in everyday life. It has proven itself in numerous real-world applications, including facial recognition [4], speech recognition [5], object recognition [6], malware detection [7], and spam filtering [8]. In addition, the main areas of application can be found in safety-critical fields of application in which errors have a tremendous impact (e.g., autonomous driving [9] or healthcare and medicine [10,11,12,13]).

The exponential increase in AI in recent years, along with the progressive development of the associated approaches, has been driven by various factors—including the availability (i.e., quality and quantity) of data, which decisively influences the success of AI models [14]. There have often been limited data in the past, but today, modern technologies often generate vast amounts of large and complex datasets [15]. In contrast to the original problem with the availability of data, there are now difficulties in coping with the sheer volume of data [16]. Parallel to the increase in the amount of available data, immense developments have taken place in the field of modeling, which have contributed to models becoming more and more accurate in general when considering a variety of core tasks [1]. In addition, the implementation of ML models is becoming easier and more accessible for a large number of people thanks to user-friendly open-source programming libraries such as scikit-learn [17] or PyTorch [18]. Furthermore, hardware developments, the associated increase in available computing power, and the consequent acceleration of calculations play a decisive role [19]. These drivers have ensured that AI has opened up many new possibilities and gradually found its way into sports contexts and applications [20,21]. Hence, the present work is based on an understanding of the field of AI in sports as the application of methods from the general field of AI in the context of sports science and practical sports issues.

The USA is considered to be a pioneer in using AI in sports, both in research and in practice. AI has been used for game and player analysis for several years, especially in American football and baseball [22]. For example, studies have shown that the analysis of various types of data can help support coaches’ tactical decision-making [23], as well as training and competition planning [24]. Furthermore, this kind of analysis has great potential for preventing injuries [20] which, in turn, is accompanied by a reduction in negative consequential costs and cancellations of competitions. Increasing numbers of major football clubs are also using AI and data analytics. First and foremost, FC Barcelona and TSG Hoffenheim have their own research labs (Barça Innovation Hub and TSG Research Lab, respectively). Liverpool is also implementing data analytics solutions and AI to support managerial decisions [25]. Even fan interactions can be monitored and analyzed for better performances [26]. On the topic of available data for analysis purposes, players are monitored via numerous different sensors during matches, competitions, and training. For example, with regard to the sub-disciplines of sports science, applications can be found in the fields of training science [24], biomechanics [27], and sports medicine [28]. Against the backdrop of the ever-increasing amounts of data, AI has become a key technology.

The research situation with respect to AI technologies in sports has also led to a large bibliographic dataset, which is impractical for manual investigation. There is a need to obtain optimal insights into the available publication data using appropriate analysis techniques since it is crucial to analyze the relevant research and structures in depth to uncover existing knowledge, challenges, and research deficits in order to derive and justify new research necessities. A promising approach, which is becoming increasingly important in many fields of research, is bibliometric analysis. For example, bibliometric analyses are used for uncovering emerging trends, collaboration characteristics, and research constituents, as well as for exploring the intellectual structures of specific domains based on their respective literature base [29,30]. Bibliometric techniques (as well as meta-analyses) seem especially promising and suitable for making sense of the large volumes of literature, as classic review methods need to focus on a manageable amount of literature [31]. Such techniques are superior to meta-analysis if the publications in question are very heterogeneous (see https://elicit.org (accessed on 19 December 2022) as a tool for research questions in this context). Furthermore, unlike classic systematic literature reviews, which tend to be affected by the interpretation bias of scholars, bibliometric analysis (or meta-analysis) relies on more quantitative methods; thus, it can avoid or at least reduce such bias [29].

Regarding the basic review situation, several works have been published to date in the context of AI and sport-specific fields of its application [21,32,33,34,35]. In addition to AI topics, numerous general works can be found in the context of sports using bibliometric methods (e.g., football [36,37,38]; sports tourism [39,40]; sports sustainability [41]; sports management [42]; sports nutrition [43]; sports science in general [44]). However, looking at different literature analysis methods applied in the context of sports-related AI research shows that only a few works have been published to date using bibliometric techniques. The authors of [45] analyzed literature regarding big data and artificial intelligence in sports using a bibliometric approach; however, their focus was restricted to articles in the CNKI database. The authors of [46] analyzed trends of AI applications in sports, with a focus on patents. Moreover, some works have used bibliometric techniques in the context of sports, focusing on topics that overlap with AI. For example, [47] analyzed publicly available baseball data in the context of big data characteristics. Furthermore, [48] investigated applications of wearable devices in sports using bibliometric techniques. In addition to bibliometric methods, even machine learning was used itself to examine details on citations, keywords, co-authors, and important topics in the context of sports [49].

To the best of our knowledge, there is still a research deficit in the bibliometric analysis of available literature in the field of AI in sports using a broad literature base. Thus, great potential remains, as the overall scientific landscape has not been thoroughly analyzed so far. Consequently, to guide researchers in identifying currently relevant research topics, research gaps, as well as collaboration opportunities, the general aim of the present work is to provide a better understanding of the research situation regarding the use of AI in sports.

Therefore, the present work aims to address the following research questions in the field of AI in sports using bibliometric methods: (i) What are important research issues? (ii) How is the conceptual structure of the research field structured? (iii) What are evolutionary trends and direction of hot topics related to key issues? Due to the presence of large bibliographic datasets, as well as the broad and heterogeneous literature in the context of machine learning research in sports, the application of a bibliometric analysis appears to be excellently suited to the intended object of investigation.

The following contributions of research constituents were analyzed using bibliometric methods through general performance analysis, and their relationships were identified using science mapping. Longitudinal development and current trends were further analyzed by looking at the occurrence of keywords over time and in the latest research publications. For in-depth background information on the applied analysis techniques, see [29], for example.

2. Materials and Methods

Although several bibliometric studies use the World of Science database [50,51], data from the Scopus database were used for this study (accessed on 6 September 2022). This decision is based on the platform’s broad geographic and thematic coverage that it offers. In particular, with regard to the intended research aim, Scopus seems advantageous as it has a larger journal coverage compared to Web of Science, especially in the fields of social sciences and arts and humanities [52]. Additionally, this database has been extensively used in sports related bibliometric analyses [53,54,55,56]. To map the most prominent subcategories of artificial intelligence, the terms “machine learning” and “deep learning” were included in the search, and “artificial intelligence” was the general term. It should be noted that the Scopus search engine is not case-sensitive and considers hyphens as spaces.

The document types selected were articles and reviews, as commonly performed in literature searches [40,57], with a focus on works of the highest quality. As usual in bibliometric analyses, we focused on one language only (i.e., English). Due to the broad range of different subject areas addressed by publications in the context of sports, the exclusion of inappropriate literature based on subject areas of the database was not suitable for the aims of this work. Inappropriate literature in this context can be defined as publications that use the term “sport*” in the sense of a conceivable application of their research but do not carry out the research in the direct context of sports. However, in the present work, the focus was on works that are closely related to sports. We dealt with this problem using a literature search focusing on titles and the keywords, searching only for publications that included the term “sport*” in their title or keywords. Overall, this procedure allowed us to ensure that sports were the direct focus of the research rather than a side issue. However, there was also the potential for inclusion of papers whose association with sports is debatable, depending on the individual and cultural understanding of the nature of the sport. The final search procedure is shown in detail in Figure 1, resulting in 1215 documents from 428 journals (1157 articles, 58 reviews).

Based on the final search, a csv file was exported. Erroneous or incomplete entries, along with duplicates, were removed. Furthermore, abbreviations were replaced with their full designations (e.g., AI → artificial intelligence; ML → machine learning), while grammatical numbers were unified in the keywords (e.g., sport → sports) in the exported csv file. Incorrect and inappropriate keywords were manually removed.

Bibliometric analysis was performed using VOSviewer [58] and the R package bibliometrix, as well as the corresponding web interface biblioshiny [59]. VOSviewer and bibliometrix are commonly used in bibliometric research [57,60,61]. For a detailed overview of the software tools used, see [62]. Accordingly, the selection of the used software tools can be justified since bibliometrix stands out due to the large number of different analysis methods available, while VOSviewer provides outstanding network visualizations.

In addition to the general performance analysis of the contributions of research constituents (e.g., articles, journals, institutions, authors, countries) to the considered field—using publication counts reflecting productivity and citations for measuring impact and influence [29]—the following science mapping methods were applied to identify the relationships between the research constituents: Co-authorship analysis focuses on revealing social interactions and relationships between authors, institutions (co-authorship institutions), and/or countries (co-authorship countries). Co-citation analysis identifies relationships among cited publications to reveal underlying themes in the research field. The co-occurrence of two publications in another publication’s reference list forms a connection between them [29]. As co-citation analysis focuses only on highly cited publications, bibliographic coupling was further applied to increase the visibility of recent and niche publications [29]. The analysis of bibliographic coupling shows which articles and authors are correlated with one another through multiple citations. A bibliographic coupling is present if two documents cite the same reference [63]. As it is recommended to use bibliographic coupling for specific timeframes [64], our analysis focused on the two recent periods, 2019–2020 and 2021–2022.

Scopus uses the indexed keywords (standardized keywords chosen by Scopus; take into account, e.g., synonyms, spellings, plurals) for the search; however, our keyword results are presented for the author keywords (keywords chosen by the authors of the documents). This decision was based on our objective of starting from a broad, exploratory point of view in order to make the best use of any potentially relevant information and then focusing on aspects seen as relevant from the perspective of the scientific community. In addition to the general analysis of the total keyword occurrence, a longitudinal analysis of the keywords’ occurrence per year was also performed. This analysis was intended to reveal trending topics of AI research in sports based on the author keywords, which are generally sufficient to conclude topical aspects in a research field, as they are connected to the content of the publication [65].

Furthermore, based on the author´s keywords, a thematic map of AI in sports was constructed. In general, the aim was to obtain insights into the current status of the field as well as to provide knowledge regarding the possible future development of research themes within the field. Therefore, the author´s keywords and their interconnections were clustered into themes. Characterization was performed through the property’s density (i.e., measurement of the cohesiveness among the nodes, or internal cohesion; vertical axis) and centrality (i.e., measurement of correlation between different topics, or external cohesion; horizontal axis) [66]. The more relationships a node shows with others in the network, the greater its centrality and importance. The density representing the cohesiveness among the nodes gives clues as to the capability of a research field to develop and sustain itself [67].

3. Results

3.1. General Performance Analysis

In terms of the various Scopus subject areas, “Computer Science” (n = 741) ranked in the first place, followed by “Engineering” (n = 517) and “Mathematics” (n = 242). The first publications were recorded in the year 1988 [68]. There was a steady increase in the annual number of publications, with the first major increase in 2012 and a second sharp and continuous increase from 2017 onwards (see Figure 2). Therefore, this research field can be thought of as being relatively young, especially in relation to the recent increase in publications. The context of the COVID-19 pandemic in the years 2020–2022 seems to have had little impact, with no visible drop in publications compared to other research fields, e.g., life sciences [69]. However, due in part to long publication times, it remains unclear whether the pandemic´s effects will become apparent with a shift in the next few years.

Article citation analysis is the most widely used method to study the impact of authors, journals, and articles, as it identifies the key works in a research area [70]. Overall, on average, every document had 9.53 citations based on the global citation counts using Scopus data. Along with the total increase in the number of published documents over time, a parallel increase in the average number of citations per year was observable (see Figure 2).

An anomaly was detected for the year 2007 in terms of the total citations per document. This can be explained by the appearance of the most cited article, “Checkers Is Solved” (Schaeffer et al., 2007) (see also:

Table 1. The top 20 most cited works. TC = total citations; normal. = normalized; = “hot papers” published in the last four years.

Title	Journal	First Author	Publication Year DOI		TC	TC per Year
Deep Learning Algorithms for Human Activity Recognition Using Mobile and Wearable Sensor Networks: State of the Art and Research Challenges	Expert Systems with Applications	H. F. Nweke	2018 10.1016/j.eswa.2018.03.056		404	80.8
Using Text Mining and Sentiment Analysis for Online Forums Hotspot Detection and Forecast	Decision Support Systems	N. Li	2010 10.1016/j.dss.2009.09.003		365	28.08
Learning Sparse Representations for Human Action Recognition	IEEE Transactions on Pattern Analysis and Machine Intelligence	T. Guha	2011 10.1109/ TPAMI.2011.253		296	26.91
Checkers Is Solved	Science	J. Schaeffer	2007 10.1126/science.1144079		277	17.31
A Deep Learning Approach to On-Node Sensor Data Analytics for Mobile or Wearable Devices	IEEE Journal of Biomedical and Health Informatics	D. Ravi	2016 10.1109/JBHI.2016.2633287		254	42.33
Learning to Select and Generalize Striking Movements in Robot Table Tennis	The International Journal of Robotics Research	K. Mülling	2013 10.1177/0278364912472380		241	24.1
Big Data and Tactical Analysis in Elite Soccer: Future Challenges and Opportunities for Sports Science	SpringerPlus	R. Rein	2016 10.1186/s40064-016-3108-2		205	29.29
Haptic-Feedback Smart Glove as a Creative Human–Machine Interface (HMI) for Virtual/Augmented Reality Applications	Science Advances	M. Zhu	2020 10.1126/sciadv.aaz8693		184	61.33
Recognizing Daily and Sports Activities in Two Open Source Machine Learning Environments Using Body-Worn Sensor Units	The Computer Journal	B. Barshan	2014 10.1093/comjnl/bxt075		179	17.9
Recognizing Human–Object Interactions in Still Images by Modeling the Mutual Context of Objects and Human Poses	IEEE Transactions on Pattern Analysis and Machine Intelligence	B. Yao	2012 10.1109/ TPAMI.2012.67		174	15.82
A Review of the Evolution of Vision-Based Motion Analysis and the Integration of Advanced Computer Vision Methods Towards Developing a Markerless System	Sports Medicine—Open	S. L. Colyer	2018 10.1186/s40798-018-0139-y		154	30.8
Deep Learning for Fall Detection: Three-Dimensional CNN Combined With LSTM on Video Kinematic Data	IEEE Journal of Biomedical and Health Informatics	N. Lu	2019 10.1109/JBHI.2018.2808281		138	34.5
Spatio-Temporal Analysis of Team Sports	ACM Computing Surveys	J. Gudmundsson	2018 10.1145/3054132		133	22.17
Learning to Track and Identify Players from Broadcast Sports Videos	IEEE Transactions on Pattern Analysis and Machine Intelligence	W.-L. Lu	2013 10.1109/ TPAMI.2012.242		132	13.2
Machine Learning for Activity Recognition: Hip versus Wrist Data	Physiological Measurement	S. G. Trost	2014 10.1088/0967-3334/35/11/2183		127	14.11
Real-Time Posture Reconstruction for Microsoft Kinect	IEEE Transactions on Cybernetics	H. P. H. Shum	2013 10.1109/TCYB.2013.2275945		109	10.9
Machine and Deep Learning for Sport-Specific Movement Recognition: A Systematic Review of Model Development and Performance	Journal of Sports Sciences	E. E. Cust	2019 10.1080/02640414.2018.1521769		104	26
A Machine Learning Framework for Sport Result Prediction	Applied Computing and Informatics	R. P. Bunker	2019 10.1016/j.aci.2017.09.005		97	24.25
Multiple Player Tracking in Sports Video: A Dual-Mode Two-Way Bayesian Inference Approach With Progressive Observation Modeling	IEEE Transactions on Image Processing	J. Xing	2010 10.1109/TIP.2010.2102045		97	8.08
Relationships Between Internal and External Training Load in Team-Sport Athletes: Evidence for an Individualized Approach	International Journal of Sports Physiology and Performance	J. D. Bartlett	2016 10.1123/ijspp.2015-0791		94	15.67

), which caused an atypical total citation count in relation to the early development of AI research in sports during that time. Checkers is a strategy board game that can be understood as a sport, similar to chess. In terms of citations per document, an unequal distribution was present, where the majority of works were only cited a few times. The proportion of papers with two or fewer citations was 53.33%. There are two potential reasons for this: the articles were published very recently, or they do not generate sufficient academic interest to be cited by the research community.

Overall, the reviewed documents were published in 428 different journals. The 15 most productive journals in terms of the number of published documents are presented in Figure 3. The top 10 journals are responsible for 28.64% of all published documents. The most productive journal was Computational Intelligence and Neuroscience (h-index: 61; 2021 impact factor: 3.64), followed by Sensors (h-index: 196; 2021 impact factor: 4.35) and IEEE Access (h-index: 158; 2021 impact factor: 4.34). It is notable that the top-ranked journals include not only journals with specific aims and scopes, but also those with a broader possible range (e.g., PLOS ONE). However, journals that focus on information technology and engineering comprise the majority. This indicates that AI in sports is interdisciplinary because most journals do not focus on sports. The Journal of Sport Science is only ranked 10th, which can be taken as a sign that AI is not the main topic in sports science at all.

3.2. Top 20 Overall Most Cited Works

The top 20 most cited works (see Table 1) account for 32.52% of the total citations. The most common theme in the top 20 most cited works is human activity or action recognition (e.g., [32,71,72,73,74,75]), followed by predictive systems (e.g., [76,77]) and systems for object or player tracking (e.g., [78,79,80]). Only one study investigated AI in the direct context of motion analysis [81]. Most studies developed an algorithm or a new methodological approach (e.g., [82,83,84]). The other studies focused on reviewing the literature [32,71,78,81,85]. The focus of the data was mainly on wearables (e.g., [71,73,75,84]), followed by visual data (e.g., [79,80,81,86]). Team sports were analyzed more frequently compared to other sports (e.g., [77,78,85]).

The publications that were a maximum of four years old and featured among the top 20 most cited publications in this context (referred to here as “hot papers”; see also: Table 1) could take on special importance in future research if they continue to be cited over the next few years.

3.3. Countries, Organizations, and Global Structures

AI in sports is a global research topic. Analysis of the total authorship data of each document showed that 128 countries contribute to the research on this topic. However, China leads in terms of the total number of publications by a considerable margin, followed by the USA and Germany (see

Table 2. Top 10 most productive countries based on the authors’ countries, total citations (TC) per country based on document citations, and mean number of authors per country based on the authors´ countries.

Countries	N Publications	TC	Mean Number of Authors
China	1244	1723	2.72
USA	591	1031	5.27
Germany	204	842	4.40
UK	170	671	4.39
Australia	169	815	4.54
South Korea	149	395	3.56
Japan	134	139	5.39
Spain	134	159	5.02
Canada	125	796	4.23
India	112	107	3.76

). On average, each publication has 3.65 co-authors. Regarding the mean number of authors from the most productive countries, China has the fewest authors per document. This can be explained by the fact that 27.33% of the publications with at least one Chinese author had only one author; in comparison to the USA and Germany—with 7.38% and 2.99% single authorship, respectively—this is a relatively high proportion. It should be noted that 21.15% of international co-authorship is also present; therefore, a single publication can be counted in several countries.

With 38 articles, the Hospital for Special Surgery (New York City, NY, USA) takes first place among the most productive organizations, followed by Victoria University (Melbourne, Australia) and Ghent University (Ghent, Belgium) (Figure 4a). Surprisingly, although China is the most productive country in terms of publication output, the leading Chinese university only ranks in sixth place (Beijing Sport University). This indicates that many institutions in China seem to be involved in the publications without any individual institution being particularly active. The aforementioned high proportion of single authorship in Chinese publications, without collaborations between institutions, also appears to be related to this.

It is notable that the trend toward multi-university publications has continued to increase, especially in recent years (Figure 4b). This could be interpreted as a sign that AI in sports has moved out of its initial niche status. An examination of the collaboration between organizations (Figure 5) shows that publications are often produced through cooperation between individual universities. There are larger associations of collaborating universities as well as islands formed through small collaboration networks. The resulting networks may be due to the need to bundle expertise, since individual universities mainly provide sporting or IT expertise, which must be brought together to process the research object in the sense of interdisciplinary or transdisciplinary research. According to the existing small, unconnected collaborative networks, networking still has great potential insofar as there is a fit in terms of the content and methodology of the research institutes.

Examination of the international collaboration, which indirectly depicts the connections of the authors from different countries, shows that five clusters of collaboration groups are present (Figure 6). The top three most productive countries (i.e., China, the USA, and Germany) are in different clusters. In the blue cluster, the most productive countries are China, the Republic of Korea, and India. In the yellow cluster, the USA is the most active country, followed by Spain and Canada. In the red cluster, the most productive countries are Germany, Turkey, and Japan. In the green cluster, the United Kingdom is the leader, followed by Australia and Slovenia. It is striking that geographic proximity is a strong explanatory factor but is not necessarily related to cluster affiliation. For example, Germany collaborates strongly with the nearby countries of the Netherlands and Belgium, but also with Japan. The same applies to the cooperation between the USA and Spain.

Substantial collaboration is visible between the different clusters. However, the collaboration between China and the other main clusters is lower than the collaboration of the remaining clusters. Using the terms of the World Bank for the classification of the countries [87], it seems that high-income countries and middle-income countries collaborate more within each cluster.

3.4. Authors and Social Structures

The research field of AI in sports is characterized by a large number of authors. In total, 3411 different authors published their works during the period of time investigated in this review. Only 188 authors of single-authored publications were present. On average, each publication shows 3.65 co-authors, and 21.15% of the papers are written by international author groups. The majority of the most productive authors are located in China. Due to the relatively high level of correspondence between names according to the country-specific characteristics in China [88], the most productive authors are not presented here so as to rule out any possible errors or injustice since authors were not identified on the basis of a clearly identifiable ID (e.g., ORCID ID) but on the basis of their first and last names which, due to the abovementioned situation, may have led to an incorrect merging of authors with the same first and last names. However, as it is reported that co-citation analysis is relatively immune to author name ambiguity issues [89], the corresponding mapping was performed to reveal the social structures of the authors (Figure 7).

In the center of Figure 7, we can see a large number of strongly networked and collaborating authors who come from China in particular. In the periphery, however, some authors do not show any networking, or small groups are formed that are not further connected to other authors. Unused potential in networks could be the reason for these islands’ formation. On the other hand, it also seems conceivable that the small, separate groups are pursuing very specific research areas, and, therefore, there are no networking opportunities with researchers from other subject areas with methodological or content-related similarities.

3.5. Co-Citation Analysis and Bibliographic Coupling

The co-citation analysis revealed the presence of six clusters (see Figure 8). Two clusters were not connected with other clusters (shown in green and blue). The overall most cited publications in the clusters were those of Breiman (2001) with the random forest algorithm [90] and Hochreiter and Schmidhuber (1997) with long short-term memory (LSTM) [91].

Focusing on the main topics of each cluster, it is hard to find significant differences. The first cluster, shown in red, could be grouped under primary algorithmic and machine learning literature—most notably the paper of Breiman (2001) [90].

Many papers in the purple cluster deal with deep convolutional neural networks as well as human action or image recognition. Furthermore, the works of Zhang et al. (2021) [92] and Cai et al. (2021) [93] (green cluster), which are not connected to other clusters, can be attributed to image segmentation. The publications in the orange cluster (e.g., LeCun et al., 2015) [94] mainly deal with the themes of wearables and deep learning.

The results of the bibliographic coupling are the following: A look at the most frequently cited publications in the recent period shows that only a few thematically distinguishable clusters can be formed for the years 2021 and 2022. With a minimum of three citations, the bibliometric analysis obtained 14 clusters. The first cluster contains 17 items, mainly concerning the themes of machine learning applications in the context of sports and health, as well as decision-making; in this cluster, the most representative paper was published by Cao et al. (2021) [95]. The second cluster includes 16 items and could be summarized under the topic of motion recognition; Cui et al. (2021) [96] published the most cited paper in this cluster. The third cluster consists of 14 items; most of these publications address the topic of AI in medicine; the study of Ramkumar et al. (2021) [97] was cited the most, with 20 citations.

Thirteen clusters were obtained via bibliometric coupling analysis for 2019 and 2020. The first cluster contains 21 items, most of which can be assigned to the topic of team sports; the most cited paper in this cluster was published by Bunker et al. (2019) [76]. The second cluster, with 16 items, can be summarized under the topic of sports recognition or recognition methods; the most representative paper in this cluster was published by Cust et al. (2019) [32]. The third and fourth clusters pertain to research with detections/tracking and big data, as well as machine learning; the most cited papers in the third cluster were authored by Yoon et al. (2019) [98] and Minhas et al. (2019) [99]; in the fourth cluster, the review by Claudino et al. (2019) [20] was cited the most.

3.6. Keywords/Themes: General Trends and Longitudinal Evolution

The overall occurrence of the 200 most frequently occurring author keywords is shown in Figure 9. The top four keywords map the search terms used for the Scopus search. It is notable that “machine learning” is the most frequently used term; the subarea “deep learning” is second, followed by “artificial intelligence”. A possible reason for this order is the more precise and specific scientific language [100] that specifically names the subfields instead of speaking of AI in a general way. The fact that the keyword “sports” only came in fourth place is a sign that the authors do not use sport as a keyword, in contrast to Scopus. This could be explained by two possible mechanisms: on the one hand, it is possible that authors do not use sport itself as a key term, referring instead to football, biomechanics, injury prevention, etc., as subsets of sport. Thus, sport(s) would be too general or too obvious as a keyword. On the other hand, it is also possible that authors do not see to the relevance of sports in their research compared to the AI topics.

The 25 most frequently occurring author keywords can be manually grouped according to their meaning. The keywords “artificial neural networks”, “convolutional neural networks”, and “support vector machine” can be subsumed under the term algorithms (a). The keywords “inertial measuring units”, “wearable sensors”, “accelerometer”, “wearables”, and “sensors” can be assigned to measuring methods or tools (b). It is striking that all of these methods can also be used outside of laboratory settings in the field and make huge amounts of data relatively easily accessible, which favors the emergence of “big data” characteristics [101]. Specific AI tasks (c) are mapped with the terms “classification”, “activity recognition”, “action recognition”, and “prediction”. Different research fields (d) are mapped with the terms “sports analytics”, “computer vision”, “data mining”, “internet of things”, and “biomechanics”. Of the 25 most frequently occurring author keywords, only “soccer” and “football” name specific sports categories (e). A closer analysis shows that football is used synonymously with soccer in most of the publications included, in accordance with country-specific conventions. Other sports were also mentioned as specific keywords; however, their frequency was significantly lower. Thus, it can be seen that the sport of soccer/football is very dominant, which can also be explained by the significant financial interests behind this sport [102].

Overall, according to the meaning of the dominant keywords, a high overlap with the contents of the most cited publications listed in Table 1 could be observed. It is notable that when the topics that are covered by the keywords or the most frequently cited publications are assigned to sub-disciplines of sports science, their primary focus is on sports informatics, training science, biomechanics, and sports medicine. Research activities in the fields of sports pedagogy, sports sociology, and sports economics were not evident from our analyses.

Figure 10 presents the corresponding thematic map of AI research in sports. The themes are divided into four quadrants by the horizontal and vertical dashed lines: motor/driving themes (upper right; Q2), basic/underlying themes (lower right; Q4), niche/very specialized themes (upper left; Q1), and disappearing or emerging themes (lower left; Q3). Themes in Q4—such as in general AI (inclusive of its subareas), different specific sports, and subfields of the AI pipeline—are basic, foundational themes that are very important for the field’s general development, characterized by high centrality but low density. Various measuring technologies (e.g., accelerometers, inertial sensors, gyroscopes) can be subsumed under the driving themes of Q2; this makes sense because they often provide complex data, making AI techniques more and more relevant due to static inference methods being limited in those contexts [103]. Looking at Q3, including the subsequent results of trending topics, sports medicine can be identified as an emerging theme. The themes of Q1 (e.g., “simulation”, “hidden Markov model”) are highly specialized and rare, creating internal bonds, as indicated by their low centrality but high density. The theme “performance analysis” is located between all quadrants, indicating that some aspects are basic and necessary for file development, some are emerging or declining, some are fundamental drivers, and some are highly specialized niche themes.

Overall, it is notable that the manual grouping of the most frequently occurring keywords fits with the data-driven clustering of the thematic map to a high degree; this also speaks to the coherence of the obtained solution.

Figure 11 shows the evolution of author keywords, which has attracted particular attention in recent years. In addition to the neural networks that have been in use for a long time, as well as the support vector machine (see Figure 9), the random forest algorithm has recently been given special attention. This is consistent with the results of the co-citation analysis, where a large cluster was found around the work of Breiman (2001) [90]. According to methodological aspects in the context of the AI workflow, attention has also recently been paid to “feature extraction” and “transfer learning”.

The biomechanical aspects of “pose estimation” and “kinematics” have only recently attracted particular attention. Furthermore, the field of “sports medicine” and the matching topic of injury prevention are mentioned frequently—especially recently. The sport of running and e-sports have also received attention in connection with AI in recent years.

Unsurprisingly, “COVID-19” also attracted attention in 2021 and 2022, during the period when the world was still struggling deeply with the pandemic. The future infection situation will probably be decisive for the progression or decline of this topic. Another aspect that has emerged—especially in 2022—is the ethical discussion of AI in sports. Please refer to Figure 11 for information on further trend topics.

At this point, it should be noted that our analysis probably does not depict topics that have arisen very recently. For example, few recent publications have addressed the black-box character of AI models [104,105], which does not comply with the General Data Protection Regulation (GDPR) [106], using explainable AI. However, this theme does not appear as a trend topic using the author´s keywords at present, although it may be visible in a bibliometric analysis in the next few years.

4. Limitations and Further Research

The bibliometric research situation with regard to the object of AI in sports is currently too limited to carry out intensive discussions of the results, including comparisons with other bibliometric works. This work should therefore form a first step towards the bibliometric analysis of the research subject and stimulate comparisons between different studies.

In addition to the aforementioned advantages of our research aims, the bibliometric methodology is also associated with some inherent limitations. The analysis is based on technical decisions, e.g., the selection of English as the search language or the focus on the highest-quality papers (i.e., articles and reviews), which may have resulted in the exclusion of other research papers. Therefore, future research should also consider the inclusion of other document types, e.g., conference papers or preprints, as these may provide better insights into the very latest research themes because their publication times are often shorter. Articles in anthologies or monographs could also be interesting since social science sub-disciplines of sport use them as publication opportunities. Furthermore, it has been shown that different databases (e.g., Google Scholar, Scopus, Web of Science) have certain differences in their search results [107]. Therefore, a multisource comparison and mutual supplementation of the different databases could result in a broader literature overview. Reported issues with the ambiguity of author names could be addressed in the future by using ORCID IDs to prevent potential incorrect merging of authors with the same first and last names [89].

It is important to disclose that we only used the titles and keywords and not the abstracts for the literature search in order to ensure that sports are the direct focus of the research and not a side issue. Although we were able to focus on literature highly related to sports with this approach, it should not go unnoticed that this could also have led to the exclusion of other potentially relevant literature. In addition, the used search terms could be expanded to include further aspects that are associated with the term “sports”, in order to identify publications that deal with topics that can be assigned to sports without mentioning the term itself (e.g., physical activity, exercise, training). At this point, the problem of defining the term sports should not go unmentioned. One should be aware that, especially in everyday life, there are a wide variety of (individual and cultural) ideas and definitions of sports—and this applies not only to the term “sports”, but also to terms that are often associated with sports, such as the term “training” [108,109]. For example, a central point of discussion is whether e-sports constitute sports [110]. This problem was dealt with in this study in such a way that, in simplified terms, it was assumed that publications containing the term “sport*” could be assigned to the research interest of sports science. Therefore, publications from the field of e-sports or—in some cases—from the field of board games (e.g., “Checkers is Solved” [82]) were also included. It should be noted that, depending on the individual or context-specific understanding of the definition of sports, these publications might also be classified as unrelated to sports.

5. Conclusions

The present study provides a detailed analysis of the leading journals, authors, institutions, keywords, and research themes associated with interconnections between the fields of AI and sports. However, it seems that AI has not yet taken a leading role in sports science but has been gaining more and more importance recently, and sport is an area of application for AI. Our key findings are as follows: (a) The literature and research interest concerning AI and its subcategories is growing exponentially, with considerable potential for further growth in the coming years. (b) The top 20 most cited works account for 32.52% of the total citations, including the presence of “hot papers”, which are still relatively young and have already received considerable numbers of citations. (c) The top 10 journals are responsible for 28.64% of all published documents and indicate a trend of interdisciplinary research. (d) In addition to strong collaborative relationships, delimited small collaboration networks of individual institutions are identifiable. (e) The three most productive countries are China, the USA, and Germany; collaborations appear to be taking place more within the clusters of high- and middle-income countries. (f) Different research themes can be characterized using author keywords, with potential current trend topics including the fields of biomechanics, injury prevention or prediction, new algorithms, and learning approaches. AI research activities in the fields of sports pedagogy, sports sociology, and sports economics seem to have played a subordinate role so far.

These findings provide a better understanding of the research situation regarding the use of AI in sports and may aid researchers in identifying currently relevant research topics, as well as research gaps, through revealing fields of research with much evidence and those that have only been considered marginally thus far. Furthermore, they may be helpful in improving the obviously necessary networking in the context of interdisciplinary and transdisciplinary research in order to combine competencies from different disciplinary approaches. Hence, this work could be a starting point for systematically initiating stronger networking between authors and their countries/institutions, which might offer yet-untapped potential in their interconnection according to the present study’s results.

Today, AI seems to play an important role not only in sports research but also in sports practice. Numerous companies (e.g., Apple, Microsoft, Enduco, iDIERS, Aaptiv) are advertising commercial offers that use AI technologies and, thus, have partially opened up areas that research has not yet sufficiently investigated. Therefore, it seems important that research does not lag behind in questioning these economic interests with scientific evidence.