Artificial intelligence in sports: A systematic review of performance and injury prevention

Protocol

This systematic review was performed following the guidelines established by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) ²⁵ to ensure transparency, comprehensiveness, and reproducibility. In this review, we utilised the PICOS framework in our analysis, which includes: participants (specific or general sports activity), intervention (data analyses using Machine learning), Comparators (not applicable), outcomes (injury prevention or performance enhancement) and the study design as data collected from participants in any sports events. The review protocol was registered with the International Prospective Register of Systematic Reviews (PROSPERO; Registration ID: [CRD420251129702], registered on [19 August 2025]). The registration was completed retrospectively after the review process had commenced; however, the review objectives, eligibility criteria, and search strategy had been predefined prior to study selection and remained unchanged throughout the review.

Search strategy

A comprehensive search was done in IEEE Xplore, ScienceDirect, Web of Science, and PubMed to identify studies that were exploring the use of artificial intelligence (AI) or machine learning (ML) in sports performance and preventing injuries. The search was performed between January 2014 and January 2025. Only peer-reviewed journal articles in the English language were included. Reference lists of included studies were manually screened to identify additional relevant publications (Table 1).

OPEN IN VIEWER

Table 1.

Database and initial search results per source.

Database	Search string (Boolean keywords)	Articles
IEEE Xplore	(“artificial intelligence” OR “machine learning”) AND (“sports performance” OR “athletic performance”) AND (“injury prevention” OR “injury detection”)	257
Science Direct	(“artificial intelligence” OR “machine learning”) AND (“sports performance” OR “athletic performance”) AND (“injury prevention” OR “injury detection”)	181
PubMed	(“artificial intelligence” OR “machine learning”) AND (“sports performance” OR “athletic performance”) AND (“injury prevention” OR “injury detection”)	35
Web of Science	(“artificial intelligence” OR “machine learning”) AND (“sports performance” OR “athletic performance”) AND (“injury prevention” OR “injury detection”)	98

Eligibility criteria

The included studies in the present review followed the subsequent inclusion criteria: (1) original research articles published in peer-reviewed journals; (2) studies involving human participants who were athletes at any competition level (elite, sub-elite, or recreational); (3) studies reporting the application or formal evaluation of an artificial intelligence (AI) or machine learning (ML) model in a sports context with a primary focus on performance enhancement and/or injury prevention; (4) studies reporting at least one measurable outcome relevant to sports performance or injury prevention, including model outputs, evaluation metrics, or performance-related indicators; (5) studies published in English; and (6) studies published between January 2014 and January 2025.

The exclusion criteria were as follows: (1) reviews, meta-analyses, conference abstracts, theses, editorials, white papers, or case studies; (2) studies focusing solely on perceptions or attitudes towards AI/ML without actual application in performance or injury prevention; (3) studies reporting only algorithm or model development without a sport-specific application; (4) studies conducted on non-human subjects; and (5) studies addressing sports domains outside of performance or injury prevention (e.g., fan engagement, sports marketing, or broadcasting).

Until January 2025, the search was limited to original articles that were published online. Titles and abstracts were initially screened and excluded based on the predetermined criteria. The final inclusion or exclusion decision was made after a critical review of full-text articles. Any disagreements between the two reviewers were resolved through discussion or, when necessary, by consultation with a third reviewer. Additional relevant secondary sources were also subjected to the same screening procedure.

Study selection

The references retrieved were imported into Zotero reference management software to organise them. The same records were located and deleted with the help of the Rayyan systematic review platform. Two reviewers independently screened the titles and abstracts of all the records searched according to the pre-specified inclusion criteria. Those studies which were obviously not meeting the inclusion criteria were very briefly excluded at this point. The full text of the remaining articles was then obtained and assessed for eligibility by the same two reviewers. The detailed study selection process mentioned in the PRISMA flow diagram concisely summarises the number of studies identified, screened, excluded and included (Figure 1).

OPEN IN VIEWER

Figure 1.

Flow diagram of study identification, screening, and inclusion.

Data extraction

Data were independently extracted by two reviewers with the aid of a pre-defined spreadsheet that was developed specifically for this review. The extraction protocol was pilot-tested on a small number of studies that were eligible for inclusion to check for clarity, completeness, and consistency before use for all included articles. For each study, the following information was recorded: author(s) and year of publication; country of origin; study design; variables examined; instrument(s) used; statistical method(s) applied; sport or domain; participant age; type of machine learning (ML) approach; specific ML techniques used; inclusion criteria; exclusion criteria; sample size; and key findings. The abstracted information manifested the context-specific methodological feature as well as the substantive results of the individual study, enabling a detailed synthesis of how AI/ML approaches have been applied in sports performance and injury prevention. Disagreements between the data extraction of the two reviewers were resolved by discussion; if consensus could not be reached, the opinion of a third reviewer was consulted. In addition, the sample size variable was further categorised to distinguish between dataset volume (e.g., number of frames, images, or segments) and human participant samples to ensure clarity in reporting.

Quality assessment

The methodological risk of bias of the identified studies was evaluated using the design-specific, validated instruments, in line with the contemporary methodological guidance in conducting systematic reviews with heterogeneous designs of studies. For studies using machine learning-based prediction models, the Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence (PROBAST + AI) ²⁶ was used. PROBAST + AI assesses risk of bias and applicability in four areas, including participants and data source, predictors, outcome, and analysis, with a specific focus on overfitting, model validation, data handling and applicability of the prediction model. In addition, the two studies employing a randomized controlled trial (RCT) design were further appraised using the revised Cochrane Risk of Bias tool (RoB 2) in order to capture trial-specific sources of bias that are not fully addressed by prediction model appraisal alone, the Cochrane Risk of Bias tool (RoB 2) was revised and applied to evaluate potential bias across five domains: the randomisation process, non-compliance with the intended interventions, lack of outcome data, outcome measurement, and the choice of the reported outcome. ²⁷ All risk-of-bias evaluations were performed by two reviewers. Any differences were resolved through discussion or consultation with a third reviewer. Assessment outcomes are reported in terms of domain-level judgments according to the criteria of each assessment tool and they are summarised using traffic light plots Figure 2a and 2b, Figure 3a and 3b.

OPEN IN VIEWER

Figure 2.

(a) Traffic light plot summarizing the PROBAST + AI assessment of included AI/ML-based prediction model studies. (b) Summary plot of PROBAST + AI assessment showing the proportion of studies with low risk, some concerns, and high risk across the four domains.

OPEN IN VIEWER

Figure 3.

(a) Traffic light plot summarizing the risk of bias assessment of the included randomized controlled trials using the Cochrane Risk of Bias tool (RoB 2) across five domains. (b) Summary plot of the risk of bias assessment of the included randomized controlled trials using the Cochrane Risk of Bias Tool (RoB 2). The plot shows the proportion of studies judged as low risk and some concerns across the five domains.

Data synthesis

As sport domains, populations of participants, AI/ML techniques, sources of data, and outcome measures were very heterogeneous, no meta-analysis was implemented. In doing so, a narrative synthesis was conducted. Studies were characterised initially by their primary application, sports performance enhancement or injury prevention and then by sport domain and type of AI/ML technique (examples include supervised learning, unsupervised learning, deep learning, ensemble methods, unsupervised clustering methods, and Hybrid Approaches). The characteristics of the studies, methodological approaches and the reported outcome measures (e.g., accuracy, precision, recall, F1-score, RMSE) are compared descriptively in order to draw similarities, methodological strengths, and research gaps relevant to the use of AI/ML in sports performance and injury prevention.

Study selection

The initial database search identified 571 records, comprising IEEE Xplore (n = 257), Science Direct (n = 181), Web of Science (n = 98), and PubMed (n = 35). After removing 34 duplicates, 537 records were screened by title and abstract, with 513 excluded for not meeting our inclusion criteria. Then, 24 full-text articles were retrieved for eligibility assessment. This study excluded 8 articles because they lacked a machine learning (ML) component (n = 3), an inappropriate study design (n = 4), or a lack of relevance to sports application (n = 3). 14 studies met all inclusion criteria and were included in the final analysis.

Study characteristics

The studies reviewed were published between 2014 and 2025 and represented a global distribution. China is contributing the largest share (n = 8), followed by the USA (n = 1), Nigeria (n = 1), Australia (n = 1), Turkey (n = 1), and one collaborative study involving China, Japan, and the USA. The sports covered in these studies verified: football (n = 4), basketball (n = 2), handball (n = 1), tennis (n = 1), and multi-sports or general athletic applications (n = 5).

The majority of the included studies employed laboratory-based or experimental designs (n = 10), of which two randomised controlled trials (RCTs. The remaining studies comprised observational designs (n = 2) and modelling-based approaches (n = 2). Sample sizes varied widely, ranging from small-scale model development datasets with as few as (n = 17 participants) to large datasets capturing over three million motion frames, with participants’ ages spanning from youth athletes to adults aged 60–75 years. Data sources were diverse and included video footage, motion capture systems, wearable sensors, Inertial Measurement Units (IMUs), infrared thermal imaging, event stream datasets, and physiological monitoring devices.

The studies employed a wide range of machine learning approaches, with deep learning models such as Convolutional Neural Networks (CNN), Backpropagation Neural Networks (BPNNs), Dynamic Graph Convolutional Networks (DGCNs), and U-Net architectures being frequently and widely used for image-and video-based analysis. Supervised learning algorithms (Support Vector Machines, Random Forest, k-Nearest Neighbors, Gradient Boosting). Often complemented by ensemble methods, including (LightGBM, XGBoost, adaBoost), were also common, alongside occasional applications of unsupervised clustering (DBSCAN). Reported outcome measures included classification accuracy, precision, recall, F1-score, mean squared error (MSE), area under the receiver operating characteristic curve (AUC), and regression coefficients (Table 2).

OPEN IN VIEWER

Table 2.

Study characteristics of the included studies.

Author(s), year	Country	Study design	Sport / Population	Sample size	ML type	ML techniques used	Variables / Measures	Key findings
Datasets-Based studies
Zhang et al. (2025) 28	China	Experimental (AI model evaluation)	General movement datasets (Human3.6 M, MPI-INF-3DHP)	H3.6M: ∼3.7 M frames; MPI: ∼13 K frames	DL	DGCN, STIA, VLTE, Embodied AI	Pose estimation metrics (MPJPE, PCK, AUC)	Achieved MPJPE 18.9 mm (H3.6 M), 29.1 mm (MPI); PCK 98%; demonstrated robustness and computational efficiency.
Kusmakar et al. (2020) 29	Australia	Observational	MLS Soccer matches	13 matches (241 segments) + validation (42,860 segments)	Supervised	SVM	Possession events, entropy indices	Achieved 75.2% segment prediction accuracy; match outcomes predicted with 81.9% accuracy.
Liao & Fu (2025) 30	China	Experimental + model development	Youth football	28,000 training images, 3533 validation	DL	CNN, Heatmap Regression	Keypoint detection accuracy	Model achieved 94.6% accuracy, AP 84.6%; effective in school football programmes.
W. Liu et al. (2022) 31	China	Experimental	Men's doubles tennis	2 matches (∼1 h each)	Unsupervised & Supervised	CenterNet, DBSCAN	Player detection, behavioural anomalies	DBSCAN outperformed CenterNet; R = 0.982, MAE = 0.013; improved decision-making fairness.
Human Participant-Based Studies
Cao et al. (2023) 32	China	AI-enhanced RCT	Football (aged 8–13 years)	116 (Exp: 60; Ctrl: 56)	DL	BPNN, PCA, SVM, CNN	Strength, speed, sensitivity, posture	Functional strength training improved strength and sensitivity; kicking improved by 49.73%; accurate posture detection achieved.
Zou (2025) 33	China	Experimental (algorithm comparison)	Multi-sport professional athletes	Not reported	Supervised, Ensemble, DL	SVM, RF, RNN, GBM	Physiological, biomechanical, recovery outcomes	RNN achieved 90% prediction accuracy; AI models outperformed traditional injury risk methods.
X. Liu et al. (2025) 34	China	Application-based experimental	Multi-sport college athletes	50	DL	U-Net, DTW, SVD, Kalman filter	HR, HRV, muscle heat distribution	AI-integrated thermal imaging improved fatigue and injury detection; enabled real-time training feedback.
L. Li et al. (2025) 35	China	AI-based injury prevention modelling	Competitive aerobics	140 surveyed; 3 modelled athletes	DL, Supervised	KO-DNN, DNN, RF, AdaBoost, XGBoost, LightGBM	Thermal distribution patterns, injury risk	KO-DNN accuracy 86.67%; effectively predicted injury risk and recommended personalised training.
Chellamuthu et al. (2025) 36	India	Observational (retrospective injury data)	University-level Football, Cricket, Basketball	120	Supervised	KNN (best), RF, AdaBoost, ROS	Psychological, neuromuscular, asymmetry measures	KNN + ROS achieved AUC 0.87, TPR 100%; ankle injuries most prevalent (52.6%).
Adetiba et al. (2017) 37	Nigeria	Experimental	Football & Basketball players and controls	40 participants; 400 ECG vectors	DL	ANN (Levenberg–Marquardt, SCG)	ECG statistical features (mean, variance, etc.)	Classification accuracy 98.02%, sensitivity 98.06%, specificity 99.35%; supports early detection of SCD risk.
Yang & Xu (2022) 38	China	Experimental (video-based CNN)	Basketball	50	DL	CNN, Adaptive NN + Genetic Algorithm	Video-based injury detection	Achieved 90.4% accuracy; region detection contributed more than denoising to performance.
Oytun et al. (2020) 39	Turkey (N. Cyprus)	Experimental	Female handball (Super League)	59	Supervised	LR, DT, SVR, BP, RBFNN, LSTM	Anthropometric, physiological performance tests	RBFNN achieved R2 = 0.86–0.97; age and body composition predicted agility and endurance.
Guo et al. (2024) 40	China/Japan/USA	RCT, double-blind	Older adults (60–75 yrs) with sarcopenia	93	Supervised	LightGBM, XGBoost, RF, LR, etc.	Grip strength, sarcopenia metrics	Hybrid training (strength + Tai Chi) reversed sarcopenia in 45.5% (TCSG); LightGBM accuracy 88.4%.
Zhang et al. (2014) 41	USA	Lab-based experimental	Healthy adults, walking under fatigue	17	Supervised	SVM (linear, polynomial, RBF)	Gait kinematics, acceleration, jerk cost	Fatigue classification accuracy up to 97% (intra-subject); fatigue significantly altered gait parameters.

Note: (ML = Machine Learning; DL = Deep Learning; SVM = Support Vector Machine; ANN = Artificial Neural Network; CNN = Convolutional Neural Network; PCA = Principal Component Analysis; RF = Random Forest; GBM = Gradient Boosting Machine; RNN = Recurrent Neural Network; KO-DNN = Knowledge-Oriented Deep Neural Network; ROS = Random Over-Sampling).

Methodological quality

The included studies were re-evaluated using design-specific risk-of-bias tools based on study type. Studies involving machine learning–based prediction models were assessed using PROBAST, while randomised controlled trials were assessed using the Cochrane Risk of Bias tool (RoB 2). Overall, the included studies raised methodological concerns, particularly regarding participant selection, small sample sizes, limited validation procedures, and incomplete reporting of analytical methods. Nevertheless, most studies provided a clear description of their objectives, input variables, and the artificial intelligence or machine learning methods used, thereby supporting their relevance to the review.

AI application for sports performance enhancement

Seven studies focused on performance-related outcomes through ML integration. These investigations report a range of objectives, including pose estimation, skill assessment, tactical analysis, and match outcome prediction. For example, utilised Support Vector Machine (SVM) to classify IUM-derived gait data to distinguish fatigue from non-fatigue states, achieving intra-subject accuracy up to 97%. In the context of football training ³² reported a 49.73% improvement in kicking accuracy following functional strength training, with posture recognition via Backpropagation Neural Network (BPNN) model enhancing movement analysis efficiency.

In team sport analytics ²⁹ introduced entropy-based performance metrics for soccer, achieving 75.2% accuracy in segment prediction and an AUC of 0.84 during external validation. Similarly, in handball ³⁹ showed that a radial basis function neural network achieved R2 values ranging from 0.86 to 0.97 for performance prediction using physiological and anthropometric data. Overall, performance-oriented studies reported accuracy rates above 75% with deep learning and hybrid ensemble models providing the highest predictive capacities.

AI application for sports injury prediction and prevention

Seven studies focused on injury risk prediction, detection, and rehabilitation optimisation. These often control physiological monitoring, historical injury records, or biomechanical profiling data. Zou (2025) ³³ compared several AI algorithms for prediction in professional athletes, reporting that recurrent neural networks achieved 90% accuracy, outperforming SVM, Random Forest, and gradient boosting models. Similarly ⁴² applied kO-DNN models with infrared thermal imaging in aerobics athletes, achieving 86.67% accuracy in predicting injury risk based on thermal anomaly patterns. Chellamuthu et al. (2025) ³⁶ integrated neuromuscular, psychological, and biomechanical variables within a KNN + ROS framework, yielding 100% true positive rate and an AUC of 0.87 for injury classification in university athletes. Likewise ³⁸ identified video-based CNN models to detect basketball injury events with 9.4% accuracy. AI-powered wearable devices and thermal imaging systems, as revealed. ³⁴ Further showcased the capacity for real-time fatigue and injury detection, enabling proactive interventions. Across these injury-focused studies highlight that AI-driven approaches consistently surpass traditional analytics approaches in accuracy, precision, and sensitivity (Table 3 and 4).

OPEN IN VIEWER

Table 3.

Studies on AI for sports performance enhancement.

Authors (year)	Participants	Data acquisition technology	Variables	Aim of prediction	Type of machine learning	Practical application from predicting
Zhang et al. (2025) 28	Adults (dataset frames)	Pose datasets (Human3.6 M, MPI-INF-3DHP), environmental sensors	Pose estimation metrics (MPJPE, PCK, AUC)	Accurate and robust human pose estimation	Deep learning (DGCN, STIA, VLTE)	Enhanced motion tracking for athletic performance
Zhang et al. (2014) 41	17 healthy adults	IMU sensors, Force plates, Motion capture	Gait kinematics, fatigue state variables	Classify fatigue vs non-fatigue gait	SVM (linear, polynomial, RBF)	Monitoring fatigue and fall risk in athletes
Cao et al. (2023) 32	116 youth footballers (8–13y)	BPNN posture detection, PCA, SVM	Strength, speed, sensitivity, posture accuracy	Improve football skill performance	Deep learning (BPNN, CNN)	Optimising training with AI posture recognition
Kusmakar et al. (2020) 29	Soccer match data	Event stream datasets, public soccer logs	Possession events, entropy metrics	Predict match segment outcomes	SVM	Improved tactical decision-making in soccer
Oytun et al. (2020) 39	59 female handball players	Skinfold calipers, force plates, sprint timing	Physiological & anthropometric variables	Predict performance metrics	RBFNN, SVR, LR, DT	Talent identification and training optimisation
Guo et al. (2024) 40	93 older adults (60–75y)	Grip strength, CT imaging	Muscle mass, muscle fat area, functional tests	Predict sarcopenia reversal	Ensemble ML (LightGBM, RF, XGBoost)	Optimising training for muscle health
Liao & Fu, (2025) 30	Youth football video data	Video + wearable sensors	Keypoint detection, physiological data	Improve training analysis	CNN (heatmap regression)	Automated technical feedback in football
W. Liu et al. (2022) 31	Tennis match videos	Drones, CenterNet, DBSCAN	Net passes, abnormal behaviour metrics	Enhance match fairness & analysis	DL + Unsupervised clustering	Reduced referee error and improved fairness

Note: AI = Artificial Intelligence; AUC = Area Under the Curve; BPNN = Back Propagation Neural Network; CNN = Convolutional Neural Network; CT = Computed Tomography; DBSCAN = Density-Based Spatial Clustering of Applications with Noise; DGCN = Dynamic Graph Convolutional Network; DL = Deep Learning; DT = Decision Tree; IMU = Inertial Measurement Unit; LightGBM = Light Gradient Boosting Machine; LR = Linear Regression; MPJPE = Mean Per Joint Position Error; PCA = Principal Component Analysis; PCK = Percentage of Correct Keypoints; RF = Random Forest; RBFNN = Radial Basis Function Neural Network; STIA = Spatio-Temporal Information Aggregation; SVM = Support Vector Machine; SVR = Support Vector Regression; VLTE = View-invariant Latent Temporal Encoding; XGBoost = Extreme Gradient Boosting.

OPEN IN VIEWER

Table 4.

Studies on AI on sports injury prevention and prediction.

Authors (Year)	Participants	Data acquisition technology	Variables	Aim of prediction	Type of machine learning	Practical application from predicting
Adetiba et al. (2017) 37	40 participants	Arduino ECG system	Heart condition classes, ECG features	Early detection of cardiac risk	Deep learning (ANN)	Prevent sudden cardiac death in athletes
Zou, (2025) 33	Professional athletes	Wearables, motion capture, medical records	Physiological & biomechanical data	Predict injuries & customise rehab	RNN, SVM, RF, GBM	Proactive injury prevention
X. Liu et al. (2025) 34	50 athletes (multi-sport)	Infrared thermal sensors, PPG	Heart rate, muscle heat, HRV	Detect fatigue & injury	Deep learning (U-Net)	Real-time feedback for injury risk
L. Li et al. (2025) 35	140 surveyed aerobics athletes	Infrared thermal imaging, sensors	Body temperature distribution	Predict injury risk	KO-DNN, RF, XGBoost	Personalised training recommendations
Chellamuthu et al. (2025) 36	120 male university athletes	Balance boards, Y Balance Test, questionnaires	Psychological, neuromuscular, biomechanical variables	Classify injury occurrence	KNN + ROS	Screening tool for injury prevention programs
Yang & Xu, (2022) 38	50 basketball players	Video injury detection (CNN)	Image/video frames	Detect basketball injuries	CNN with GA	Automated injury detection from match footage

Note: ANN = Artificial Neural Network; CNN = Convolutional Neural Network; GA = Genetic Algorithm; GBM = Gradient Boosting Machine; HRV = Heart Rate Variability; KNN = K-Nearest Neighbors; KO-DNN = Knowledge-Oriented Deep Neural Network; PPG = Photoplethysmography; RF = Random Forest; RNN = Recurrent Neural Network; ROS = Random Over-Sampling; SVM = Support Vector Machine; U-Net = CNN architecture for image segmentation.

AI for sports performance enhancement

Applications of AI to enhance performance have primarily been on skill analysis, tactical optimisation, and biomechanical feedback. Studies have been done on the use of BPNN-based posture recognition system and have shown considerable technical progress including a reported 49.73% increase in football kicking accuracy. ³² Moreover, predictive model based on entropy has also been applied in the realisation of patterns of possession in soccer and prediction of match results, assisting in the process of data-based tactical planning. ²⁹ These findings are consistent with the wider body of sports analytics research that emphasises the potential use of AI-based motion tracking to accelerate skill acquisition and reduce the use of objective measures. ⁴³ Nevertheless, sports performance is multifactorial in nature and it is a complex interplay of biomechanical, physiological, and neuromuscular systems. Existing AI systems are typically based on predominantly kinematic or outcome-based data and they have limited incorporation of underlying neuromuscular processes. Performance analysis is an advanced field of the need to include electromyography (EMG) signals and musculoskeletal modelling to obtain muscle activation patterns, joint kinetics, and inter-segmental coordination. ⁴⁴ This type of integration would allow AI systems to go beyond the superficial analysis of motion to the more in-depth analysis of movement efficiency and skill performance, especially in highly motorized activities like racket games and multi-joint movement sports. In addition, although they are accurate, most high-performance models are black boxes that limits their practical use in sports since they do not give interpretable biomechanical information (e.g., joint kinematics, force distribution, or movement coordination), thus limiting their application to interpretable coaching and training interventions.^23,45 The insufficient explainability limits the practical use of AI systems, since predictions that cannot be explained mechanistically are hard to apply to the coaching and athlete development systems.

AI for injury prevention and detection

Research on injury prevention and detection proved that artificial intelligence can identify early signs of muscle strain and classify injury incidents as well as predict the possibility of sustaining an injury. As an example, the wearable sensors detect a change in the body prior to the occurrence of injury by applying thermal imaging. ³⁵ CNN-based video analysis was able to identify basketball injuries with 90.4% accuracy. ³⁸ Hybrid models that use both psychological, neuromuscular, and biomechanical data have been found to have an ideal classification rate in a controlled environment. ³⁶ One important aspect to note with regard to AI-based injury prediction is the reliance on large and healthy baseline datasets. Most of the models are known to have high predictive accuracy, but the quality and the extent of input data is the key determinant of the reliability of the models. An efficient model should be developed on the ground of comprehensive retrospective cohort profiling (injury history, training load, biomechanical patterns, physiological markers in diverse populations of athletes. ⁴⁶ Such profiling allows determining the underlying risk patterns and increases the overall predictive model predictability. ⁴⁴

As well, not every highly accurate model in controlled laboratory settings can be equally accurate in highly dynamic competitive settings due to contextual variability, environmental effects and athlete specific adaptation. ⁵ This indicates that there is a need to validate AI systems in reality within the sporting environment. Ethical concerns in AI-based sports applications should be dealt with formal regulatory and governance frameworks. The collection and treatment of sensitive information of the athletes should be carried out in accordance with the laws on data privacy, informed consent and principles of biometric data governance. The frameworks such as the General Data Protection Regulation (GDPR) and the Health Insurance Portability and Accountability Act (HIPAA) should be followed to be able to use the data responsibly and avoid losing the trust of the athletes. ⁴⁷ Thus, the successful and ethical application of AI in sports injury prevention and management requires the integration of a strong data governance strategy in addition to the high-quality datasets and real-life validation.

Strengths, limitations, and research gaps

The studies included in this review demonstrate notable strengths, particularly in their holistic evaluation of AI applications in sports, encompassing both performance and injury-related outcomes; however, certain limitations remain. A number of studies employed a small sample size, datasets were often not highly diverse and limited research had experimented AI systems under real competition situations, raising questions on their real-world activities. A major challenge is the lack of clarity in how AI models make decisions. Without understandable explanations, practitioners may hesitate to rely on these predictions in practice. Moreover, none of the studies that have been conducted with multi-season, field, validation is limited to the understanding of long duration reliability. The gaps in the field of sports require further research to develop large and multicentre datasets, such as a broad set of sports and demographics of athletes to strengthen and increase the generalisability of models. An explainable artificial intelligence (XAI) system can be adopted to encourage practitioner confidence and acceptance. As these approaches may provide human-interpretable descriptions of AI decisions, they enhance the trust of practitioners and easily bridge the gap into clinical practice. The correlation between different approaches to the same datasets in benchmarking studies would be a powerful recommendation in the selection of the most appropriate models.

Future direction and practical implications

Future research should focus on building more reliable and generalisable AI models by using large, diverse, and longitudinal datasets that capture not only biomechanical and physiological factors but also neuromuscular aspects of performance. There is also a clear need to test these models in real sporting environments, as many current findings are based on controlled settings that may not reflect actual practice. Improving interpretability is equally important, as coaches and practitioners need to understand how and why a model makes a prediction in order to use it effectively. At the same time, greater attention must be given to the lack of standardisation in sensor technologies, data collection methods, and modelling approaches, which currently limits comparison across studies and reduces confidence in findings. ⁴⁸ Emerging developments such as wearable-integrated systems, edge computing, and digital twins offer promising opportunities for real-time and personalised insights; however, their practical value will depend on proper validation and consistent methodological frameworks. This review used broad search terms to maximise coverage, which may have limited the inclusion of some highly specialised studies. Future work should consider more specific keywords. Overall, advancing AI in sports will require a balanced approach that combines technical development with practical relevance, transparency, and ethical responsibility.