Design and development of a model for tennis elbow injury prediction and prevention using Artificial Neural Network (ANN) and Adaptive Neuro-Fuzzy Inference System (ANFIS) approaches | BMC Musculoskeletal Disorders

Tennis elbow affects individuals performing repetitive arm movements such as painters, plumbers, and factory workers. The condition results from chronic overuse of the extensor tendons, leading to microtears, inflammation, and tendon degeneration shown in Fig. 2. Understanding its multifactorial etiology is essential for developing effective clinical prevention and management strategies [25].

This study examined how three key factors—activity level (Factor A), equipment or technique (Factor B), and environmental conditions (Factor C)—influence the severity of tennis Experimental data were analyzed using a full central composite design (FCCCD), with injury outcomes categorized as per the scheme presented in Table 4. Table 5 summarizes the observed injury levels, and representative MRI images are shown in Figs. 3, 4, 5 and 6. The values for key factors for the Experimental tennis elbow injury level are shown in Table 6.

Table 7 provides a schematic summary that links the clinical assessment of injury severity with imaging findings and computational predictions made by the Artificial Neural Network (ANN). The injury severity is categorized into Mild, Moderate, and Severe; each associated with typical MRI findings used by clinicians for diagnosis.

Mild injuries show minimal structural abnormalities (e.g., minor edema, no ligament tear) and correspond to high biomechanical input values (e.g., flexion angle, torque, strength), which the ANN maps to low or no injury risk.

Moderate injuries reflect partial tissue damage visible in MRI scans and are linked to moderate biomechanical stress levels, leading the ANN to output a moderate injury risk range. Severe injuries demonstrate extensive damage (e.g., complete ligament rupture, joint effusion) and are associated with low mechanical input values, which the ANN interprets as a high injury risk. This mapping illustrates how objective MRI data and biomechanical inputs feed into the ANN to yield interpretable injury risk predictions, supporting clinical decision-making. Activity level (Factor A) was identified as the most influential factor. Both extremely low and high levels of physical activity were associated with more severe injury grades. This suggests that improper workload, either excessive or inadequate, may compromise tendon health, aligning with existing clinical evidence that sudden increases or decreases in load can provoke tendinopathy [26].

Equipment or technique (Factor B) significantly impacted injury levels as well. Participants using suboptimal tools or improper techniques showed higher rates of moderate to severe injury, reinforcing the importance of ergonomic practices and proper form in both occupational and sports settings. Environmental factors (Factor C) such as temperature, humidity, or surface condition—had a less consistent but noticeable effect. While not as influential as activity or technique, environmental extremes may exacerbate tendon stress, and should be considered in clinical advice, especially for athletes training in variable conditions [31].

The interplay of these three factors underscores the complex, multifactorial origin of tennis elbow. The most severe injuries were observed when all three factors were unfavorable, indicating a compounding effect. Conversely, when all three factors were optimized, the risk of injury significantly decreased. These insights can guide personalized intervention strategies, where modifiable risk factors are adjusted based on individual profiles [32].

Further, analysis using an Artificial Neural Network (ANN) described in Table 6 provided a high-accuracy model to predict injury risk based on biomechanical inputs such as flexion angle, elbow torque, and elbow strength. This predictive model can support early diagnosis and tailored rehabilitation, particularly when combined with clinical assessments [33].

While our findings offer valuable insights, limitations exist in terms of sample size and the range of contributing variables. Future studies with larger and more diverse populations are needed to validate these trends. Moreover, factors such as training intensity, individual biomechanics, and recovery protocols should be explored further. These results also suggest the potential for AI-supported tools to inform real-time risk assessment and targeted therapy for patients with repetitive strain injuries.

The Table 8 Present ANN analysis reveals a nuanced relationship between flexion angle and the risk of tennis elbow injuries. Flexion angle refers to the angle formed between the forearm and upper arm, representing the degree of bending at the elbow joint during activities. Across various runs of the ANN model, we observe that certain configurations of flexion angle, particularly at extreme values (10° and 140°), correspond to higher injury probabilities. This finding suggests that both excessive and insufficient flexion may contribute to increased strain on the elbow tendons, predisposing individuals to injury.

Elbow torque, the rotational force acting on the elbow joint, emerges as another critical determinant of tennis elbow susceptibility. Our results indicate that higher levels of torque, particularly at 90 N•m, are associated with elevated injury probabilities. This observation aligns with biomechanical principles, as increased torque imposes greater stress on the elbow structures, potentially exceeding their tolerance limits. Furthermore, interactions between flexion angle and torque may exacerbate injury risks, highlighting the importance of considering multiple biomechanical parameters concurrently. Elbow strength, representing the tensile capacity of the elbow tendons and muscles, emerges as a key modulator of injury likelihood. Higher strength values, indicative of greater tissue robustness, are generally associated with reduced injury probabilities across our ANN simulations. Conversely, lower strength levels correspond to heightened injury risks, particularly when combined with unfavorable flexion angles and torque magnitudes. These findings underscore the significance of muscular integrity in safeguarding against tennis elbow injuries and underscore the importance of strength training in injury prevention protocols. One of the notable aspects of our analysis is the identification of complex interactions and synergistic effects among flexion angle, torque, and strength in influencing injury outcomes. Certain combinations of these variables exhibit nonlinear relationships, wherein the cumulative effect on injury probability surpasses the individual contributions of each parameter. For instance, configurations involving extreme flexion angles, high torque, and compromised strength manifest disproportionately elevated injury risks. Few biomechanical factors can generate insights for proper injury prevention strategies.

The findings from our ANN analysis have significant implications for the development of targeted interventions aimed at preventing and managing tennis elbow injuries. By elucidating the intricate interplay between flexion angle, torque, and strength, clinicians and coaches can tailor exercise regimens, ergonomic modifications, and technique refinement to mitigate injury risks. Furthermore, the predictive capabilities of ANNs enable the identification of high-risk scenarios, facilitating early intervention and personalized care for individuals susceptible to tennis elbow. Based on this study provides valuable insights into tennis elbow pathogenesis without limitations. Factors like sample size and diversity will affect The ANN model’s predictions and also quality of the datasets. Additionally, the inherent complexity of musculoskeletal dynamics necessitates ongoing refinement and validation of predictive models. Future research endeavors may explore additional biomechanical parameters, incorporate real-time monitoring technologies, and employ longitudinal study designs to enhance the predictive accuracy and clinical utility of injury risk assessment tools. Figure 2 shows the ANN performance charts. The performance evaluation of the Artificial Neural Network (ANN) model on a regression task reveals promising results across multiple datasets. Beginning with the training data, the scatter plot depicts a strong positive correlation between predicted and actual values, indicated by a correlation coefficient (R value) of 0.9868. This alignment around a straight line with a slight positive slope signifies the model’s adeptness in capturing underlying patterns within the training data, showcasing its ability to learn complex relationships effectively.

Also, for the validation dataset, the scatter plot illustrates an even tighter clustering of data points along a perfect straight line (Y = T), with an exceptionally high correlation coefficient of 0.99999. This near-perfect fit indicates that the model generalizes well to unseen data, mitigating the risk of overfitting to the training dataset. Similarly, the scatter plot for the testing data demonstrates a notable alignment of data points around a straight line, albeit with slightly more dispersion compared to the validation dataset.

Despite this, the correlation coefficient remains high at 0.99813, reaffirming the model’s strong performance on unseen data and its ability to capture underlying trends accurately. Combining all datasets further strengthens the model’s performance, with data points clustering tightly around a straight line and a correlation coefficient of 0.99009. The consistency observed across all datasets underscores the model’s robustness and reliability in maintaining accuracy across diverse data instances. Additionally, the nearly identical equations for the best fit lines for both training and testing datasets highlight the model’s generalizability. However, a slight bias present in the equations, indicated by non-zero constant values, suggests that the model’s predictions tend to be slightly higher than the target values on average. Despite this minor bias, the model’s overall performance is commendable, showcasing its efficacy in predictive modeling tasks. Further refinements to address any biases observed would enhance the model’s accuracy and broaden its applicability across various domains. In short, our ANN-based analysis provides valuable insights into the biomechanical determinants of tennis elbow injuries, emphasizing the roles of flexion angle, elbow torque, and strength in injury susceptibility. Also, Fig. 7 illustrates better agreement between Experimental and FLS results.

In the present work, Fuzzy Logic Systems (FLS) for injury assessment offers a novel approach to understanding the multifaceted interplay between biomechanical parameters and injury susceptibility. Here, we discuss the insights gleaned from applying FLS to analyze the relationship between flexion angle, elbow torque, elbow strength, and the likelihood of tennis elbow injuries. Fuzzy Logic Systems enable the modeling of complex, uncertain relationships inherent in musculoskeletal biomechanics, providing a framework for integrating imprecise inputs and generating meaningful outputs. By incorporating linguistic variables and fuzzy rules, FLS can capture the inherent ambiguity and variability present in injury risk assessment. The FLS analysis reveals nuanced interactions between flexion angle, elbow torque, and strength in influencing tennis elbow susceptibility. Flexion angle, representing the degree of elbow joint bending during activity, emerges as a significant determinant of injury risk. In the present work, total 27 fuzzy rules are defined with membership functions as mentioned in Fig. 7. Table 6 illustrates the results obtained from fuzzy logic system (Fig. 8).

The fuzzy inference engine delineates optimal and suboptimal ranges of flexion angle, elucidating critical thresholds beyond which injury likelihood escalates. Similarly, elbow torque, reflecting the rotational force acting on the elbow joint, exhibits a nonlinear relationship with injury probability. FLS elucidates torque thresholds beyond which tendon strain surpasses tolerable limits, predisposing individuals to tennis elbow. Moreover, the role of elbow strength, indicative of tendon and muscle integrity, is dynamically assessed by the fuzzy inference system, highlighting its protective effect against injury occurrence. FLS facilitates the incorporation of linguistic variables such as “low,” “medium,” and “high” to characterize the degree of injury risk associated with specific combinations of biomechanical parameters. Linguistic terms encapsulate the qualitative aspects of injury susceptibility, enhancing the interpretability of FLS-generated outputs and facilitating informed decision-making. The FLS-generated injury levels provide valuable insights into the complex interplay of flexion angle, torque, and strength in tennis elbow pathogenesis. Also, Fig. 9 illustrates better agreement between Experimental and FLS results. By delineating fuzzy membership functions and rule-based inference mechanisms, FLS elucidates subtle variations in injury likelihood across diverse input configurations.

Contributing factors and predicting the injury level will affect effective prevention and treatment strategies. The model presenting the relationship between inputs and outputs is The Adaptive Neuro-Fuzzy Inference System. We can have the table showing the relationship between flexion angle, elbow torque, elbow strength, and the injury level as predicted by ANFIS. Three key parameters were considered in this study: flexion angle (measured at 10°, 80°, and 140°), elbow torque (measured at 10, 50, and 90 Nm), and elbow strength (measured at 88, 113, and 142 MPa). The results Table 9 displays 20 runs with varying conditions and the corresponding injury levels as predicted by the ANFIS model. Table 6 illustrates the results obtained from fuzzy logic system. The injury levels varied even under consistent conditions (A = 0, B = 0, C = 0), with runs 1, 2, 3, 6, 17, and 19 showing levels ranging from 0.95 to 2.88. This variability suggests that other underlying factors influence injury levels, indicating the complexity of predicting tennis elbow injuries. When examining the impact of the flexion angle (A), negative flexion angles (A = −1) in runs 4, 7, 10, 11, and 20 showed varying injury levels from 0.91 to 2.87. Run 7, with B = 1 and C = 1, showed a low injury level of 0.91, suggesting a potential protective effect of higher torque and strength at this angle. Positive flexion angles (A = 1) in runs 5, 8, 12, 14, and 18 showed injury levels ranging from 1.25 to 2.91. Run 8, with B = 1 and C = 1, indicated a lower injury level (1.25), implying that increased torque and strength might mitigate the risk at this angle.

The Table 10 represents the results for Experimental and ANFIS values, for elbow torque (B), negative torque (B = −1) in runs 10, 11, 12, 14, and 16 resulted in injury levels ranging from 0.92 to 2.87. The lower levels observed in runs 12 and 16 (1.00 and 0.92, respectively) suggest that reduced torque could decrease injury risk depending on other conditions. Positive torque (B = 1) in runs 7, 8, 14, 15, and 18 exhibited injury levels from 0.91 to 2.91 (Table 11).

The variability in these results indicates that increased torque alone does not consistently predict injury level, highlighting the need for a multifactorial approach. Regarding elbow strength (C), negative strength (C = −1) in runs 11, 12, 13, 18, and 20 showed injury levels ranging from 1.00 to 2.91. Run 13 (A = 0, B = 0) demonstrated a low injury level of 1.00, suggesting that lower strength might not always correlate with a higher injury risk. Positive strength (C = 1) in runs 7, 8, 9, 10, and 14 exhibited injury levels from 0.91 to 2.87. Also, Fig. 10 illustrates better agreement between Experimental and ANFIS results. The graph comparison between ANN, FLS and ANFIS is illustrated into Fig. 11.

Medical interpretation of results based on biomechanical variables

The ANN, FLS, and ANFIS analyses presented in this study provide insight into the biomechanical mechanisms contributing to lateral epicondylitis (tennis elbow) — a degenerative condition marked by micro tears at the origin of the extensor carpi radialis brevis (ECRB) tendon [34].

Flexion angle

Extreme elbow flexion angles (10° and 140°) were associated with a higher risk of injury in the models. Clinically, such positions either cause excessive stretch or compression of the ECRB tendon during repetitive wrist extension, a known mechanism for lateral epicondylitis [35]. Research by Bisset et al. (2006) confirms that abnormal joint positioning under load is a major factor in tendon overload syndromes [36].

Elbow torque

Elbow torque represents the rotational force applied during forearm movements. High torque values (e.g., 90 N·m) resulted in higher predicted injury probabilities. This reflects prior studies showing that increased torque amplifies stress at the lateral epicondyle, especially in occupations or sports involving forceful grip and wrist extension [37, 38].

Elbow strength

Elbow strength, as a surrogate for tendon and muscle integrity, was inversely related to injury risk. Low strength values corresponded to higher predicted injuries, aligning with studies showing that insufficient muscular support increases tendon strain, while strength training enhances tendon load tolerance [39].

Multifactorial interaction

Our models reveal that combinations of extreme flexion, high torque, and poor strength magnify injury risk — a hallmark of cumulative overuse injuries [40]. For instance, high torque applied at extreme flexion angles on a weak tendon can exceed its physiological loading threshold, accelerating tendinopathy (Fig. 12) [41].

The predictive models thus align with clinical understanding that repetitive strain, poor technique, and reduced tissue resilience are primary contributors to tennis elbow. These insights support the development of personalized prevention protocols, such as ergonomic adjustments, progressive resistance training, and early load management strategies.

Data splitting and isolation

The dataset collected from Suyog Imaging and Physiotherapy Centre (Mehsana, Gujarat) was divided using a stratified approach to preserve class distribution, ensuring no overlap between training (70%), validation (15%), and testing (15%) subsets.

Cross-validation

To verify model generalizability, we conducted fivefold cross-validation on the training data. The consistently high R² scores across folds confirmed that the model performance was stable and not confined to a particular data split.

Regularization and early stopping

In the ANN, we applied L2 regularization and early stopping to prevent overfitting. The training was stopped when validation loss stopped decreasing for 10 consecutive epochs.

Evaluation on unseen data

Final model metrics were calculated only on the test dataset, which was isolated from the training and tuning process. The test R² of ~0.99, although high, was consistently reproduced across different random seeds and data reshuffles.

Model complexity control

For ANFIS, the number of fuzzy rules was restricted (27 rules), and for ANN, the architecture was shallow (only one hidden layer with 3 neurons). This reduced the risk of memorization and over-parameterization.

Interpretability and external validation

The outputs from the ANN and ANFIS models were further validated against experimental observations, as shown in Figs. 7 and 9, confirming physiological plausibility and external agreement with domain-specific patterns.

Although high R² values (≈0.99) were obtained across training, validation, and testing datasets, we applied multiple safeguards to prevent overfitting or data leakage. These include careful data partitioning, 5-fold cross-validation, regularization, and early stopping. Moreover, the models were deliberately designed with limited complexity to avoid overfitting, and the performance was further validated against experimental data. The consistency and physiological alignment of the results suggest that the models are capturing meaningful patterns in the biomechanical data rather than overfitting noise.

Model interpretability

While Fuzzy Logic Systems (FLS) offer a degree of interpretability through rule-based reasoning, we acknowledge that the current analysis only presents rule structures and outputs without a deep dive into individual rule interactions or their relative impact on the final prediction. For example, among the 27 fuzzy rules employed, those involving moderate torque and mid-range flexion angles consistently contributed to lower injury level outputs—implying a stabilizing role of mid-level loading conditions. However, a systematic ranking or contribution analysis of each rule was not performed.

Moreover, modern model-agnostic interpretability techniques such as SHAP (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations) were not used in the present study. These methods can provide feature attribution insights for black-box models like ANN or ANFIS by quantifying the contribution of each input feature (e.g., flexion angle, torque, strength) toward a specific prediction. Incorporating these tools in future work could enhance transparency and trust in biomechanical injury prediction systems, particularly for clinical end-users.We recognize this as a limitation and an opportunity for future expansion of the interpretability framework of our models.

Limitations

While the proposed ANN, FLS, and ANFIS models demonstrate high predictive accuracy for tennis elbow injury severity, several limitations must be considered:

1. 1.

   Generalization beyond lab settings: The data used for model training and validation were acquired in controlled settings at Suyog Imaging and Physiotherapy Centre (Mehsana, Gujarat). This controlled environment may not capture the full variability encountered in real-world clinical or occupational settings, limiting model generalizability.

2. 2.

   Potential Biases due to uncontrolled factors: Demographic and Factors such as age, gender, dominant hand, occupational history, and previous injury were not systematically balanced in the dataset. These factors are known to influence musculoskeletal injury susceptibility and recovery and may introduce unquantified bias into the models.

3. 3.

   Lack of external validation: The models were trained and tested on the same institutional dataset without validation against the independent external cohort. This limits our ability to assess their performance across different populations or imaging protocols.

4. 4.

   Small to moderate sample size: While the dataset covered a broad range of biomechanical inputs (e.g., flexion angle, torque, strength), the number of distinct subjects was relatively modest. These constraints the statistical readiness and increased overfitting risk. Future studies should expand sample size and diversity to enhance strengthen reliability and reduce model variance.

5. 5.

   Simplified injury thresholds: Although clinically relevant thresholds were incorporated, the biomechanical complexity of tendon injury progression (e.g., involvement of neural feedback, compensatory movement patterns) was simplified in the current model.

Future work will focus on addressing these limitations by incorporating demographic diversity, longitudinal tracking, wearable sensor data, and validation across multi-center clinical datasets.

Ethical and safety considerations

The deployment of predictive models like ANN, FLS, and ANFIS in clinical or amateur sports contexts must be approached with caution. While these models can assist in identifying individuals at higher risk of tennis elbow injuries based on biomechanical parameters, they are not a substitute for clinical diagnosis or medical judgment. Misinterpretation of model output, especially in non-expert settings—could lead to incorrect decisions about training load, therapy, or rest, potentially causing harm.

Moreover, there are ethical considerations around the use of personal biomechanical data, including informed consent, data privacy, and the risk of over-reliance on algorithmic predictions. Ensuring that athletes or patients understand the limitations of such tools is essential. Any application of these models in practice should involve medical professionals to interpret results within the broader clinical and individual context.

Future iterations of the system should incorporate transparent communication strategies, include clinical validation studies, and implement user-interface safeguards to prevent misuse or misinterpretation. Ethical review board (IRB) clearance and clear data handling protocols should be integral to such applications.