From data to prevention: A systematic review of artificial intelligence applications in sports injury prediction

The study populations demonstrated an increasing trend toward demographic and geographic diversity. While earlier research predominantly featured professional male athletes, more recent studies have expanded to include female and other non-traditional cohorts. For example, Jauhiainen et al. developed an ACL injury prediction model using data from female elite handball and soccer players, whereas Merrigan et al. examined noncontact lower-body injury predictability in female NCAA Division I athletes using countermovement jump force-time metrics.^26,28 In addition, the geographic scope of recent studies has broadened beyond traditionally dominant North American and European settings. For instance, Saberisani et al. investigated football injury prediction in Iranian professional players, Tabben et al. analyzed subsequent injury patterns in professional football players from the Qatar Stars League, and Guo et al. examined ACL injury prediction in male collegiate basketball players in an Asian context.^24,29,32 Furthermore, the demographic scope has expanded beyond elite professionals to include collegiate athletes and more diverse competitive contexts, providing broader insight beyond traditional male professional cohorts. Overall, these developments suggest a gradual move toward greater contextual diversity, although male athletes remain overrepresented in the current literature.

3.3.2 Methodological quality and validation strategies

A significant validation gap persists within the field of injury prediction. Notably, most included studies relied solely on internal validation strategies, such as k-fold cross-validation or random data splitting, rather than external validation in independent cohorts. For example, Jauhiainen et al. explicitly reported repeated cross-validation procedures and emphasized the instability of prediction estimates across repetitions, ultimately concluding that statistical predictive ability did not necessarily translate into clinical usefulness. ²⁸ Only a small minority of studies reported more rigorous testing procedures or uncertainty estimates. Castellanos et al., for instance, evaluated concussion prediction using a held-out test set and reported AUROC with 95% confidence intervals, while the study by Lu et al. was one of the few studies in the review to provide confidence intervals for model performance metrics.^17,30 Additionally, reporting quality was often limited, with many studies failing to clearly describe stratification procedures, temporal separation strategies, or safeguards against data leakage in longitudinal datasets. Collectively, these issues suggest that many reported performance metrics may be optimistic and may not fully reflect real-world model generalizability.

3.3.3 Technical frameworks and predictors

To provide a more structured synthesis of the heterogeneous evidence, the included studies can be broadly grouped according to model type, predictor domain, and sport context. In terms of model type, tree-based ensemble methods, including Random Forest, XGBoost, and CatBoost, were commonly used for handling non-linear injury risk factors in tabular datasets involving workload, injury history, clinical, and biomechanical variables. For example, Guo et al. reported that Random Forest achieved the highest predictive performance for ACL injury risk among male basketball players, ³² while Weng et al. identified CatBoost as the best-performing model in baseball-related upper extremity injury prediction. ²¹ Support Vector Machine and logistic-regression-based approaches were also frequently applied, particularly in smaller datasets or studies emphasizing interpretable classification. Emerging research has also begun to apply deep learning and temporal modeling approaches. Cohan et al., for instance, used a deep learning framework to forecast injuries in NBA basketball based on longitudinal injury and game-related data. ³¹

In terms of predictor domains, the included studies generally drew on four overlapping categories: workload and exposure variables, previous injury and clinical history, biomechanical or neuromuscular measures, and subjective or contextual indicators such as sleep quality, academic stress, or questionnaire-based measures. Previous injury history and training load metrics remained among the most frequently reported and consistent predictors across studies. Tsilimigkras et al. further extended this line of work by incorporating both external load and internal load variables, including heart-rate-related measures, into a soccer injury risk model, highlighting the importance of physiological overload alongside mechanical demand. ³³ Meanwhile, more complex models increasingly incorporated biomechanical and neuromuscular indicators. Guo et al. integrated sport-specific side-cutting biomechanics and electromyography, ³² while Bergeron et al. demonstrated the added value of subjective psychological indicators such as sleep quality and academic stress. ³

Finally, sport contexts varied across the included evidence, with football/soccer, basketball, baseball, rugby, and collegiate multi-sport cohorts represented. This structured grouping suggests that model choice was often shaped by data structure and sport context, but the lack of standardized predictor definitions and validation procedures limits direct comparison across studies. Together, these findings suggest an increasing emphasis on holistic and multimodal injury prediction frameworks.

3.3.4 Model performance metrics

Reported model performance was variable, with AUC values in many studies falling approximately between 0.65 and 0.85, suggesting modest to moderate discriminative ability in most cases. For example, Jauhiainen et al. reported a mean AUC-ROC of 0.63 for ACL injury prediction in female elite athletes, emphasizing that statistically significant prediction above chance may remain insufficient for clinical implementation. ²⁸ By contrast, Claros et al. reported an AUC of 0.82 for predicting post-concussion musculoskeletal injury risk in collegiate athletes, suggesting that more comprehensive and clinically structured variable sets may improve predictive discrimination. ²³

However, performance metrics should be interpreted cautiously. Merrigan et al. reported a high overall model accuracy of 85.6% but a low AUC of 0.659 and a recall of 0.0%, indicating that no injuries in the testing dataset were correctly classified. ²⁶ Similarly, Haller et al. observed apparently high classification accuracy in a holistic monitoring framework, yet also acknowledged low precision and a considerable number of false positives. ²⁷ These examples highlight that accuracy alone is insufficient for evaluating injury prediction models, particularly in imbalanced datasets, and that clinically meaningful interpretation requires consideration of recall, precision, AUC, and generalizability.

4.2.1 Methodological quality and validation gap

A major finding of this review is the persistence of a clear validation gap. Most included studies used internal validation strategies such as random train-test splits, k-fold cross-validation, or bootstrap-based resampling, whereas external validation across independent cohorts was uncommon. This is important because internally validated models may perform well within the development dataset yet fail to generalize when applied to new teams, seasons, institutions, or athlete populations. Jauhiainen et al., for example, explicitly demonstrated that although prediction performance was statistically better than chance, the resulting ACL injury model still lacked sufficient clinical utility. ²⁸

This concern is consistent with the broader methodological literature. Majumdar et al. noted that football injury prediction studies are characterized by substantial heterogeneity in variable selection, data treatment, balancing procedures, and validation approaches, making unified interpretation difficult and limiting practical translation. ³⁴ They further emphasized that further progress in the field will require not only better-performing algorithms but also multiple seasons of data, careful handling of class imbalance, and more interpretable analytical pipelines.

The validation gap also affects how high reported performance values should be interpreted. Some studies reported strong accuracy or discrimination metrics, but these estimates were often derived from small samples, event-sparse datasets, or single-center cohorts. Under such conditions, even apparently strong model performance may partly reflect overfitting, optimistic sampling, or hidden leakage rather than robust predictive ability. Accordingly, future studies should treat external validation as a core design requirement rather than an optional enhancement.

4.2.2 Reporting quality and interpretation of performance metrics

A second important issue concerns reporting quality. Across the included studies, model reporting was inconsistent, particularly with respect to uncertainty estimates, calibration, data preprocessing steps, and the handling of missing values. Only a very limited number of studies reported confidence intervals for key performance metrics. This omission substantially limits the ability to judge the precision, stability, and reproducibility of reported models.

Equally important, this review highlights that model performance cannot be interpreted on the basis of accuracy alone. In injury prediction, class imbalance is often substantial because injury events are much less frequent than non-injury observations. Under these conditions, a model may achieve high overall accuracy while still performing poorly in identifying injured athletes. The study by Merrigan et al. provides a clear example: despite relatively high accuracy, the model's AUC was modest and recall was 0.0%, meaning that no injuries in the testing dataset were correctly classified. ²⁶ Haller et al. likewise reported apparently strong classification results but also acknowledged low precision and a notable false-positive burden. ²⁷

This interpretation aligns closely with the concerns raised by both Cohan et al. and Majumdar et al., who emphasized that sports injury datasets are typically imbalanced and that inappropriate evaluation metrics can overstate practical usefulness.^31,34 Therefore, future research should routinely report a broader range of metrics, including AUC, recall/sensitivity, specificity, precision, F1-score, and calibration-related information, ideally with confidence intervals.

4.2.3 Predictor domains: From foundational variables to multimodal frameworks

Although algorithmic diversity has increased, the main predictor domains used in injury prediction models have remained relatively consistent. Previous injury history and workload-related measures continue to emerge as foundational predictors across multiple studies and sports. This pattern is biologically plausible and practically meaningful: prior injury likely reflects residual impairment, incomplete recovery, or persistent vulnerability, whereas workload variables capture cumulative and acute mechanical and physiological stress.

At the same time, the literature is increasingly moving beyond these core variables. Tsilimigkras et al. incorporated both external and internal load, including heart-rate-derived variables, suggesting that physiological overload may add predictive value beyond purely mechanical training metrics. ³³ Guo et al. integrated sport-specific biomechanics and EMG during unanticipated side-cutting, illustrating the value of capturing movement patterns at biomechanically critical phases. ³² Claros et al. combined demographic, injury-specific, and concussion-assessment variables to model subsequent musculoskeletal injury after concussion, while Bergeron et al. highlighted the importance of psychological and contextual variables such as sleep quality and academic stress.^3,23

These developments are consistent with a broader shift in sports biomechanics and athlete monitoring. Dhahbi emphasized that biomechanical analysis, wearable technologies, and neuromuscular monitoring are increasingly central not only to performance optimization but also to injury mitigation and rehabilitation planning. ³⁵ In this sense, the growing move toward multimodal models is theoretically coherent. However, the inclusion of more variables does not automatically guarantee better real-world performance. Larger and more heterogeneous datasets also create new challenges related to missingness, feature redundancy, dimensionality, and transportability. Thus, the field should pursue multimodal integration cautiously and with strong methodological discipline.

4.2.4 Gender imbalance and population generalizability

Another important observation is the continued overrepresentation of male athletes. Although more recent studies have begun to include female elite athletes and NCAA women's sport cohorts, the literature remains predominantly male-centered. This imbalance matters because female athletes may differ from male athletes in hormonal milieu, neuromuscular control, biomechanics, exposure patterns, and injury mechanisms. Models derived primarily from male cohorts may therefore not generalize well to female populations without dedicated validation.

This issue extends beyond sex to broader questions of generalizability. The geographical profile of the literature has become more diverse, with recent studies from Iran, Qatar, Greece, and broader Asian contexts, yet representation remains uneven. Gaddour et al. argued that AI-driven football injury prediction research remains geographically incomplete, particularly in underrepresented regions such as Africa, where contextual, infrastructural, and athlete-specific factors may differ substantially from those of Europe or North America. ³⁶ This point is relevant more broadly: the external validity of injury prediction models likely depends not only on sport type but also on local training culture, competition structure, environmental conditions, and resource availability. This concern also extends to structured physically demanding populations outside elite sport. For example, Dhahbi et al. reported a retrospective cohort of 979 newly recruited male police cadets during an initial training phase, showing that musculoskeletal injury profiling in tactical cohorts is both feasible and clinically relevant, with nearly half of recruits sustaining at least one injury and most injuries occurring between Weeks 2 and 5. ³⁷ Taken together, these observations suggest that future research should prioritize more inclusive and context-sensitive datasets rather than assuming that models developed in one setting will be universally portable.

4.2.5 Clinical translation, explainability, and implementation challenges

From a practical standpoint, the current evidence suggests that AI models should be viewed as decision-support tools rather than replacements for clinician or practitioner judgment. This is particularly true for injury prediction, where model outputs must be interpreted in context and where false positives and false negatives may carry meaningful consequences for both athlete welfare and performance planning. The implementation challenge is therefore not simply to maximize discrimination, but to develop models that are interpretable, operationally usable, and trustworthy.

Explainability is especially important here. Some recent studies have already begun to incorporate explainable AI techniques, such as SHAP-based model interpretation, to clarify which variables contribute most strongly to risk classification. This approach is valuable not only for building user trust, but also for generating clinically meaningful hypotheses and identifying modifiable intervention targets. Weng et al., for instance, used SHAP to interpret model predictions in baseball-related injury forecasting, illustrating how explainability can bridge predictive analytics and individualized prevention. ²¹

In addition, the increasing use of wearable devices, computer vision, force plates, and other sensor-based platforms introduces a separate but related challenge: technological standardization. This concern has been emphasized in recent methodological commentary on AI and sport technologies. In “The Algorithmic Athlete,” Dhahbi and Chamari argued that AI-driven and sensor-based systems require more rigorous and standardized evaluation extending beyond model performance alone to include assessment of the underlying measurement technologies, including verification, analytical validation, and practical or clinical validation. ³⁸ This point is highly relevant to sports injury prediction. Even a theoretically strong model may be difficult to interpret or reproduce if the underlying measurement system lacks standardization, validity, or consistent reporting. Accordingly, progress in injury prediction will depend not only on better algorithms, but also on better measurement infrastructure and stronger reporting standards.