Non-contact lower limb injuries constitute a notable subset of musculoskeletal sports injuries and are of particular importance because they typically arise in the absence of external impact, making them more challenging to predict and more closely linked to modifiable intrinsic and biomechanical risk factors (Belkhelladi et al., 2025; Whittaker et al., 2025). Across youth and adult athletes, such injuries frequently lead to time loss and more than half of anterior cruciate ligament (ACL) injuries in team sports arise from non-contact mechanisms such as cutting or sudden deceleration (Chia et al., 2022; Guan et al., 2021). These injuries are particularly prevalent in sports with repeated high-intensity directional changes—most notably soccer, basketball and rugby—where epidemiological studies consistently report elevated non-contact injury rates (Achenbach et al., 2021; Ekstrand et al., 2011; Evans et al., 2024; López-Valenciano et al., 2020). In elite soccer, for example, over 90% of lower-limb muscle injuries occur through non-contact mechanisms (Ekstrand et al., 2011). Importantly, non-contact injuries are broadly considered preventable, with evidence showing that neuromuscular and strength-focused injury-prevention programs can substantially reduce their incidence (Al Attar et al., 2017; Rössler et al., 2018; Webster and Hewett, 2018; Yu and Garrett, 2007). Beyond their high incidence, non-contact injuries also impose meaningful economic burdens; in the Australian Football League, the annual financial loss per club reaches AUD$188k to 333k, with missed matches due to hamstring strain injuries, predominantly non-contact, increasing by 71% between 2003 and 2012 (Hickey et al., 2014; Lu et al., 2021).

The lack of consensus on the risk factors for non-contact lower limb sports injuries poses a considerable challenge to accurately identifying their underlying causes. Traditional univariate analytical approaches are inherently limited, as their conclusions are often fragmented and fail to account for the complex interactions among multidimensional factors within dynamic sporting environments(Ruddy et al., 2019). Increasing evidence indicates that injuries emerge from nonlinear interactions among physiological, biomechanical, psychological, and environmental variables rather than from any single determinant(Green et al., 2020; Liveris, 2025). This recognition has prompted a shift from linear, single-cause analyses toward more comprehensive and systematic modeling approaches (Bittencourt et al., 2016), enabling identification of critical combinations of risk factors and providing a stronger scientific foundation for individualized injury prediction and prevention strategies.

In recent years, the field of sports science has increasingly adopted machine learning (ML) approaches to uncover latent patterns within large-scale and complex datasets, demonstrating substantial utility in areas such as competition outcome prediction, performance optimization, and tactical decision-making(Horvat and Job, 2020; Hubáček et al., 2019; Ou-Yang et al., 2025; Sampaio et al., 2024; Watson et al., 2021). These advances are gradually reshaping the landscape of sports medicine. However, conventional statistical techniques (primarily logistic regression) struggle to model nonlinear relationships and are prone to biased performance when faced with the pronounced class imbalance typical of prospective injury datasets. As a result, these models frequently classify the majority of non-injury cases correctly while showing substantially reduced sensitivity and limited discriminative capacity for the minority injury outcomes (Lopez-Valenciano et al., 2018; Oliver et al., 2020; Rossi et al., 2018; Ruddy et al., 2018; Ruiz-Perez et al., 2021). By comparison, ML techniques can accommodate nonlinear relationships and complex feature interactions within high-dimensional, multimodal datasets, enabling a more nuanced characterization of injury-related patterns. While not uniformly superior across all applications, ML approaches have shown potential to yield improved sensitivity and more informative risk stratification in certain contexts (Ayala et al., 2019). Furthermore, ML offers a unique advantage in its ability to integrate a broad range of athlete-specific variables, including sport experience, training load characteristics, biological sex, performance level, prior injury history, and sport-specific biomechanical demands, into unified predictive frameworks (Bogaert et al., 2022; Musat et al., 2024; Rommers et al., 2020). This capacity to model complex, individualized risk profiles is especially relevant for non-contact lower limb injuries, which arise from multifactorial and predominantly intrinsic mechanisms. Although challenges remain due to substantial inter-individual variation in tissue tolerance and adaptive capacity (Nassis et al., 2023), continued progress in multimodal data fusion, feature engineering, and rigorous model validation is steadily enhancing the precision and practical relevance of ML-based injury risk estimation. These developments are expected to support more dependable individualized assessments and contribute to more targeted, evidence-informed prevention strategies (Bartlett et al., 2017; Rossi et al., 2018; Wilkerson et al., 2018; Willy, 2018).

Recent reviews have explored machine learning applications in sports injury prediction, including the systematic review by Van Eetvelde et al. (2021), the scoping review by Leckey et al. (2025), and the narrative review by Yuan et al. (2025). These studies provided important overviews of general ML developments and highlighted shared challenges such as heterogeneous data sources, inconsistent injury definitions, small sample sizes, and limited interpretability and external validation, but they largely evaluated ML at a global level across multiple injury types and body regions. Leckey et al. (2025) provided a broad evidence synthesis of ML methods across sports but did not perform an anatomically or mechanism-focused analysis, and Van Eetvelde et al. (2021) emphasized the need for future work to focus on interpretable ML models and injury-specific analyses, yet their review did not provide region- or mechanism-targeted evaluations. Yuan et al. (2025) structured their narrative around the workflow of injury prediction model development, highlighting methodological challenges encountered during model development, such as data preprocessing, feature selection, and model evaluation, but without stratifying findings by anatomical region or injury mechanism. Therefore, a focused, up-to-date synthesis that examines ML applications specifically for non-contact lower-limb injuries, attending to data modalities, class-imbalance strategies, temporal prediction windows, injury-type heterogeneity, and interpretability practices, is warranted to generate more actionable, domain-relevant guidance. To clarify how machine learning methods interface with injury mechanisms, data modalities, and prediction tasks, we present a conceptual framework summarizing the key components of ML-based non-contact lower-limb injury prediction (Figure 1). This framework also serves to situate the scope of the present review within the broader methodological landscape.

In light of these developments and the growing need for precise, individualized risk assessment, the present systematic review aims to synthesize current evidence on ML applications for lower limb non-contact injury prediction and address the following objectives:

1. Summarize the main ML approaches employed in injury prediction and their methodological characteristics;
2. Evaluate their effectiveness in terms of predictive performance; and
3. Examine the role of interpretability techniques in existing studies and assess their implications for clinical translation and practical application.

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines(Page et al., 2021). The review protocol was prospectively registered in PROSPERO (ID: CRD420251070408).

As of January 20, 2025, a comprehensive literature search was performed across three electronic bibliographic databases: Web of Science, PubMed, and SPORTDiscus (EBSCO), using the following search terms: (‘athletic injuries’ OR ‘sports injuries’) AND (‘machine learning’ OR ‘transfer learning’) AND (‘lower extremity’ OR ‘lower limbs’) (the complete search strategy is available in the Supplementary Table 1). Additionally, three reviewers (JY, YZ, and QZ) independently conducted the database search and cross-checked the reference lists of relevant studies.

Studies were included if they met the following criteria: (1) published in English in peer-reviewed journals; (2) original research articles; (3) applied ML techniques to predict non-contact lower limb injuries in humans; and (4) involved athletes or physically active populations, with study designs encompassing prospective, retrospective, or cross-sectional approaches; and (5) published within the last ten years (January 2015 to January 20, 2025). Exclusion criteria were as follows: (1) full text not available in English; (2) studies that did not employ ML methods for non-contact lower limb injury prediction (e.g., those limited to traditional regression analyses); (3) studies that did not report the number of injury cases; and (4) review articles, conference abstracts, or editorials.

All records were screened following a predefined protocol. Grey literature (e.g., theses, dissertations, non-peer-reviewed reports) was excluded a priori, as the review focused exclusively on peer-reviewed scientific evidence. Two independent reviewers (JY and QZ) screened the titles and abstracts of all retrieved studies to determine eligibility based on the predefined inclusion criteria. Inter-rater agreement during screening was assessed using Cohen’s κ coefficient, with discrepancies resolved through discussion with a third reviewer (YZ) until consensus was reached. For the studies that met the inclusion criteria, the reviewers independently extracted relevant data using a predesigned standardized data extraction form(Fernandez-Felix et al., 2023), followed by cross-checking to ensure accuracy. The extracted information included: (1) study characteristics (study design, authors, year of publication, study population, and sample size); (2) machine learning methodology (model type, feature variables, data preprocessing methods, training strategies, performance metrics, and interpretability); and (3) injury-related information (type of injury, anatomical location, number of injury events, and injury definition). If certain information was not reported in a study, it was recorded as “not reported.”

Two independent reviewers (JY and QZ) assessed the risk of bias for the included studies, with discrepancies resolved through arbitration by a third reviewer (YZ) until consensus was reached. Assessment was conducted using the Prediction model Risk Of Bias Assessment Tool (PROBAST)(Wolff et al., 2019), which evaluates four domains: participants, predictors, outcomes, and analysis, comprising 20 signaling questions to determine both domain-specific and overall risk of bias (low, unclear, high). Key considerations included participant representativeness and inclusion criteria; predefinition and reliability of predictors; objectivity and blinding of outcome assessment; and analysis-related issues such as model overfitting, data leakage, failure to address class imbalance, and validation strategy. Applicability was assessed based on the relevance of study participants, predictors, and outcomes to the review question.

The outcomes of the search strategy and study selection process are illustrated in Figure 2. A total of 86 potentially relevant studies were initially identified through systematic searches of the Web of Science, PubMed, and SPORTDiscus databases. An additional 6 articles were retrieved through manual searching, yielding a total of 92 records. After removing 15 duplicates, 77 unique articles remained. Following title and abstract screening, 42 studies were deemed eligible based on the inclusion criteria. Inter-rater reliability for screening was substantial (Cohen’s κ = 0.72). The results from the independent screenings were subsequently consolidated, and any discrepancies were resolved through discussion among the three reviewers (JY, YZ, and QZ). Ultimately, 15 studies were included in the final review. The descriptive characteristics of the included studies are summarized in Supplementary Table 2; Supplementary Table 3 provides the ML and statistical definitions referenced therein.

Among the 15 included studies, four were rated as having a low overall risk of bias, eight as unclear, and three as high risk (see Figure 3 and Figure 4). Bias was predominantly observed in the analysis domain, with key issues including insufficient sample sizes, unreported sensitivity and specificity, and inadequate handling and reporting of missing data. Regarding applicability, nine studies were rated as low concern, four as unclear, and two as high concern. The main applicability limitations were unclear inclusion and exclusion criteria and limited relevance of the predicted outcomes to actual lower-limb muscle injury risk, which may reduce the practical utility of the findings for real-world sports or clinical prevention.

The distribution of publications by year showed a trend of initial growth followed by a subsequent decline: 2018 (n = 2), 2019 (n = 2), 2020 (n = 2), 2021 (n = 2), 2022 (n = 4), and 2023 (n = 3). Regarding sport type, four studies focused on soccer(Ayala et al., 2019; Javier Robles-Palazon et al., 2023; Kolodziej et al., 2023; Oliver et al., 2020), three on basketball(Huang et al., 2022; Huang et al., 2023; Lu et al., 2022), and one study each on football(Ruddy et al., 2018), futsal(Ruiz-Perez et al., 2021), and military personnel(Connaboy et al., 2019). Five studies included multiple sport populations(Bogaert et al., 2022; Henriquez et al., 2020; Jauhiainen et al., 2022; Jauhiainen et al., 2021; Lopez-Valenciano et al., 2018).

Sample sizes ranged from 16 to 2103 participants (Table 1). In terms of sex distribution, six studies recruited only male participants (40%), four included mixed-gender participants (27%), three recruited only female participants (20%), and two did not report participant sex. Regarding age or population characteristics, seven studies involved adult athletes (47%), six involved adolescent athletes, one included recreationally active individuals, and one involved military personnel.

Among the 15 included studies, the majority (n = 11, 73%) evaluated the predictive ability of machine learning models for injuries occurring in any region of the lower limb, while the remaining studies focused on specific anatomical sites, including hamstring strain injuries (HSI, n = 2) (Ayala et al., 2019; Ruddy et al., 2018), knee (n = 1)(Jauhiainen et al., 2022)and one study predicting both knee and ankle injuries (Jauhiainen et al., 2021). Most studies (n = 10) targeted traumatic injuries, three addressed overuse injuries (Bogaert et al., 2022; Huang et al., 2022; Huang et al., 2023), and two did not specify the injury mechanism (Connaboy et al., 2019; Henriquez et al., 2020) (Table 2). Regarding predictor variables, the most frequently collected features could be classified into three major domains: demographics and injury history, psychological and perceptual variables, and physical performance measures. Thirteen studies (87%) incorporated demographic information such as age, height, weight, competitive level, and prior injury history, with nine studies explicitly including previous injury as a predictor. Five studies assessed psychological and perceptual factors, most commonly sleep quality (n = 4)(Ayala et al., 2019; Huang et al., 2022; Lopez-Valenciano et al., 2018; Ruiz-Perez et al., 2021), alongside other constructs such as sport anxiety, team cohesion, and stress levels. Less commonly, innovative predictors such as urinary biomarkers (n = 2)(Huang et al., 2022; Huang et al., 2023) and match performance indicators (n = 1)(Lu et al., 2022) were also reported.

In terms of data preprocessing, 13 studies reported explicit procedures. The most common steps included data normalization (n = 8, using either Z-score standardization or Min-Max scaling) to harmonize feature scales, and data imputation to handle missing values. Additionally, four studies used the Weka software package for preprocessing (Javier Robles-Palazon et al., 2023; Lopez-Valenciano et al., 2018; Oliver et al., 2020; Ruiz-Perez et al., 2021), including imputation and discretization. Two studies did not report any preprocessing.

Feature selection or dimensionality reduction techniques were reported in eight studies (53%). Expert-based feature selection was used in one study(Jauhiainen et al., 2021), while the Attribute Selected Classifier from Weka was applied in two(Javier Robles-Palazon et al., 2023; Ruiz-Perez et al., 2021). Feature importance ranking (Mean Decrease Accuracy) was adopted in one study(Henriquez et al., 2020). Dimensionality reduction techniques included principal component analysis (PCA, n = 1)(Bogaert et al., 2022), recursive feature elimination (RFE, n = 1)(Lu et al., 2022), linear discriminant analysis (LDA, n = 1)(Huang et al., 2023), and least absolute shrinkage and selection operator (LASSO, n = 1)(Kolodziej et al., 2023). The remaining seven studies did not report any such methods.

A key characteristic of the included datasets was class imbalance. Based on the sample counts reported in the original studies, we recalculated the imbalance ratio (IR = minority/majority, where minority refers to injury cases and majority to non-injury cases). The average IR across studies was 0.35, with values ranging from 0.08 to 0.77. To address this issue, ten studies adopted imbalance-handling strategies, which could be broadly grouped into two categories: (i) resampling methods, including SMOTE (n = 4)(Ayala et al., 2019; Huang et al., 2022; Huang et al., 2023; Jauhiainen et al., 2022; Lopez-Valenciano et al., 2018; Ruddy et al., 2018) and under-sampling bagging (n = 2)(Javier Robles-Palazon et al., 2023; Ruiz-Perez et al., 2021); and (ii) cost-sensitive learning (n = 2)(Bogaert et al., 2022; Oliver et al., 2020). All included studies adopted cross-validation methods such as 5-fold, 10-fold, or leave-one-out.

Among the 15 included studies, 4 (27%) employed a single ML model for predictive modeling, whereas the remaining studies compared multiple models to identify the one with optimal predictive performance (Table 3). Specifically, 1 study evaluated 2 models, 3 studies evaluated 3 models, 5 studies evaluated 4 models, and 2 studies assessed more than 4 models. Across all studies, random forest (RF) and logistic regression were the most frequently applied algorithms, each appearing in 8 studies (53%), followed by support vector machine (SVM), which were used in 7 studies. In addition, decision trees and their variants (e.g., C4.5, SimpleCart, ADTree, CHAID) were applied in 6 studies. By contrast, extreme gradient boosting (XGBoost) was less commonly used, reported in only 2 studies. Overall, tree-based models and their ensemble methods emerged as the most prevalent approaches for sports injury prediction.

Among the 15 included studies, four (27%) identified decision trees (DT) as the best-performing models(Ayala et al., 2019; Connaboy et al., 2019; Lopez-Valenciano et al., 2018; Oliver et al., 2020), followed by SVM (n = 4)(Bogaert et al., 2022; Jauhiainen et al., 2022; Javier Robles-Palazon et al., 2023; Ruiz-Perez et al., 2021) and logistic regression (LR, n = 2)(Jauhiainen et al., 2021; Kolodziej et al., 2023). Notably, one study employing the CHAID variant of DT reported the highest predictive performance across all studies (AUC = 0.91)(Connaboy et al., 2019). Overall, six studies (40%) demonstrated that tree-based algorithms, including RF, XGBoost, and DT variants were the most effective, underscoring their advantage in balancing interpretability, generalizability, and stability.

With respect to model evaluation, the area under the curve (AUC) was the most widely used metric, reported in 14 studies (93%). Among the studies reporting AUC, seven (47%) fell within the “poor” range (0.50-0.69)(Bogaert et al., 2022; Henriquez et al., 2020; Jauhiainen et al., 2022; Jauhiainen et al., 2021; Kolodziej et al., 2023; Oliver et al., 2020; Ruddy et al., 2018) (0.50-0.69), three (20%) were rated as “fair” range (0.70-0.79)(Javier Robles-Palazon et al., 2023; Lopez-Valenciano et al., 2018; Ruiz-Perez et al., 2021), another three as “good” range (8.80-0.89)(Ayala et al., 2019; Huang et al., 2023; Lu et al., 2022), and only one reached the “excellent” level (≥ 0.90)(Connaboy et al., 2019). The mean AUC across all studies was 0.73. In addition to AUC, sensitivity was the second most frequently reported metric, appearing in eight studies (53%), with values ranging from 0.35 to 0.92 and a mean of 0.63. Specificity was reported in six studies, ranging from 0.62 to 0.84 with a mean of 0.74, while precision was reported in only two studies.

Among the 15 included studies, 13 (87%) reported interpretability analyses. Eight studies relied on inherently interpretable models (“white-box” algorithms(Belle and Papantonis, 2021)), primarily DT, LR, and RF. Five studies used post-hoc interpretability techniques, including SHAP (n = 4) and logistic regression applied to SVM (n = 1)(Bogaert et al., 2022).

Across the studies conducting interpretability analysis, body mass index (BMI) and previous injury history were consistently identified as important predictors across multiple studies (BMI: 4 studies(Connaboy et al., 2019; Jauhiainen et al., 2021; Javier Robles-Palazon et al., 2023; Oliver et al., 2020); previous injury history: 3 studies(Ayala et al., 2019; Lopez-Valenciano et al., 2018; Lu et al., 2022)). In addition, biomechanical features—particularly range of motion (ROM), muscle strength, and neuromuscular control—were identified as relevant predictors in several studies. These variables were primarily obtained from laboratory-based assessments using isolated screening tests, such as isokinetic or isometric strength testing(Jauhiainen et al., 2021; Kolodziej et al., 2023), goniometric or motion-capture–based ROM evaluation(Ayala et al., 2019), and balance or perturbation tasks to assess neu-romuscular control. In fewer cases, validated field-based protocols (e.g., the ROM-Sport battery(Ruiz-Perez et al., 2021)) were used to capture these capacities in applied settings.

This systematic review synthesized 15 studies investigating ML approaches for non-contact lower limb injury prediction. Overall, tree-based algorithms were the most frequently applied and often achieved the highest predictive performance, with one study using the decision-tree variant CHAID reaching an AUC of 0.91 (Connaboy et al., 2019), exceeding the mean AUC (0.73) across studies by 25%. While AUC was the primary evaluation metric in most studies (93%), sensitivity values, reported in a subset of studies, varied widely (0.35-0.92, mean = 0.63), highlighting differences in models’ ability to identify actual injury cases.

A notable feature of the included studies is that most (79%, n = 11) generalized the prediction target to “any lower limb injury event.” Although this approach increases statistical power in smaller datasets, it reduces clinical specificity because different injury types (e.g., ACL tears, ankle sprains, hamstring strains) have distinct biomechanical mechanisms, risk factors, and intervention pathways. This limitation underscores the need for injury-specific prediction models and contextualizes both model performance and the interpretation of feature importance.

Among the included studies, AUC was the most frequently reported metric for evaluating model performance, primarily reflecting the ability of a model to discriminate between positive and negative cases across varying thresholds. However, in the highly imbalanced context of lower limb injury prediction (average imbalance ratio = 0.35), a high AUC does not necessarily indicate satisfactory identification of the minority class, namely the actual injury cases (Van Eetvelde et al., 2021). To address this limitation, several studies additionally reported sensitivity and specificity to provide a more comprehensive assessment of clinical utility (Ayala et al., 2019; Javier Robles-Palazon et al., 2023; Kolodziej et al., 2023; Lopez-Valenciano et al., 2018; Oliver et al., 2020; Ruiz-Perez et al., 2021). In real-world sports injury prevention, practitioners often adopt a strategy of “erring on the side of caution”—prioritizing the identification of high-risk individuals even at the cost of increased false positives—thereby making higher sensitivity particularly important (Florkowski, 2008). Nevertheless, among the six studies in this review that reported both sensitivity and specificity, sensitivity values ranged from 0.35 to 0.78, whereas specificity ranged from 0.62 to 0.84. Notably, only one study demonstrated higher sensitivity than specificity (Ruiz-Perez et al., 2021), while the remaining studies showed the opposite pattern, including one with a sensitivity as low as (Kolodziej et al., 2023). From a clinical perspective, such imbalances indicate that many models are more effective at correctly identifying non-injury cases than detecting minority injury events, which may limit their utility for timely injury prevention and early intervention—settings where high sensitivity is particularly important.

While machine learning models generally demonstrate competitive predictive performance, they do not consistently outperform traditional statistical approaches. For example, Jauhiainen et al.(Jauhiainen et al., 2021) reported that LR achieved a slightly higher AUC (0.65) than RF (0.63) in youth athletes, and Oliver et al.(Oliver et al., 2020) similarly found LR (AUC = 0.69) to marginally exceed a DT model (AUC = 0.66) in elite youth soccer players. However, when evaluating performance beyond AUC, substantial differences emerged. In Oliver et al., the DT achieved markedly higher sensitivity (55.6%) compared with LR (11.1%), despite similar AUC values. This discrepancy highlights a critical issue: under class-imbalanced conditions common in injury datasets, AUC alone may mask models’ ability to correctly identify injury cases. Thus, the mixed findings do not indicate a fundamental limitation of LR per se, but rather emphasize that model evaluation must account for metrics sensitive to minority-class detection when comparing ML with traditional methods.

In prospective injury prediction studies, class imbalance is a pervasive challenge, as injury cases are typically much less frequent than non-injury cases. Addressing this imbalance is therefore critical for robust model development. Among the 15 studies included, 10 (67%) applied specific strategies to handle imbalance, primarily resampling or cost-sensitive learning. Resampling was the most common (80%), involving techniques such as synthetic minority oversampling (e.g., SMOTE) to generate new “injury” samples, or under-sampling combined with ensemble learning (e.g., under-sampling bagging) to reduce “non-injury” samples. Although SMOTE was applied in six studies, three of them reported that its use did not improve predictive performance (Jauhiainen et al., 2022; Lopez-Valenciano et al., 2018; Ruddy et al., 2018). This pattern reflects a broader limitation of over-sampling in injury prediction: when synthetic samples are generated from nearest neighbors, the minority class may be overly homogenized, masking rare but clinically informative patterns and increasing overfitting risk (Carvalho et al., 2025; Fernández et al., 2018). In contrast, López-Valenciano et al. (2018) observed marginal gains using random under-sampling, which avoids synthetic noise but removes substantial majority-class information that may be essential for stable decision boundaries. Together, these findings illustrate a central methodological challenge in injury prediction: conventional resampling techniques often fail to capture the complex, low-prevalence nature of injury events. This suggests that using data-driven recommendation systems, such as those based on dataset complexity measures, to automatically identify the most appropriate resampling strategy may offer a more effective solution (Carvalho et al., 2025).

A frequently cited example is the study by Rommers et al. (2020), which prospectively monitored 734 elite youth soccer players (U10 - U15) across a full competitive season. Their models achieved balanced predictive performance (accuracy, sensitivity, and specificity all = 0.85) when forecasting both acute and overuse injuries. The study adopted a clear and standardized injury definition, recording any physical complaint that required evaluation by medical or paramedical staff; medical personnel were present at every training session and match, ensuring complete medical-attention reporting. Injuries included both event-related acute cases and overuse injuries without a single causal incident, and predictions covered injuries across the entire body rather than focusing on a specific anatomical region. Two factors likely contributed to the model’s favorable performance. First, the dataset was unusually well balanced (50.1% injured vs. 49.9% non-injured), which helped minimize the class imbalance issues that typically challenge injury prediction models (Javier Robles-Palazon et al., 2023; Kolodziej et al., 2023; Ruiz-Perez et al., 2021). In a balanced dataset, ML models may be better positioned to learn injury-related patterns because the minority class is more adequately represented during training. Evidence from youth soccer injury-prediction studies suggests that models developed from relatively balanced class distributions (e.g., Rommers et al., 2020), IR ≈ 1.0) tend to report higher AUC values (≈ 0.85) compared with those trained on more imbalanced datasets (IR = 0.21-0.39, AUC = 0.66-0.70)(Javier Robles-Palazon et al., 2023; Oliver et al., 2020). While these findings do not establish a causal relationship, they indicate that class balance can contribute to improved predictive performance under certain conditions. Second, the adolescent sample (mean age = 11.7 ± 1.7 years) falls within a developmental period where injury risk shows clear age-related variation, with the 13 to 15 age range identified as the peak-incidence period (Rumpf and Cronin, 2012)—making such patterns easier for machine learning algorithms to detect (Jauhiainen et al., 2022). The study predicted injuries across the entire body, which may have improved overall model accuracy and stability. However, this broad classification reduces the ability to provide actionable guidance for specific anatomical sites. Predicting injuries by region would allow for more targeted prevention strategies and tailored interventions, which are typically more relevant in clinical practice.

In the clinical translation of ML for lower limb injury prediction, although some studies have reported strong model performance over extended prediction windows (AUC ≥ 0.8)(Ayala et al., 2019; Connaboy et al., 2019; Huang et al., 2023), their clinical applicability remains limited. These models frequently adopt long-term injury outcomes (e.g., across a season or a year) as labels. While this approach facilitates the accumulation of sufficient injury cases and mitigates the problem of “extreme class imbalance,” it may compromise the temporal validity of predictions. On the one hand, athletes’ risk status dynamically fluctuates with variations in training load and physiological condition (Bache-Mathiesen et al., 2022; Johnston et al., 2019). Because most injury-prediction studies in this field adopt a prospective design (Van Eetvelde et al., 2021), the predictor data are collected before the injury occurs. However, data obtained several months prior to the injury may still fail to reflect the athlete’s immediate pre-injury condition. On the other hand, excessively long prediction windows reduce the actionable value of risk alerts, thereby constraining their utility for training monitoring and rehabilitation management. To enhance clinical feasibility, future research should investigate modeling strategies based on periodic screenings (e.g., monthly or per training cycle) to capture risk features closer to injury onset, thereby improving both the timeliness and practical relevance of predictions.

Notably, attempts have been made to develop short-term injury prediction models. For example, Briand et al. (Briand et al., 2022) proposed a framework for predicting injuries within 1-7 days, but its average sensitivity was only 0.35 ± 0.19, underscoring the methodological challenges associated with sample distribution and feature sensitivity in short-term predictions. More recently, a four-year longitudinal study in professional football applied machine learning to internal (RPE) and external (GPS-derived) workload data from the two-week and four-week periods prior to injury (Martins et al., 2025). Using a four-week window, the KStar classifier achieved a sensitivity of 0.69, a specificity of 0.76, and an AUC of 0.81. The two-week models delivered slightly lower but still meaningful predictive performance, with the MLP yielding a sensitivity of 0.75, a specificity of 0.69, and an AUC of 0.79.Collectively, these findings further demonstrate that short-term injury risk prediction is achievable when leveraging multidimensional workload indicators.

A further challenge that directly influences clinical translation, yet is often overlooked in existing reviews, is the heterogeneity of injury mechanisms included in model development. Previous syntheses (Leckey et al., 2025; Van Eetvelde et al., 2021; Yuan et al., 2025) did not systematically distinguish between contact and non-contact injuries in their inclusion criteria. This lack of differentiation leads to pooled evidence combining fundamentally different etiological pathways: contact injuries are frequently driven by external forces or collisions, whereas non-contact injuries are more closely linked to intrinsic factors, neuromuscular control, and biomechanical patterns (Dauty et al., 2022; Yu and Garrett, 2007). Aggregating these mechanisms may obscure true model performance, alter feature importance profiles, and reduce the generalizability of findings. By contrast, the present review adopts a strictly defined non-contact lower-limb injury criterion, reducing etiological heterogeneity and enabling a more coherent evaluation of prediction models within a mechanistically consistent category. This focus provides clearer insight into which data modalities, feature representations, and ML architectures are effective for non-contact injury risk and strengthens the translational relevance of the synthesized evidence.

Looking forward, emerging methodological frameworks provide promising avenues for improving model timeliness and contextual relevance. The Weighted Cumulative Exposure (WCE) approach, implemented within Piecewise Exponential Additive Mixed Models allows researchers to model how past training loads accumulate and exert time-dependent effects on injury risk (Zumeta-Olaskoaga et al., 2025). These models flexibly estimate the time window during which previous exposures meaningfully contribute to current injury hazard, enabling predicttions that better reflect the evolving load patterns experienced in real-world training environments.

In the field of lower limb injury prediction, ML models have demonstrated promising predictive performance; however, their practical utility extends beyond conventional metrics such as accuracy or AUC. A critical issue is whether these models can be reliably trusted in clinical or sports settings. Trustworthiness depends not only on predictive capability but also on model interpretability and reliability. Cross-validation plays an essential role in this context, providing a more robust estimate of model generalizability and reducing the risk of overfitting. Notably, all 15 studies included in this review employed cross-validation procedures, underscoring its role as a standard methodological safeguard. Nevertheless, cross-validation alone does not guarantee clinical or applied reliability. Conventional cross-validation can produce overly optimistic performance estimates when data exhibit temporal dependence, as is common in training-load-based injury prediction (Roberts et al., 2017). This highlights the need for time-aware validation strategies and, more broadly, for external validation on independent cohorts. Complementing cross-validation with external validation and domain-relevant interpretability is therefore essential to ensure real-world trustworthiness(Ramspek et al., 2021). Ultimately, the goal of injury prediction is not only to identify high-risk individuals but also to reveal actionable mechanisms underlying injury risk. Analogous to the established link between smoking and cancer, interpretable models can inform targeted intervention strategies (Wang et al., 1999). Consequently, interpretability constitutes a key prerequisite for translating ML models from research into practice.

Existing literature shows considerable variability in how interpretability is conceptualized within ML-based injury prediction. Prior reviews, such as those by Leckey et al. (2025) and Yuan et al. (2025), have largely centered their discussion on post-hoc explanation techniques, particularly SHAP, to interpret complex “black-box” models including XGBoost, neural networks and SVM. While these methods are valuable for quantifying feature contributions, they represent only one dimension of model interpretability. In contrast, the present review underscores the importance of inherently interpretable “white-box” models such as DT and RF. These algorithms offer transparency by design, enabling direct inspection of decision pathways and feature relevance without external interpretability tools (Belle and Papantonis, 2021). This is one reason why tree-based models remain prevalent in injury-prediction research, as their structure supports accessible metrics of feature importance (for example, split frequency or impurity-based measures) that facilitate clear identification of salient risk factors. Empirical studies further illustrate the advantages of these models. López-Valenciano et al. (2018) used DT classifiers to highlight previous injury history and strength asymmetries as primary determinants of lower-extremity injury risk. Similarly, Ruiz-Pérez et al. (2021) applied RF and identified workload and neuromuscular parameters as dominant predictors based on impurity-based importance scores. These examples demonstrate how tree-based approaches not only reveal influential variables but also clarify how these features interact to stratify athletes into different risk profiles. Such transparency is particularly valuable in applied sport settings where practitioners must interpret and justify risk assessments.

Compared with the SHAP-centric approach in previous syntheses, our broader framing highlights that model interpretability can arise either from intrinsic model structure or from post-hoc explanation techniques applied to more complex architectures. Recognizing both pathways provides a more comprehensive understanding of how ML outputs can inform mechanism-oriented interpretations and guide evidence-based intervention design (Kulshrestha et al., 2021; Majumdar et al., 2022).

At the feature level, several studies have identified relatively stable risk factors. When demographic variables were included, BMI and previous injury history frequently emerged as key predictors, consistent with broader musculoskeletal injury literature (Hecksteden et al., 2023; Rommers et al., 2020). In addition to commonly used demographic and biomechanical variables, many studies have incorporated psychological measures into their models (Ayala et al., 2019; Javier Robles-Palazon et al., 2023; Lopez-Valenciano et al., 2018; Ruiz-Perez et al., 2021). Notably, Lipps Lene et al. (2024) directly compared models with and without psychological factors and found that adding these variables significantly improved predictive performance (p < 0.001).

However, evidence across studies also shows considerable variability in the relative importance of individual predictors, which complicates their clinical use. Ruddy et al. (2018) examined whether supervised learning models using preseason eccentric hamstring strength, age, and previous HSI history could accurately predict hamstring strain injuries in elite Australian footballers. Although the models were trained on the same dataset, performance fluctuated widely (AUC 0.24-0.92) due to minor changes in training-testing partitions. This instability reflected meaningful season-to-season differences in cohort characteristics: injured players were substantially weaker than uninjured players in 2013, whereas no strength differences were observed in 2015 despite similar HSI incidence. These findings demonstrate that the influence of commonly cited risk factors is highly context dependent and shaped by variations in conditioning status, training load, and population profiles. Consequently, predictors identified in one season or team may not generalize reliably to others.

Further evidence of contextual fluctuation is provided by Ayala et al. (2019), who integrated neuromuscular, personal, and psychological variables into an injury-specific model. They observed that no single predictor consistently dominated across classifiers. Instead, variables such as sleep quality, hip flexion range of motion, and angle-specific torque contributed variably, reflecting the inherently multifactorial nature of HSI etiology. Importantly, their injury-specific modeling strategy produced stronger predictive performance than studies relying on limited or non-specific feature sets, suggesting that predictor stability improves when models are grounded in mechanisms directly relevant to the injury being predicted.

Taken together, these findings indicate that the relevance of individual predictors varies substantially across seasons, populations, and modeling frameworks. Therefore, machine-learning-derived predictors should not be assumed to generalize across contexts unless they are rooted in injury-specific mechanisms and validated across multiple cohorts. For clinical application, this underscores the importance of developing models that incorporate comprehensive, injury-relevant features and that undergo external validation before being used to guide risk-mitigation strategies.

Despite systematically reviewing current advances in applying ML to lower limb injury prediction, several limitations should be acknowledged. First, although all included studies used some form of internal cross-validation such as k-fold or leave-one-out, considerable methodological heterogeneity remained across studies in terms of study populations, injury types, feature engineering strategies, and prediction windows. More importantly, most studies relied only on internal validation and did not conduct independent external validation, which limits the generalizability of model performance and may contribute to inconsistencies in the reported findings. Second, the transparency of ML methodology in the included studies was limited. Many studies provided insufficient detail regarding model development pipelines, hyperparameter tuning procedures, software toolboxes, and code availability. Differences in how model interpretability was conceptualized and implemented, together with variation in injury sites studied, further hinder cross-study comparisons and reduce the feasibility of systematic integration and clinical translation. Third, this review included only peer-reviewed publications written in English, excluding non-English articles, theses, conference papers, and grey literature. Although this approach enhances methodological rigor, it may also have resulted in the omission of relevant evidence. Finally, this review synthesized findings qualitatively and did not perform a meta-analysis. The absence of pooled effect estimates prevents direct quantitative comparisons of ML algorithm performance. Therefore, the findings should be interpreted cautiously, and future research, especially large-scale multicenter studies with transparent methodological reporting and external validation, is needed to strengthen and extend these conclusions.

This review demonstrates that ML holds considerable potential for predicting non-contact lower limb injuries; however, its clinical utility depends not only on predictive performance but also on interpretability and reliability. White-box algorithms offer inherent transparency, enhancing clinical comprehensibility, whereas black-box models, despite achieving higher predictive accuracy, face limitations in trustworthiness due to their opacity. Therefore, future research should strive to balance predictive performance with interpretability by integrating post-hoc explanation techniques and hybrid modeling frameworks to facilitate clinical translation. Moreover, standardized data collection and feature selection, integration of multi-source information, and large-scale prospective studies are critical for enhancing model robustness and generalizability across populations. Overall, only through the coordinated development of predictive performance, interpretability, and methodological rigor can ML truly support precision injury prevention and rehabilitation in sports practice.