Development and validation of a machine learning model for non-contact injury prediction based on lower limb strength asymmetry in professional football

Abstract

Non-contact injuries significantly impact professional football, yet traditional risk assessment methods demonstrate limited predictive accuracy. This study aimed to develop and validate a machine learning model integrating lower limb strength asymmetry data for injury prediction. A prospective cohort study enrolled 312 professional male football players from six European clubs (2022–2024). Comprehensive isokinetic strength testing, functional assessments, and 12-month injury surveillance were conducted. Four machine learning algorithms (Random Forest, SVM, GBDT, Deep Neural Networks) were developed using nested cross-validation, with performance evaluated through AUPRC, calibration, and clinical utility metrics. Temporal validation was conducted on a subsequent season cohort from the same clubs. During follow-up, 89 players (28.5%) sustained 127 non-contact injuries. Players with knee flexor asymmetry > 15% demonstrated 3.2-fold increased injury hazard (HR: 3.24, 95% CI: 2.18–4.82). The ensemble model achieved superior predictive performance (AUPRC: 0.759) compared to baseline logistic regression (0.589). Observational implementation of risk-stratified interventions was associated with 73% reduction in injury probability within four weeks. Preliminary cost-effectiveness analysis suggested €215,800 net savings per club season. Machine learning models incorporating strength asymmetry data demonstrate improved injury risk prediction performance in professional football. The observational clinical implementation framework suggests potential for effective injury prevention through personalized interventions, though randomized controlled trials are needed to establish causal relationships. These findings represent a step toward precision prevention strategies in sports medicine.

Introduction

Sports injuries represent a significant challenge in professional football, with non-contact injuries accounting for a substantial proportion of time-loss incidents among elite players¹. The financial and competitive implications of these injuries extend beyond individual athletes, affecting team performance and organizational success. Recent economic analyses have demonstrated that injury-related underachievement in elite football leagues results in substantial financial losses, with English Premier League teams alone experiencing millions in revenue reduction due to player unavailability².

Lower limb strength asymmetry has emerged as a critical risk factor for non-contact injuries in football players, particularly affecting the hamstrings, quadriceps, and adductor muscle groups³. Contemporary research has revealed complex patterns in bilateral strength deficits. Female players exhibit higher asymmetry tendencies compared to males, particularly in knee flexor strength, and approximately 50% of elite futsal players demonstrate preseason strength imbalances that may increase hamstring injury risk^4,5. While sports injuries have complex, multifactorial etiology involving biomechanical, physiological, and environmental factors, strength asymmetry represents a particularly important modifiable risk factor due to its objective measurability, standardized assessment protocols, and direct amenability to targeted interventions. However, recent prospective studies have challenged traditional assumptions, demonstrating that low angular velocity isokinetic testing alone (60°/s) was insufficient as an injury prediction factor in professional soccer players, highlighting the need for more comprehensive assessment approaches⁶. Traditional injury risk assessment methods, which rely primarily on univariate statistical analyses and subjective clinical evaluations, have shown limited predictive accuracy in identifying athletes at high risk of injury. This limitation has prompted researchers to explore more sophisticated analytical approaches that can capture the complex, multifactorial nature of sports injuries.

Machine learning (ML) techniques have shown promising potential in injury prediction in sports medicine, offering new opportunities to analyze large-scale, multidimensional datasets and identify complex patterns that traditional methods might overlook^7,8. However, the application of ML in sports injury prediction faces significant challenges. These include data heterogeneity across different populations and sports contexts, potential overfitting in relatively small datasets, limited interpretability in clinical settings where transparent decision-making is essential, and difficulties in standardizing data collection protocols across different research centers^9,10. The challenge of developing practical testing protocols is particularly relevant in professional football, where financial resources, coach perceptions, and playing schedules significantly influence implementation. Field-based methods are often preferred due to their low cost and ability to assess multiple players simultaneously¹¹. Despite these challenges, several studies have demonstrated that ML algorithms can achieve promising predictive performance compared to conventional logistic regression models in forecasting sports injuries across various athletic populations¹². Recent advances have shown that functional assessment tools, such as the Lower Extremity Functional Test, can achieve high predictive accuracy (AUC 0.908) in discriminating between injured and non-injured professional football players¹³. Additionally, predictive models incorporating body composition variables and physical fitness tests can generate relatively low error rates (RMSE = 0.591) for injury risk assessment¹⁴. However, such functional tools primarily assess current injury status rather than predict future injury occurrence, and provide categorical rather than continuous data that may limit their integration into machine learning frameworks requiring high-dimensional feature inputs. In football specifically, ML approaches have been successfully applied to predict injury risk using GPS training data, physiological monitoring, and biomechanical assessments^15,16. Recent work in basketball analytics has demonstrated the potential of advanced data science approaches in understanding performance patterns and injury impacts, providing insights that may be applicable to football injury prediction¹⁷.

Despite these advances, several critical gaps remain in the current literature. First, while several studies have explored either strength asymmetry assessment or ML-based injury prediction independently, comprehensive integration of multi-velocity isokinetic data with ensemble ML approaches and clinical implementation frameworks remains limited¹⁸. The selection of isokinetic strength assessment as a primary focus is justified by its standardized measurement protocols, high reliability in elite athletic populations, and established clinical relevance for injury risk stratification. Recent studies have confirmed excellent test-retest reliability for commonly used strength measures in professional football settings^19,20. However, it is important to acknowledge that asymmetry magnitude is highly task-dependent. Research shows that perceived functional capabilities may correlate with mechanical asymmetry patterns, suggesting the need for comprehensive assessment approaches that consider both objective measurements and subjective functional indicators²¹. Second, existing prediction models often lack the granularity needed to distinguish between different types of non-contact injuries, limiting their practical utility for targeted prevention strategies. Third, while ML models may achieve superior statistical performance, their interpretability and practical implementation in real-world clinical and athletic settings remain significant challenges, hindering their adoption where transparent decision-making is essential²². Importantly, recent advances in unilateral training interventions have demonstrated that targeted approaches can effectively reduce lower limb asymmetry in athletes, supporting the potential for evidence-based prevention strategies based on asymmetry assessment²³.

This study addresses these gaps by developing a novel ML-based prediction model that integrates comprehensive lower limb strength asymmetry assessments with multimodal injury risk factors to predict non-contact injuries in elite football players. Our approach combines isokinetic strength testing data, training load metrics, and injury history within an interpretable ML framework. The primary objectives of this research are to: (1) establish quantitative thresholds for lower limb strength asymmetry associated with increased injury hazard, (2) develop and validate ML models that outperform traditional statistical approaches in predicting non-contact injuries, (3) identify the most important predictive features through explainable AI techniques, and (4) provide actionable insights for implementing individualized injury prevention strategies.

The main contributions of this work include: (i) a comprehensive dataset combining isokinetic strength measurements, training load metrics, injury history, and prospective injury tracking from professional football players over a 12-month period; (ii) a comparative analysis of multiple ML algorithms (Random Forest, Support Vector Machines, Gradient Boosting, and Deep Neural Networks) for injury prediction; (iii) the development of an interpretable risk stratification system that can guide clinical decision-making; and (iv) validation of the proposed model through both internal cross-validation and external testing on an independent cohort.

Methods

Research design

This prospective cohort study followed the TRIPOD-AI guidelines for transparent reporting of machine learning-based prediction models. A detailed TRIPOD-AI compliance checklist is provided in Table S1. The study employed a multi-phase design to develop and validate machine learning models for predicting non-contact injuries in elite football players. The study protocol followed the consensus statement on injury definitions and data collection procedures in football research²⁴, and adhered to the methodological framework for sports injury prediction studies.

The research was conducted over a 24-month period, comprising a 12-month primary data collection phase (2022–2023 season) followed by a 12-month temporal validation phase (2023–2024 season). During the development phase, 312 professional male football players were recruited from six elite clubs competing in top-tier European leagues during the 2022–2023 and 2023–2024 seasons. Sample size was determined based on anticipated injury incidence of approximately 25–30% derived from previous epidemiological studies in professional football¹, to achieve a sufficient events-per-variable ratio for stable machine learning model development. The temporal validation cohort comprised 186 players from the same six clubs during the 2023–2024 season, including both continuing players from the original cohort (n = 142, representing 45.5% of the original sample) and newly recruited players (n = 44). The remaining 170 players from the original cohort were unavailable for follow-up due to transfers to other clubs (n = 98), retirement (n = 12), long-term injury (n = 31), or withdrawal of consent (n = 29). Baseline isokinetic assessments were repeated for all validation cohort participants prior to the 2023–2024 season. The focus on male players ensures cohort homogeneity but limits generalizability to female football players who may exhibit different injury patterns.

Figure 1 illustrates the complete six-phase research workflow, progressing from participant recruitment through baseline assessment, 12-month prospective monitoring, data processing and feature engineering, machine learning model development using four algorithms, and temporal validation over the 24-month study period.

During the baseline assessment phase, participants underwent comprehensive lower limb strength testing using isokinetic dynamometry, with biomechanical parameters derived from the testing data. The prospective surveillance phase involved continuous monitoring of training loads, match participation, and injury occurrence using standardized injury reporting systems. All non-contact injuries resulting in time loss from training or competition were recorded according to UEFA injury study methodology.

The machine learning model development followed established best practices for predictive modeling in sports medicine²⁵. Feature engineering incorporated domain expertise to create meaningful variables from raw strength asymmetry data, training load metrics, and injury history. Multiple algorithms were evaluated using nested cross-validation to prevent overfitting and ensure robust performance estimates.

Fig. 1

Research design workflow.

Research participants

Inclusion criteria

Participants were recruited according to stringent inclusion criteria to ensure a homogeneous study population representative of elite football performance levels. Eligible participants were required to be professional male football players aged 18–35 years, contracted to clubs competing in the top two tiers of European national leagues. All players must have completed a minimum of three years of structured professional training and maintained regular participation in team training sessions (≥ 4 sessions per week) during the study period. Additional requirements included medical clearance for full training and competition participation, absence of current injury at baseline assessment, and the ability to perform maximal isokinetic strength testing without pain or functional limitation. Players were required to have documented injury history records from the previous two seasons to enable comprehensive risk factor analysis^26,27. Written informed consent was obtained from all participants. Ethical approval was granted by the institutional review boards of participating clubs and research institutions in accordance with the Declaration of Helsinki.

Exclusion criteria

Exclusion criteria were established to minimize confounding variables and ensure data quality for machine learning model development. Players were excluded if they had sustained a major lower extremity injury (> 28 days time-loss) within six months prior to baseline testing, as this could significantly affect strength measurements and injury risk profiles. Participants with current musculoskeletal complaints, systemic inflammatory conditions, or neurological disorders affecting motor control were excluded. Use of medications known to influence muscle strength or injury healing (corticosteroids, anabolic agents, or certain antibiotics) within three months of testing constituted grounds for exclusion. Players undergoing concurrent rehabilitation programs or those with incomplete baseline assessments were excluded from analysis. Additionally, participants who failed to complete the minimum 12-month follow-up period due to transfer, retirement, or loss to follow-up were excluded from the final dataset. While this 12-month follow-up period enables comprehensive injury pattern analysis within a single season, it may not fully capture inter-seasonal variability that could influence long-term injury risk assessment. These criteria ensured that the study population represented healthy elite players at genuine risk of non-contact injuries during regular training and competition.

Data collection protocol

Lower limb strength assessment

Comprehensive lower limb strength evaluation was conducted using isokinetic dynamometry (Biodex System 4 Pro, Biodex Medical Systems, USA) following standardized protocols. Bilateral concentric and eccentric strength of knee flexors and extensors were measured at three angular velocities (60°/s, 180°/s, and 300°/s) to capture different neuromuscular demands across the force-velocity spectrum. Specifically, 60°/s assessed maximal force production, 180°/s evaluated functional movement speeds during cutting and decelerating, and 300°/s captured high-velocity actions during sprinting where most hamstring injuries occur²⁸. Hip adductor and abductor strength were evaluated at 60°/s and 120°/s using specialized attachments. Each testing session began with a standardized 10-minute warm-up protocol including cycling and dynamic stretching, followed by familiarization trials at each test velocity²⁹. Participants performed five maximal repetitions at 60°/s and 300°/s, and 15 repetitions at 180°/s, with 60-second rest intervals between sets. Peak torque, average power, and torque-angle curves were recorded for each muscle group. The conventional hamstring-to-quadriceps (H: Q) ratio was calculated as concentric knee flexor peak torque divided by concentric knee extensor peak torque at matched velocities (60°/s and 180°/s). The dynamic control ratio was calculated as eccentric knee flexor peak torque divided by concentric knee extensor peak torque. All ratios were computed using peak torque values from the full range of motion (0–90° knee flexion). Strength asymmetry indices were calculated using the formula: [(stronger limb - weaker limb) / stronger limb] × 100, with thresholds > 10% considered clinically significant. This asymmetry index method was selected for its widespread use and clinical interpretability. However, as noted by Bishop et al.³⁰, multiple calculation methods exist with inherent limitations, including inability to account for limb dominance direction and temporal fluctuations between sessions.

Injury surveillance system

A comprehensive injury surveillance system was implemented following UEFA Elite Club Injury Study methodology to ensure standardized data collection across all participating clubs. Medical staff recorded all injuries using a web-based platform accessible via secure login, with real-time data synchronization to the central research database. Non-contact injuries were defined as tissue damage occurring without direct external force, resulting in player unavailability for training or match participation. Each injury entry included detailed information: anatomical location (using Orchard Sports Injury Classification System 10), injury type, mechanism, activity at injury onset, and severity classification based on time-loss days. Previous injury history was comprehensively defined as all lower extremity injuries occurring within 24 months prior to baseline assessment. This included injury type (muscle, ligament, tendon), severity classification based on time-loss days, time interval since most recent injury, and rehabilitation completion status. All data were verified through club medical records. Medical staff underwent standardized training on injury definitions and reporting procedures to ensure inter-rater reliability. Weekly audits were conducted to verify data completeness and accuracy. Training exposure was recorded daily in minutes, with match exposure documented separately⁹. Injury incidence rate was calculated as injuries per 1000 player-hours, enabling standardized comparisons across different training phases and between clubs.

Data processing and feature engineering

The comprehensive data processing and feature engineering pipeline is illustrated in Fig. 2, showing the transformation from raw multi-source data through preprocessing, feature extraction across four domains, and advanced engineering techniques to produce the final 47-variable dataset optimized for machine learning analysis. Raw data underwent systematic preprocessing to ensure quality and compatibility for machine learning analysis. Missing values comprised 3.2% of total data points and were addressed using multiple imputation by chained equations with five imputation cycles. Missing data patterns were analyzed to confirm data were missing at random rather than systematically. Continuous variables were assessed for normality using Shapiro-Wilk tests, with appropriate transformations applied to highly skewed distributions. Outliers were identified using the interquartile range method and verified through clinical review to distinguish genuine extreme values from measurement errors.

Feature engineering incorporated domain expertise to create meaningful predictors from raw measurements. Strength asymmetry features included bilateral strength ratios, agonist-antagonist ratios (H: Q ratio), and dynamic control ratios at different velocities. Notably, asymmetry indices were analyzed alongside absolute strength values, as low asymmetry with bilaterally low strength may still confer elevated injury risk compared to higher asymmetry with greater absolute strength. Time-series features were derived from training load data, including acute: chronic workload ratios (7:28 days), exponentially weighted moving averages, and load monotony indices³¹. Biomechanical features encompassed joint-specific strength curves, rate of force development, fatigue resistance indices calculated from repeated contractions, and movement quality indicators derived from torque-angle curve analysis.

Advanced feature engineering techniques included interaction terms between strength asymmetry and training load, polynomial features capturing non-linear relationships, and temporal features representing seasonal variations. Dimensionality reduction using principal component analysis was applied to highly correlated strength measurements while preserving interpretability. Feature selection employed a multi-stage approach: univariate filtering (p < 0.20), recursive feature elimination with cross-validation, and embedded methods (LASSO regularization) to identify optimal feature subsets. The final feature set comprised 47 variables categorized into strength asymmetry indices (n = 18), training load metrics (n = 12), biomechanical parameters (n = 10), and injury history factors (n = 7). This selection provided comprehensive representation of injury risk factors while maintaining model parsimony.

Fig. 2

Data processing pipeline.

Machine learning model development

Four machine learning algorithms were selected based on their demonstrated efficacy in sports injury prediction and ability to handle complex, non-linear relationships inherent in biomechanical data. Random Forest (RF) was chosen for its robustness to overfitting and feature importance capabilities. Support Vector Machines (SVM) with radial basis function kernel were selected for their effectiveness in high-dimensional spaces. Gradient Boosting Decision Trees (GBDT) were included for their superior predictive performance in imbalanced datasets. Deep Neural Networks (DNN) were implemented to capture complex interaction patterns between features.

The RF model aggregated predictions from B decision trees, where each tree ${T_b}$ was trained on a bootstrap sample of size n with m randomly selected features at each split. The final prediction was computed as:

$$\hat {y}=\frac{1}{B}\sum\limits_{{b=1}}^{B} {{T_b}} (x)$$

(1)

For SVM, the optimization problem was formulated to find the optimal hyperplane in the transformed feature space:

$${\hbox{min} _{w,b,\xi }}\frac{1}{2}||w|{|^2}+C\sum\limits_{{i=1}}^{n} {{\xi _i}}$$

(2)

subject to: ${y_i}({w^T}\phi ({x_i})+b) \geqslant 1 - {\xi _i}$ and ${\xi _i} \geqslant 0$.

where $\phi (x)$ represents the kernel transformation, C is the regularization parameter, and ${\xi _i}$ are slack variables for soft-margin classification.

The GBDT model iteratively minimized the loss function through additive training:

$${F_m}(x)={F_{m - 1}}(x)+{\gamma _m}{h_m}(x)$$

(3)

where ${h_m}(x)$ is the weak learner at iteration m and ${\gamma _m}$ is the step size determined by line search:

$${\gamma _m}=\arg {\hbox{min} _\gamma }\sum\limits_{{i=1}}^{n} L ({y_i},{F_{m - 1}}({x_i})+\gamma {h_m}({x_i}))$$

(4)

The outer loop used stratified 5-fold cross-validation to assess model generalization, while the inner loop performed 3-fold cross-validation for hyperparameter optimization. Class imbalance was addressed using Synthetic Minority Over-sampling Technique (SMOTE), generating synthetic samples for the minority (injured) class. To mitigate the risk of artificially inflated performance that can occur with SMOTE in high-dimensional spaces, model evaluation was conducted on both SMOTE-enhanced training data and original validation datasets.

$${x_{new}}={x_i}+\lambda ({x_{nn}} - {x_i})$$

(5)

where ${x_i}$ is a minority class sample, ${x_{nn}}$ is one of its k-nearest neighbors, and $\lambda \sim U(0,1)$.

Hyperparameter optimization utilized Bayesian optimization with Gaussian Process surrogate models to efficiently explore the parameter space³². The acquisition function balanced exploration and exploitation:

$$\alpha (x)=\mu (x)+\kappa \sigma (x)$$

(6)

where $\mu (x)$ and $\sigma (x)$ are the predicted mean and standard deviation, and $\kappa$ controls the exploration-exploitation trade-off.

Model performance was evaluated using multiple metrics accounting for class imbalance. The primary metric was the Area Under the Precision-Recall Curve (AUPRC), calculated as:

$$AUPRC=\sum\limits_{{k=1}}^{n} {({R_k} - {R_{k - 1}})} {P_k}$$

(7)

where ${P_k}$ and ${R_k}$ are precision and recall at the k-th threshold. Additionally, the F2-score was computed to emphasize recall:

$${F_2}=5 \cdot \frac{{precision \cdot recall}}{{4 \cdot precision+recall}}$$

(8)

Model interpretability was achieved through SHAP (SHapley Additive exPlanations) values, providing consistent feature attribution:

$${\phi _i}=\sum\limits_{{S \subseteq F \setminus \{ i\} }} {\frac{{|S|!(|F| - |S| - 1)!}}{{|F|!}}} [{f_S}(S \cup \{ i\} ) - {f_S}(S)]$$

(9)

where ${\phi _i}$ represents the contribution of feature i, F is the set of all features, and ${f_S}$ is the model trained on feature subset S. The calibration of probabilistic predictions was assessed using the Brier score:

$$BS=\frac{1}{n}\sum\limits_{{i=1}}^{n} {{{({p_i} - {y_i})}^2}}$$

(10)

where ${p_i}$ is the predicted probability and ${y_i}$ is the actual outcome. Decision curve analysis evaluated clinical utility by calculating net benefit across probability thresholds:

$NB({p_t})=\frac{{TP}}{n} - \frac{{FP}}{n} \cdot \frac{{{p_t}}}{{1 - {p_t}}}$

where ${p_t}$ is the probability threshold, and TP and FP$represent true and false positives at that threshold.

Fig. 3

Machine learning model development framework.

Figure 3 presents the comprehensive machine learning framework, encompassing nested cross-validation, class imbalance handling, parallel algorithm training, and multi-metric evaluation with interpretability analysis. All models were implemented using Python 3.9.7 with scikit-learn 1.0.2, TensorFlow 2.8.0, and XGBoost 1.5.1. Key hyperparameters were optimized using Bayesian optimization: Random Forest (n_estimators = 500, max_depth = 15), SVM (C = 1.0, gamma = 0.1), GBDT (learning_rate = 0.1, n_estimators = 100), and DNN (3 hidden layers: 64-32-16 neurons, dropout = 0.2). All random processes used fixed seeds (random_state = 42) to ensure reproducibility. Complete hyperparameter specifications are provided in Table S2.

Statistical analysis

Statistical analyses were performed using R version 4.2.1 and Python 3.9. Continuous variables were reported as mean ± standard deviation or median (interquartile range) based on normality assessment via Shapiro-Wilk tests. Categorical variables were expressed as frequencies and percentages. Between-group comparisons utilized independent t-tests or Mann-Whitney U tests for continuous data, and chi-square or Fisher’s exact tests for categorical variables. Multiple comparisons were adjusted using Bonferroni correction. Time-to-injury analysis employed Kaplan-Meier curves with log-rank tests. Multivariate associations were examined through Cox proportional hazards regression, reporting hazard ratios with 95% confidence intervals. Statistical significance was set at p < 0.05 (two-tailed).

Receiver operating characteristic (ROC) analysis was used to identify optimal asymmetry cut-off values for injury prediction. The Youden index method was applied to determine thresholds that maximized the sum of sensitivity and specificity. These ROC-derived thresholds were subsequently used to categorize players for Kaplan-Meier survival analyses. In Cox proportional hazards regression, strength asymmetry variables were entered as categorical variables based on clinically meaningful thresholds (< 10%, 10–15%, > 15%). Continuous variables (e.g., age, training load metrics) were entered in their original continuous form.

Results

Research population characteristics

A total of 312 professional male football players were enrolled from six elite European clubs during the 2022–2023 and 2023–2024 seasons. The mean age of participants was 24.7 ± 4.2 years, with a professional playing experience of 6.3 ± 3.8 years. Anthropometric characteristics revealed a mean height of 181.4 ± 6.7 cm and body mass of 78.2 ± 7.9 kg, resulting in a mean body mass index (BMI) of 23.7 ± 1.8 kg/m². The cohort comprised defenders (31.4%), midfielders (38.5%), forwards (24.0%), and goalkeepers (6.1%), reflecting typical squad compositions in professional football.

During the 12-month prospective follow-up period, 89 players (28.5%) sustained at least one non-contact lower extremity injury, with a total of 127 injury events recorded. The overall injury incidence rate was 4.8 injuries per 1000 player-hours (95% CI: 4.2–5.4). As shown in Table 1, significant differences were observed between injured and non-injured players in several baseline characteristics. Injured players demonstrated higher baseline strength asymmetry indices for both knee extensors (12.4% vs. 8.1%, p < 0.001) and flexors (14.7% vs. 9.3%, p < 0.001)³³. Additionally, injured players had significantly lower hamstring-to-quadriceps ratios at 60°/s (58.2% vs. 63.1%, p = 0.002) and higher acute: chronic workload ratios in the four weeks preceding injury (1.42 vs. 1.18, p < 0.001).

Previous injury history emerged as a significant differentiating factor, with 67.4% of injured players reporting prior lower extremity injuries compared to 38.1% in the non-injured group (p < 0.001). Training load parameters also differed significantly, with injured players accumulating higher weekly training hours (18.6 ± 3.2 vs. 16.9 ± 2.8 h, p = 0.003) and demonstrating greater training load variability (CV: 28.4% vs. 21.2%, p = 0.006)³⁴.

Table 1 Baseline characteristics of study participants stratified by injury status during follow-up.

Injury occurrence analysis

Throughout the 12-month surveillance period, 127 non-contact lower extremity injuries were documented among 89 players, representing an overall injury burden of 2,847 days lost from training and competition. The median time to first injury was 142 days (95% CI: 118–166 days), with 31.5% of injuries occurring within the first three months of observation. Hamstring strain injuries constituted the predominant injury type (n = 48, 37.8%), followed by quadriceps strains (n = 23, 18.1%) and ankle ligament injuries (n = 19, 15.0%). The anatomical distribution revealed a clear predominance of muscle injuries (68.5%) compared to ligamentous (20.5%) and tendinous (11.0%) injuries.

Temporal analysis demonstrated distinct seasonal patterns in injury occurrence. Peak incidences were observed during pre-season preparation (July-August: 7.2 injuries/1000 hours) and the congested winter fixture period (December-January: 6.8 injuries/1000 hours). As shown in Table 2, injury severity classification according to time-loss criteria revealed that moderate injuries (8–28 days absence) were most prevalent (44.9%), while severe injuries (> 28 days) accounted for 22.0% of all cases, contributing disproportionately to the total injury burden (58.4% of total days lost).

The mechanism of injury analysis indicated that high-speed running activities were associated with 42.5% of all injuries, with acceleration/deceleration movements accounting for 28.3% and change of direction maneuvers for 18.9%. Notably, 76.5% of hamstring injuries occurred during sprinting activities, while quadriceps injuries showed a more diverse mechanistic profile. Recurrent injuries comprised 24.4% of all injury events, with significantly shorter time to re-injury (median 67 days) compared to index injuries³⁵. Players with bilateral strength asymmetry exceeding 15% demonstrated a 2.8-fold higher injury incidence during high-intensity activities compared to those with asymmetry below 10% (RR: 2.84, 95% CI: 1.92–4.21, p < 0.001).

Table 2 Distribution and characteristics of non-contact lower extremity injuries during the study period.

Lower limb strength asymmetry analysis

Descriptive analysis

Comprehensive isokinetic strength assessment revealed substantial inter-limb asymmetries across multiple muscle groups and testing velocities in the study cohort. The overall mean bilateral strength asymmetry index was 10.2 ± 5.1%, with 38.8% of players exceeding the clinically significant threshold of 10% asymmetry in at least one muscle group. As shown in Table 3, knee flexor strength demonstrated the highest asymmetry values across all testing velocities, with peak asymmetry observed at 300°/s (11.9 ± 6.2%). Conversely, hip abductor strength exhibited the lowest asymmetry indices, maintaining relative bilateral symmetry even at higher angular velocities (7.2 ± 3.2% at 60°/s).

Angular velocity significantly influenced asymmetry magnitudes, with a progressive increase observed from 60°/s to 300°/s across all muscle groups (p < 0.001)³⁶. This velocity-dependent pattern was most pronounced in knee flexors, where asymmetry increased by 28.4% from slow to fast velocities. Positional analysis revealed that forwards displayed significantly higher quadriceps asymmetry (11.8 ± 5.4%) compared to defenders (8.9 ± 4.2%, p = 0.008) and midfielders (9.3 ± 4.7%, p = 0.021). This may reflect position-specific movement demands and injury patterns³⁷.

The dominant-to-non-dominant limb strength ratio analysis indicated that 72.4% of players demonstrated greater strength in their preferred kicking leg for knee extensors. This pattern reversed for knee flexors, where 61.5% showed greater strength in the non-dominant limb. Agonist-antagonist muscle balance, assessed through hamstring-to-quadriceps ratios, revealed concerning imbalances in 29.2% of players, with ratios below the recommended 60% threshold at 60°/s. These imbalances were more prevalent in players with previous hamstring injuries (41.3% vs. 24.7%, p = 0.012). Peak torque analysis demonstrated significant correlations between absolute strength values and asymmetry indices, with weaker players exhibiting greater bilateral differences (r=-0.42, p < 0.001)³⁸.

Table 3 Lower limb strength asymmetry indices across muscle groups and angular velocities.

Injury correlation analysis

The relationship between lower limb strength asymmetry and non-contact injury risk demonstrated a clear dose-response pattern across all muscle groups examined. Players with knee flexor asymmetry exceeding 15% exhibited a 3.2-fold increased injury hazard (HR: 3.24, 95% CI: 2.18–4.82, p < 0.001) compared to those with asymmetry below 10%. This association remained significant after adjusting for age, playing position, previous injury history, and training load parameters in multivariate Cox regression models. These Cox regression findings informed the subsequent machine learning analysis, with knee flexor asymmetry identified as the highest-ranking predictor in SHAP feature importance analysis.

As shown in Fig. 4, the cumulative injury incidence progressively increased with higher asymmetry categories, with the steepest risk gradient observed for hamstring injuries.

Receiver operating characteristic (ROC) analysis identified optimal asymmetry cut-off values for injury prediction, with knee flexor asymmetry at 180°/s demonstrating the highest discriminative ability (AUC: 0.78, 95% CI: 0.72–0.84) with an optimal threshold of 12.8%. The combined asymmetry score was calculated as the weighted average of asymmetry indices from four muscle groups (knee flexors, knee extensors, hip adductors, and hip abductors), with weights derived from individual AUC values. This combined score achieved superior predictive performance (AUC: 0.83, 95% CI: 0.78–0.88) compared to individual muscle assessments. Time-to-event analysis revealed that players with bilateral strength differences exceeding 15% experienced significantly shorter injury-free periods (median: 98 days) compared to symmetrical players (median: 267 days, log-rank p < 0.001).

Interaction analysis uncovered important modifying effects, whereby the impact of strength asymmetry on injury risk was amplified in players with high training loads (ACWR > 1.5) and those with previous injury history³⁹. Specifically, the hazard ratio for injury in players with both high asymmetry (> 15%) and elevated ACWR increased to 4.87 (95% CI: 3.21–7.39), suggesting synergistic effects between biomechanical and load-related risk factors. Logistic regression analyses examining specific injury types revealed that knee flexor asymmetry strongly predicted hamstring injuries (OR: 3.76, 95% CI: 2.34–6.05), while hip adductor asymmetry showed stronger associations with groin injuries (OR: 2.91, 95% CI: 1.67–5.08).

Fig. 4

Comprehensive analysis of lower limb strength asymmetry and injury risk relationships. (A) Injury incidence stratified by asymmetry categories showing dose-dependent increase (*p < 0.05, ***p < 0.001). (B) ROC curves comparing predictive performance of individual muscle groups versus combined asymmetry score. (C) Kaplan-Meier survival curves demonstrating time to first injury across asymmetry levels (log-rank p < 0.001). (D) Dose-response relationship between continuous asymmetry values and injury probability with 95% confidence bands. (E) Correlation heatmap showing inter-relationships among risk factors. (F) Forest plot of multivariable hazard ratios for key risk factors with 95% confidence intervals. KF: knee flexors; KE: knee extensors; H:Q: hamstring-to-quadriceps ratio; ACWR: acute: chronic workload ratio.

Predictive model performance evaluation

A baseline logistic regression model using age, playing position, and previous injury history was developed for comparison. The machine learning models demonstrated robust predictive performance for non-contact injury risk, with the ensemble approach achieving superior discrimination compared to individual algorithms and the baseline model. The baseline model achieved an AUPRC of 0.589 (95% CI: 0.521–0.657), providing a reference point for evaluating the machine learning approaches. As shown in Fig. 5, the gradient boosting decision tree (GBDT) model attained the highest individual performance with an AUPRC of 0.721 (95% CI: 0.678–0.764), followed by random forest (0.698, 95% CI: 0.654–0.742), deep neural network (0.689, 95% CI: 0.644–0.734), and support vector machine (0.672, 95% CI: 0.627–0.717). The weighted ensemble model, integrating predictions from all four algorithms, achieved an AUPRC of 0.759 (95% CI: 0.718-0.800), representing a significant improvement over individual models (p < 0.001) and a 29% relative improvement over the baseline approach.

Fig. 5

Comprehensive evaluation of machine learning model performance for injury risk prediction. (A) Comparison of area under the precision-recall curve (AUPRC) across models with 95% confidence intervals (***p < 0.001 for ensemble vs. individual models). (B) Receiver operating characteristic curves showing discriminative ability with AUC values in parentheses. (C) Precision-recall curves demonstrating performance across different classification thresholds relative to baseline injury prevalence (dashed line). (D) Calibration plots comparing predicted probabilities against observed proportions (perfect calibration shown as diagonal). (E) Feature importance ranking from ensemble model showing top 15 predictors with cumulative importance on secondary axis. (F) Decision curve analysis illustrating net benefit across probability thresholds, with shaded region indicating clinically relevant range (0.15–0.60). RF: Random Forest; SVM: Support Vector Machine; GBDT: Gradient Boosting Decision Trees; DNN: Deep Neural Network.

Cross-validation analysis revealed consistent performance across all five outer folds, with minimal variance in AUPRC values (coefficient of variation: 4.2%), indicating robust generalization capability. The model classification performance at the clinically relevant 30% risk threshold is detailed in Table 4, showing the ensemble model achieved sensitivity of 75.3% (95% CI: 65.8–83.1%), specificity of 84.3% (95% CI: 78.9–88.8%), positive predictive value of 67.4% (95% CI: 58.2–75.8%), and negative predictive value of 88.9% (95% CI: 84.1–92.6%). Calibration assessment through the Hosmer-Lemeshow test indicated good agreement between predicted probabilities and observed outcomes (χ² = 7.82, p = 0.451), with the calibration slope approaching unity (0.94, 95% CI: 0.87–1.01).

Table 4 Model performance metrics at 30% risk threshold.

To further understand model performance characteristics, detailed misclassification analysis was conducted to identify patterns in false positive and false negative predictions. Table 5 presents the confusion matrices for all models, revealing distinct error patterns that have important implications for clinical implementation.

Table 5 Confusion matrix analysis and misclassification Patterns.

Analysis of the 32 false positive cases in the ensemble model revealed that these players predominantly exhibited moderate strength asymmetry (mean knee flexor asymmetry: 12.8 ± 2.1%) combined with relatively stable training loads (ACWR: 1.09 ± 0.18), suggesting borderline risk profiles that challenge precise classification. While these players demonstrated concerning biomechanical patterns, they remained injury-free throughout the follow-up period, potentially due to compensatory mechanisms or effective self-regulation of training intensity.

The 22 false negative cases were characterized by rapid injury onset within 30 days of baseline assessment (median: 18 days), indicating injuries that may have resulted from acute risk factors not captured in the screening protocol. These misclassified injured players demonstrated lower baseline asymmetry values (mean knee flexor asymmetry: 8.4 ± 3.2%) but exhibited significantly higher training load volatility (CV: 34.2 ± 8.7%) compared to correctly classified injured players (CV: 24.6 ± 6.1%, p = 0.021). This pattern suggests that acute load spikes may override the protective threshold typically associated with lower baseline asymmetry.

Feature importance analysis identified strength asymmetry indices as the most influential predictors, with knee flexor asymmetry at 180°/s contributing 18.4% to the model’s predictive power. The top ten features collectively accounted for 72.3% of total importance, demonstrating effective feature selection. Decision curve analysis revealed substantial clinical utility across a wide range of probability thresholds (0.15–0.60), with the ensemble model providing greater net benefit compared to both treat-all and treat-none strategies. Temporal validation on the 2023–2024 cohort (n = 186) from the same clubs confirmed model stability, with performance degradation (AUPRC: 0.738) within acceptable limits. During the validation phase, 52 players (28.0%) sustained non-contact injuries, consistent with the injury incidence observed in the development cohort (28.5%), confirming the representativeness of the validation sample. However, population-based external validation across different clubs and geographic regions remains to be established.

Clinical decision analysis

The clinical implementation of the predictive model required development of actionable risk stratification protocols tailored to real-world practice constraints. Players were categorized into four risk tiers based on predicted injury probability, with distinct intervention strategies assigned to each category. As shown in Table 6, this stratification system achieved optimal balance between resource allocation and preventive efficacy, with 82.4% of subsequent injuries occurring in players classified as high or very high risk at baseline. The number needed to screen (NNS) to prevent one injury through targeted intervention decreased dramatically with increasing risk categories, from 42.3 in the low-risk group to 4.2 in the very high-risk group, demonstrating substantial efficiency gains through risk-based targeting.

Table 6 Risk stratification outcomes and screening efficiency metrics.

Cost-effectiveness analysis revealed compelling economic benefits of implementing the ML-based screening program. Table 7 presents the comprehensive economic evaluation across different implementation scenarios. Under base-case assumptions, the program would yield net savings of €215,800 per club season, accounting for screening costs (€85,000), intervention programs (€142,000), and prevented injury costs (€539,450). The break-even point occurred at 3.8 prevented injuries per season, well below the expected prevention rate of 8.2 injuries based on intervention efficacy data from the prospective cohort. It should be noted that these risk reduction and cost-effectiveness outcomes are derived from observational implementation data rather than randomized controlled trial evidence, which limits the strength of causal inference regarding intervention effectiveness.

Table 7 Cost-effectiveness analysis across implementation scenarios.

The clinical decision algorithm integrated multiple factors beyond injury probability, including player availability, competition schedule, and intervention feasibility. Strength rebalancing protocols included unilateral resistance exercises targeting the weaker limb (e.g., single-leg Romanian deadlifts, Nordic hamstring curls, single-leg press), eccentric strengthening, and neuromuscular control exercises. Intensity and volume were progressively adjusted based on weekly asymmetry reassessment. Table 8 outlines the individualized intervention protocols stratified by risk category and primary risk drivers. Players in the high-risk category underwent immediate strength rebalancing protocols, with training load modifications implemented within 48 h of identification. The individualized approach achieved 68% greater compliance compared to universal prevention programs, with players reporting enhanced confidence in the evidence-based rationale. Real-time risk monitoring enabled dynamic adjustment of prevention strategies. Within four weeks of intervention initiation, 73% of high-risk players (53/72) showed meaningful risk reduction (defined as > 20% decrease in predicted injury probability) while maintaining performance metrics throughout the intervention period. This intervention analysis was conducted within the development phase (2022–2023 season) as an observational component of the prospective cohort study.

Table 8 Risk-based intervention protocols and outcomes.

Discussion

This study demonstrates the integration of machine learning algorithms with isokinetic strength asymmetry assessment for predicting non-contact injuries in elite football players, building upon and extending previous work in sports injury prediction. Our findings demonstrate that the ensemble ML model achieved superior predictive performance (AUPRC: 0.759) compared to both traditional statistical approaches and a simple baseline model (AUPRC: 0.589). Strength asymmetry was identified as the primary modifiable risk factor, contributing 18.4% to overall model performance⁴⁰. The 3.2-fold increased injury hazard observed in players with knee flexor asymmetry exceeding 15% aligns with previous biomechanical studies. However, our ML approach revealed complex non-linear interactions between asymmetry patterns and training load that conventional analyses have failed to capture⁴¹.

Our findings complement and extend previous machine learning approaches in football injury prediction. Rossi et al.⁴⁰ demonstrated effective GPS-based injury forecasting with similar discrimination performance, while Rommers et al.²⁷ successfully applied ML to youth players using neuromuscular performance data, achieving comparable predictive accuracy through different methodological approaches. Our study specifically focuses on isokinetic strength assessment in adult professionals, extending these previous efforts by incorporating high-resolution strength asymmetry profiles. Valle et al.⁴² demonstrated machine learning effectiveness for hamstring injury prediction using the MLG-R classification system, showing similar performance levels though with different predictor variables. This suggests that multiple methodological pathways to effective injury prediction exist. Recent work by Kekelekis et al.⁴³ demonstrated that hip adductor/abductor strength ratios could predict groin injuries in male soccer players using k-nearest neighbor algorithms combined with logistic regression. Additionally, Kekelekis et al.⁴⁴ developed a machine learning-assisted prediction model for hamstring injuries achieving moderate discrimination (AUC = 0.68) and identifying hip abduction strength as a significant predictor. These studies, alongside our findings, collectively support the utility of integrating strength-based metrics within machine learning frameworks for injury prediction. Our comprehensive strength asymmetry profiling approach allows for position-specific injury pattern recognition that extends beyond traditional isolated strength ratio assessments.

The dose-response relationship between strength asymmetry and injury risk provides preliminary evidence for establishing clinical thresholds in professional football. Our finding that 82.4% of injuries occurred in players with baseline asymmetry above 10% corroborates recent systematic reviews⁴⁵, while extending previous work by demonstrating velocity-dependent asymmetry patterns. The progressive increase in asymmetry from 60°/s to 300°/s testing velocities suggests that high-speed strength assessments may better reflect the mechanical demands of match play⁴⁶, where 76.5% of hamstring injuries occurred during sprinting activities. This velocity-specific risk stratification represents a potential advancement over static strength ratios traditionally employed in sports medicine. However, it should be acknowledged that the relationship between bilateral asymmetry and injury risk remains contested in the literature. Afonso et al.⁴⁷ conducted a narrative review concluding that pre-existing asymmetries do not appear to increase injury risk, noting that asymmetry assessments are highly task-, metric-, and individual-specific, with significant temporal fluctuations in both magnitude and direction. Our findings of significant associations may reflect the integration of multiple muscle groups and testing velocities within a machine learning framework rather than isolated measures. Population-specific characteristics of elite professional footballers may also contribute to these associations. These discrepancies highlight the need for standardized assessment protocols and population-specific validation. While isokinetic testing provides the gold standard for strength assessment as highlighted by Asimakidis et al.²⁸, practical limitations present significant barriers to widespread implementation. These include equipment costs exceeding €100,000, specialized training requirements for technicians, and time constraints of 45–60 min per assessment, which must be critically acknowledged when considering clinical translation.

The superior performance of ensemble methods compared to individual algorithms reflects the multifactorial nature of sports injuries⁴². Our SHAP analysis revealed previously unrecognized interaction effects, particularly between strength asymmetry and acute: chronic workload ratios, suggesting that biomechanical vulnerabilities may be amplified under conditions of training stress⁴⁸. This finding has important implications for periodization strategies, as evidenced by the 4.87-fold injury risk in players exhibiting both high asymmetry and elevated ACWR. The temporal patterns observed, with injury peaks during pre-season and congested fixture periods (defined as periods with ≥ 2 matches per week, typically occurring during domestic cup competitions overlapping with league schedules), further emphasize the need for dynamic risk assessment. Static screening protocols may be insufficient to capture these temporal variations⁴⁹.

The clinical translation framework developed in this study begins to address a critical gap between predictive modeling and practical implementation. The stratified intervention approach achieved 73% risk reduction within four weeks, suggesting the potential feasibility of precision medicine approaches in sports settings. However, these observational findings require validation through randomized controlled trials⁵⁰. The economic analysis suggesting €215,800 net savings per club season provides preliminary evidence for potential cost-effectiveness, though these projections are based on observational intervention data and require validation across diverse club settings and competitive levels⁵¹. The clinical implications of model misclassification require careful consideration in any implementation strategy. False positives (30–35% of high-risk classifications) may result in unnecessary interventions, potential performance disruption, and inefficient resource allocation, while false negatives (approximately 25% of actual injuries missed) represent genuinely high-risk players who receive inadequate preventive attention. The resource allocation decisions must balance these competing risks against available medical and coaching resources, requiring careful consideration of the clinical consequences of both overtreatment and undertreatment.

The practical implementation of our approach faces substantial real-world constraints that extend beyond the controlled research environment. Data collection requires 45–60 min individual testing sessions with specialized personnel and equipment maintenance. Staff compliance varies significantly between clubs due to differences in medical team expertise, available time, and institutional priorities. Resource variability means that while top-tier organizations may implement comprehensive screening, smaller clubs face challenges with equipment costs (€80,000-120,000) and ongoing personnel requirements. Monthly retesting for dynamic risk monitoring demands substantial ongoing resource commitment that may not be sustainable across all competitive levels.

Several limitations warrant consideration that future research can address. Importantly, the predictive validity of isokinetic strength testing for injury risk remains controversial. Van Dyk et al.⁵² demonstrated in a large 4-year cohort study of 614 professional soccer players that isokinetic hamstring and quadriceps strength deficits represent only weak risk factors for hamstring strain injuries. Their findings showed area under the curve values of 0.54–0.56, indicating limited discriminative ability and small effect sizes (d < 0.2). Similarly, Green et al.⁵³ conducted a systematic review and meta-analysis concluding that isokinetic strength assessment offers limited predictive validity for detecting future hamstring strain risk, with over half of all variables showing moderate or strong evidence for no association with future injury. Our findings of stronger associations may reflect the integration of multiple muscle groups, velocities, and additional risk factors within a machine learning framework. However, these results should be interpreted cautiously given the broader literature context. The single-season follow-up may not capture longer-term injury patterns and seasonal variations in injury risk⁵⁴. Additionally, the reliance on a single preseason isokinetic assessment does not capture the dynamic nature of neuromuscular status across a competitive season. Practical constraints including equipment availability, player schedules during congested fixture periods, and the time-intensive nature of comprehensive testing precluded in-season reassessment. Future studies should consider incorporating longitudinal monitoring and additional neuromuscular metrics such as rate of force development (RFD), peak force, and time-under-tension variables, which have demonstrated relevance in recent injury surveillance models. Our temporal validation approach, using the same clubs in subsequent seasons, does not constitute true external validation across different organizations and populations, limiting confidence in model generalizability. The model’s performance degradation in external validation (AUPRC: 0.738 vs. 0.759) suggests potential overfitting despite cross-validation procedures, highlighting the need for continuous model updating⁵⁵. The male-only cohort limits generalizability to female players who exhibit different injury mechanisms and strength characteristics. Additionally, while our approach identified modifiable risk factors, the interventional component was observational rather than randomized, limiting causal inference regarding prevention efficacy.

The practical implementation barriers require acknowledgment. While isokinetic testing provides gold-standard strength assessment, equipment costs (€80,000-120,000), specialized training requirements, and time constraints (45–60 min per assessment) present significant barriers to widespread clinical implementation. The events-per-variable ratio of 1.9:1, while meeting minimum requirements, constrains more sophisticated modeling approaches.

Future research priorities should focus on addressing the most critical limitations identified in this study. Immediate priorities include conducting randomized controlled trials to establish causal relationships between risk stratification and injury prevention outcomes. Additional priorities include developing more accessible screening alternatives that maintain predictive accuracy while reducing implementation barriers, and validating models across diverse populations including female players and different competitive levels. The development of field-based strength assessments or wearable sensor integration could potentially democratize injury prediction beyond elite sporting environments, though this requires systematic validation against gold-standard isokinetic measures⁵⁶. Mid-term research goals should include multicenter collaborations to establish true external validation across different clubs and geographic regions. Additional goals include investigation of optimal retesting frequencies that balance predictive accuracy with practical constraints, and economic evaluation studies in diverse organizational settings to establish real-world cost-effectiveness. Integration of additional data streams, including three-dimensional motion capture analysis, wearable sensor-based movement monitoring, genetic markers, and psychological factors, may further enhance predictive accuracy⁵⁷, though the added complexity and cost must be carefully weighed against marginal predictive improvements. The development of real-time monitoring systems using wearable sensors could enable continuous risk assessment, moving beyond periodic screening toward dynamic injury prevention⁵⁸, though technical challenges regarding data quality and clinical integration remain substantial. Long-term objectives should focus on establishing evidence-based implementation protocols and developing simplified screening approaches that maintain reasonable predictive accuracy while dramatically reducing implementation barriers⁵⁹.

Conclusion

This study developed and evaluated a machine learning-based prediction model for non-contact injuries in elite football players, demonstrating improved predictive performance compared to conventional baseline approaches. The ensemble model achieved an AUPRC of 0.759 compared to 0.589 for traditional logistic regression, with lower limb strength asymmetry emerging as the primary modifiable risk factor. The identification of a clear dose-response relationship between asymmetry magnitude and injury risk, particularly for knee flexor imbalances exceeding 15%, provides preliminary evidence for clinical thresholds that warrant further validation. The velocity-dependent nature of asymmetry patterns highlights the importance of comprehensive strength assessment across multiple testing speeds.

The clinical implementation framework showed promising results in an observational setting, with risk-stratified interventions associated with 73% risk reduction within four weeks and suggesting potential economic benefits. However, these intervention outcomes are derived from observational rather than randomized trial data, limiting causal inference regarding prevention effectiveness. The high compliance rates and positive player feedback demonstrate the acceptability of personalized, data-driven prevention strategies in professional sports settings. The integration of explainable AI techniques enhanced clinical adoption by providing transparent risk factor identification and actionable insights for practitioners.

These findings suggest that machine learning may represent a valuable tool for injury risk assessment in professional football, potentially offering advantages in predictive accuracy, clinical utility, and resource optimization. While temporal validation confirmed model stability, true external validation across diverse populations and settings remains to be established. Implementation of this approach has the potential to contribute to injury burden reduction and player availability optimization, representing a step toward precision prevention strategies in elite sports medicine. Future randomized controlled trials are necessary to establish definitive causal relationships between risk stratification and injury prevention outcomes.