A SHAP-based machine learning approach to decoding the relationship between running posture features and tibial load in novice runners with MTSS: towards accurate and intelligent rehabilitation strategies Authors: Ting Fan, Xiaomin Wu, Zeyi Zhang, Youping Sun Published: 2026 ABSTRACT BackgroundThis study systematically investigates the relationship between running posture and tibial load in novice runners using machine learning methods from the field of artificial intelligence. The aim is to provide scientific evidence for the precise prevention and personalized rehabilitation of Medial Tibial Stress Syndrome (MTSS).MethodsA total of 174 novice runners diagnosed with MTSS were included in the study. Kinematic and kinetic data were collected during their running trials, and tibial load was subsequently calculated. Three machine learning models—Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Regression (SVR)—were employed, combined with the SHapley Additive exPlanations (SHAP) framework, to analyze the relationship between running posture and tibial load in this population.ResultsThe predictive performance of the XGBoost model outperformed both the RF and SVR models. SHAP analysis revealed that tibial load predicted by the model significantly increased when the vertical center of mass (COM) position was below 927 mm, the ankle inversion angle exceeded 2.2°, and the knee internal rotation angle surpassed 2.8°.ConclusionFor novice MTSS runners, a lower COM position, greater ankle inversion, and increased knee internal rotation were associated with elevated tibial loading and showed clear nonlinear relationships with identifiable threshold effects. These findings highlight complex nonlinear interactions among multiple biomechanical factors and demonstrate the value of SHAP in quantifying such effects. Clinically, optimizing vertical COM control and strengthening the ankle evertor and knee external rotator musculature may help reduce tibial loading during running and lower the risk of MTSS development or recurrence.Trial registrationThe entire study protocol was approved by the Medical Ethics Committee of East China Normal University (Approval No. 2024-28). Introduction Running is a low-cost and easily accessible form of exercise that can significantly promote metabolic health, improve cardiovascular function, enhance postural balance, and reduce all-cause mortality [1, 2]. As a result, an increasing number of individuals are incorporating running into their routines [3]. However, as the number of people participating in running continues to rise, the incidence of running-related lower limb injuries has also increased, particularly among novice runners who lack systematic training and scientific guidance [4, 5]. Epidemiological data indicate that the incidence of lower limb injuries in novice runners is 49.9% [6]. Among these, Medial Tibial Stress Syndrome (MTSS) is one of the most common injuries, with an incidence rate of 15% [7]. Furthermore, this injury has a notably high recurrence rate, reaching up to 79.3% [8]. Recurrent injuries not only undermine the exercise confidence and training effectiveness of novice runners but may also have lasting effects on their mental health and quality of life. Therefore, effectively reducing the recurrence risk of MTSS in novice runners has become a critical issue in the fields of sports science and rehabilitation. Existing research generally suggests that excessive tibial load is a direct cause of MTSS recurrence [9]. Abnormal kinematic characteristics during running are often regarded as key factors contributing to increased tibial load [10]. For example, Pohl et al. [10]found that excessive rearfoot eversion during running leads to premature fatigue of the tibial posterior muscles, which in turn increases medial tibial stress. Moen et al.‘s comparative study [11]indicated that an increase in ankle plantarflexion angle enhances the likelihood of forefoot strike during running. However, previous studies have confirmed that, compared to rearfoot strike, forefoot strike significantly increases posterior-medial tibial strain, thereby elevating tibial load [12]. Additionally, other studies have suggested that reduced hip internal rotation may also increase tibial load and trigger MTSS [11], although Burne et al. [13] reached the opposite conclusion. It is evident that the running posture contributing to increased tibial load in novice MTSS runners remain a subject of debate. This uncertainty is likely due to the fact that previous studies have primarily relied on univariate analyses or linear models, which struggle to capture the complex, nonlinear, and multidimensional interactions between gait characteristics and tibial load. As a result, these methods have been insufficient in providing an accurate and comprehensive understanding of the recurrence mechanisms of MTSS. In contrast, machine learning models from the field of artificial intelligence can simultaneously handle multiple variables, nonlinearities, and interaction effects, effectively addressing the limitations of traditional statistical methods [14]. Furthermore, machine learning models can be combined with methods such as the SHapley Additive exPlanations (SHAP) framework to provide in-depth analysis of the model’s decision-making process. This enables the identification of key running posture that significantly contribute to tibial load, and allows for the quantification of their influence pathways and strengths, thereby offering precise insights into the recurrence mechanisms of MTSS [15, 16]. However, research on this topic is currently limited, leaving novice runners at a higher risk of recurrence over the long term. Moreover, most existing evidence is derived from studies on professional or experienced runners, with very few systematic studies focusing on novice runners. Previous studies have shown that novice runners, lacking running experience, appear to be more susceptible to MTSS [17], and there are significant differences in running posture characteristics between novice and experienced runners [18, 19]. For example, compared with experienced runners, novice runners often exhibit greater ankle plantarflexion moments, higher instantaneous vertical loading rates, and larger ranges of motion (ROM) at the hip, knee, and ankle joints. These differences in injury incidence and biomechanical characteristics may result in novice runners exhibiting mechanical patterns during the occurrence or recurrence of MTSS that differ from those of experienced runners [18, 19]. However, the mechanisms underlying MTSS recurrence in novice runners remain insufficiently understood. In particular, there remains a lack of studies employing advanced data-driven approaches to systematically reveal the complex interactions among running posture characteristics and their influence on tibial loading in novice runners with MTSS. Therefore, the purpose of this study was to employ machine learning algorithms to systematically investigate the relationship between running posture characteristics and tibial loading in novice runners with MTSS, and to explore potential running posture patterns associated with increased tibial loading. This may provide more comprehensive and rigorous theoretical support for the precise prevention and individualized rehabilitation of MTSS. We hypothesized that multiple running posture characteristics interact in a complex and nonlinear manner to influence tibial loading in novice runners with MTSS, thereby contributing to the recurrence of the condition. Research methods Study participants This study recruited novice runners from East China Normal University through multiple recruitment methods, including WeChat public account posts, online recruitment notices, and offline poster postings. The inclusion criteria were as follows: 1) age between 18 and 30 years; 2) no prior systematic running training, defined as the absence of structured or formal running programs, and regular running exercise for less than 24 months; 3) an average weekly running distance exceeding 20 km in the 4 weeks prior to the study [20, 21], This criterion was adopted to ensure that participants had sufficient and relatively stable running exposure for investigating MTSS-related risk, and was also broadly consistent with previous literature defining novice runners as individuals with relatively short regular running history despite maintaining a minimum weekly running volume of 20 km [21]; 4) a history of MTSS within the past 6 months, with no current pain symptoms at the time of enrollment. The diagnostic criteria for MTSS include the following: 1) pain in the medial tibial region during running or weight-bearing activities, with diffuse tenderness on palpation covering an area of at least 5 cm; 2) the pain persists for at least 3 weeks; 3) stress fractures or other bony abnormalities are excluded through X-ray imaging [22]. The entire study protocol was approved by the Medical Ethics Committee of East China Normal University (Approval No. 2024-28), and all participants signed a written informed consent form approved by the ethics committee. Data collection To ensure the accuracy and reproducibility of the experimental data, the research team calibrated and standardized the testing environment prior to the participants’ arrival in the laboratory. The environmental conditions were as follows: indoor temperature maintained at 22 ± 2 °C, relative humidity controlled between 40 and 60%, and standard white lighting with a brightness of 300–500 lx. Additionally, the global coordinate origin was clearly defined. Upon entering the testing area, participants’ basic information (e.g., age) was recorded, followed by measurements of height, weight, lower limb length, knee width, and ankle width. After these procedures, participants performed a 5-minute light jogging warm-up in the testing area to activate the relevant muscle groups and stabilize their movement patterns. After the warm-up, the testers adhered reflective markers to key anatomical landmarks on the participants, including the anterior superior iliac spine, posterior superior iliac spine, knee, and ankle, following the Plug in Gait standard protocol. A total of 39 reflective markers were used [23]. During the formal testing phase, participants were instructed to perform five valid running trials on a 2 × 10 m track, with a 1-minute rest interval between each trial. Prior to the formal data collection, participants were asked to run one lap around the laboratory to reach a natural running rhythm and were then directed to continue running into the testing runway without stopping. This continuous running procedure ensured that participants had already reached a stable running pace before entering the testing area, allowing the recorded gait data to reflect steady-state running rather than the acceleration phase. Kinematic parameters during running were captured using the Vicon V2.2 three-dimensional motion capture system (Vicon Motion Systems Ltd., UK, 200 Hz) [23]. Ground reaction forces (GRF) were measured using an AMTI BMS400600 force platform (Advanced Mechanical Technology Inc., USA, 1000 Hz) [24]. To ensure accurate placement of the affected foot on the force platform, participants were allowed to adjust their starting position according to their stride length. Data processing Kinematic and kinetic data processing This study focuses on the biomechanical data of the MTSS-affected leg during the stance phase of running. The stance phase is defined as beginning when the GRF exceeds 20 N and ending when the GRF falls below 20 N within a complete gait cycle. The raw trajectories of reflective markers and GRF data were processed using the Vicon system (Vicon Motion Systems Ltd., UK). All data were filtered using a fourth-order zero-phase Butterworth low-pass filter to remove high-frequency noise, with cutoff frequencies set at 20 Hz for marker trajectories and 50 Hz for GRF data [25]. Joint angles were calculated using the “flexion-extension—abduction/adduction—internal/external rotation” Cardan rotation sequence, which describes the three-dimensional position of the distal segment relative to the proximal segment [26]. Tibial load calculation The formula for calculating tibial load is as follows [27]:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{tibia}\left(t\right)=F_{ext}\left(t\right)+F_{int}\left(t\right)$$\end{document} Where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{{\mathrm{ext}}}}$$\end{document} represents the projection of the GRF along the tibial axis, calculated using the following formula:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{{\mathrm{ext}}}}(t)=\left| {{{\overrightarrow {{\mathrm{GRF}}} }^*}(t)} \right| \cdot \cos \beta (t)$$\end{document} In this equation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overrightarrow {{\mathrm{GRF}}} ^*}$$\end{document}represents the GRF after low-pass filtering, and the angle between the force and the tibial axis is denoted as cos β (determined by the biomechanical model). The formula for calculating the internal force component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{{int}}}$$\end{document} is as follows:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{int}\left(t\right)=\frac{M_{ankle}\left(t\right)}{0.05}$$\end{document} Where\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${M_{{ankle}}}(t)$$\end{document} represents the moment of force generated by the GRF at the ankle joint, calculated using the formula:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_{ankle}(t)=COP_{ankle}\cdot GRF_z(t)$$\end{document} In this equation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$COP_{ankle}$$\end{document} denotes the anterior-posterior position relative to the reference point at the ankle joint center, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{GRF}_z}(t)$$\end{document} represents the vertical component of the GRF. Machine learning modeling Prior to modeling, the dataset was carefully reviewed for completeness. No missing data were observed in the final cleaned dataset used for analysis, following the exclusion of any trials with marker occlusion or force plate clipping [28]. Therefore, no imputation was necessary. The dataset was then split into training and testing sets in an 8:2 ratio using the train_test_split function from the scikit-learn library. To investigate the relationship between running posture characteristics and tibial load, this study developed three types of nonlinear regression models: XGBoost, Random Forest (RF), and Support Vector Regression (SVR). Model training and implementation were conducted using the scikit-learn library [29, 30]. To enhance prediction performance, optimal hyperparameters were identified through GridSearchCV combined with 5-fold cross-validation [31]. Model performance was evaluated using several metrics, including mean absolute error (MAE), mean squared error (MSE), median absolute error (MedAE), explained variance score (EVS), coefficient of determination (R²), and 5-fold cross-validation [32]. The calculation formulas for all evaluation metrics are as follows:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MAE=\frac{1}{n}\sum\limits_{{i=1}}^{n} {|{y_i}} - {\mathrm{y}}_{{\mathrm{i}}}^{\prime }|$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MSE=\frac{1}{n}{\sum\limits_{{i=1}}^{n} {({y_i} - y_{i}^{\prime })} ^2}$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MedAE=median(|{y_{^{1}}} - y_{1}^{\prime }|,...,|{y_i} - y_{i}^{\prime }|)$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$EVS=1 - \frac{{\operatorname{var} ({y_i} - y_{i}^{\prime })}}{{\operatorname{var} ({y_i})}}$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R^2}=1 - \frac{{\sum\limits_{{i=1}}^{n} {{{({y_i} - y_{i}^{\prime })}^2}} }}{{\sum\limits_{{i=1}}^{n} {{{({y_i} - \bar {y})}^2}} }}$$\end{document} Where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{y}}_{{i}}^{{\prime\:}}$$\end{document} represents the predicted value, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{y}}_{{i}}$$\end{document} represents the actual value, var denotes variance, and n represents the sample size. To gain a deeper understanding of the model’s internal mechanisms, this study employed the SHAP method to assess the importance of each feature variable. Based on game theory principles, this method comprehensively characterizes the interactions between features and quantifies their marginal contributions to the model’s predictions, thereby revealing the nonlinear response characteristics of key variables [33–35]. It is important to note that the variable thresholds computed by SHAP represent sensitivity points at the model prediction level and should not be equated with clinically significant pathological thresholds. Finally, the SHAP analysis results were integrated into the Streamlit platform to develop an online personalized tibial load prediction tool. The platform includes modules for model introduction, parameter input, and result visualization, designed to meet the individualized assessment and precise early warning needs of runners. Results Participant selection process and demographic characteristics As shown in Fig. 1, a total of 197 amateur runners were initially recruited for this study. However, 23 participants did not complete all testing procedures for various reasons. Specifically, 17 participants voluntarily withdrew from the study due to personal reasons, and 4 participants sustained acute ankle sprains prior to the official testing, which had not fully healed. Additionally, during the data processing phase, it was found that two participants had issues with marker dropout. Considering the potential impact of these factors on the data accuracy, these participants were excluded from the analysis. Therefore, a final total of 174 participants’ biomechanical data were included in the analysis. The average age of the participants was 20.79 years, with a mean height of 1.78 m and a mean weight of 70.79 kg (Fig. 1; Table 1). The descriptive statistics of the primary biomechanical variables included in this study are summarized in Table 1, including the mean, standard deviation, minimum, and maximum values. These results provide an overview of the running posture characteristics of novice runners with MTSS (Table 1). Fig. 1Flowchart of patient inclusion Flowchart of patient inclusion Table 1Demographic characteristics of participants and descriptive statistics of relevant biomechanical variablesIndicatorMeanStandard deviationMaximumMinimumDemographics Age (years)20.792.4129.0018.00 Height (m)1.780.061.921.63 Weight (kg)70.798.9592.9544.65Primary Outcome Tibial loading (N/kg)15.394.6928.909.04Feature variable Ankle plantar/dorsiflexion angle(°)26.443.7934.1118.48 Ankle in/eversion angle(°)2.872.028.99-1.37 Hip ad/abduction angle(°)4.613.8218.23-6.08 Knee flex/extension angle(°)31.989.0259.883.62 Knee ad/abduction angle(°)9.828.2732.48-8.61 Knee in/external rotation angle(°)3.738.6320.54-16.42 M/L pelvic lean(°)-0.092.937.89-7.21 Vertical pelvic lean(°)-5.504.4910.43-16.96 Vertical trunk lean(°)-6.6638.17179.66-191.85 M/L COM position(mm)686.4381.75907.28463.81 Vertical COM position(mm)963.0338.541068.09851.90 Table 1Demographic characteristics of participants and descriptive statistics of relevant biomechanical variablesIndicatorMeanStandard deviationMaximumMinimumDemographics Age (years)20.792.4129.0018.00 Height (m)1.780.061.921.63 Weight (kg)70.798.9592.9544.65Primary Outcome Tibial loading (N/kg)15.394.6928.909.04Feature variable Ankle plantar/dorsiflexion angle(°)26.443.7934.1118.48 Ankle in/eversion angle(°)2.872.028.99-1.37 Hip ad/abduction angle(°)4.613.8218.23-6.08 Knee flex/extension angle(°)31.989.0259.883.62 Knee ad/abduction angle(°)9.828.2732.48-8.61 Knee in/external rotation angle(°)3.738.6320.54-16.42 M/L pelvic lean(°)-0.092.937.89-7.21 Vertical pelvic lean(°)-5.504.4910.43-16.96 Vertical trunk lean(°)-6.6638.17179.66-191.85 M/L COM position(mm)686.4381.75907.28463.81 Vertical COM position(mm)963.0338.541068.09851.90 Demographic characteristics of participants and descriptive statistics of relevant biomechanical variables Validation and selection of machine learning models As shown in Fig. 2A and B, the correlation coefficients between the features included in this study were all below 0.8, and the variance inflation factors (VIF) were all less than 5, indicating that there was no significant multicollinearity in the data, thus meeting the basic requirements for machine learning modeling. After modeling, we conducted a comprehensive evaluation of the model’s predictive performance, as shown in Fig. 2C and D. The XGBoost algorithm outperformed both the RF and SVR models across all metrics. Specifically, the model achieved a MAE of 0.039, a MSE of 0.006, a MedAE of 0.020, an EVS of 0.888, a R² of 0.886, and a 5-fold cross-validation score of 0.964. These results demonstrate that the XGBoost model exhibits significant advantages in both fitting accuracy and generalization ability, accurately capturing the underlying patterns in the target variable. Fig. 2Feature selection and model validation process. A Correlation analysis of feature variables; B VIF values of the feature variables; C-D Model validation results Feature selection and model validation process. A Correlation analysis of feature variables; B VIF values of the feature variables; C-D Model validation results Global and local interpretations of machine learning models As observed in Fig. 3A and B, the three running posture parameters that contribute most significantly to tibial load are, in order: a lower vertical center of mass (COM) position, a significant increase in the ankle inversion angle, and an increased knee internal rotation angle. Further analysis using SHAP dependence plots reveals that when the vertical COM position is below approximately 927 mm, the ankle inversion angle exceeds about 2.2°, and the knee internal rotation angle exceeds approximately 2.8°, tibial load shows a noticeable upward trend (Fig. 3C). Additionally, the left side of Fig. 3D shows a runner with higher tibial load. The key contributing factors to the elevated tibial load are a lower vertical COM position (approximately 875 mm) and a larger knee internal rotation angle (14.6°), among others. In contrast, the runner on the right side of Fig. 3D exhibits a lower tibial load. The primary reason for the reduced tibial load is the higher vertical COM position (approximately 990 mm) and a smaller knee internal rotation angle, which may even be in an externally rotated position (− 10.8°). Fig. 3Interpretability analysis results of the machine learning model based on the SHAP method. A–B SHAP summary plot: Evaluates the importance of each feature variable to the model’ s predicted output from a global perspective, with variables arranged in descending order of contribution; C SHAP dependence plot: Displays the effect trend of a single feature on the prediction result across its different value ranges; D SHAP waterfall plot: Comparison of two representative extreme individual cases, clearly revealing the key drivers of tibial load variation and their contribution pathways (high tibial load vs. low tibial load) Interpretability analysis results of the machine learning model based on the SHAP method. A–B SHAP summary plot: Evaluates the importance of each feature variable to the model’ s predicted output from a global perspective, with variables arranged in descending order of contribution; C SHAP dependence plot: Displays the effect trend of a single feature on the prediction result across its different value ranges; D SHAP waterfall plot: Comparison of two representative extreme individual cases, clearly revealing the key drivers of tibial load variation and their contribution pathways (high tibial load vs. low tibial load) Development of the online risk prediction system Based on the results of the machine learning model, this study developed an online personalized tool for assessing the relationship between running posture characteristics and tibial load in novice runners. This tool can assist in determining an individual’s potential risk of developing MTSS, offering significant practical value and potential for widespread application. The research team has deployed the tool on the web, and users can access and use it through the following link: https://tibia-ft-xgpxws4fggxvqtds8lddpm.streamlit.app/ (Fig. 4). Fig. 4Screenshot of the online calculator for assessing the relationship between running posture characteristics and tibial load Screenshot of the online calculator for assessing the relationship between running posture characteristics and tibial load Discussion Development and selection of machine learning models This study compares the predictive performance of three machine learning algorithms, revealing that XGBoost achieves the highest accuracy. Running-related biomechanical data are typically characterized by a multivariate structure, where different kinematic variables often exhibit complex nonlinear relationships and potential interaction effects. For example, vertical COM position, joint angles, and lower-limb control strategies may jointly influence tibial loading. Such characteristics require models capable of effectively capturing nonlinear patterns and interactions among variables. The gradient-boosting framework implemented in XGBoost iteratively improves predictions by sequentially fitting residuals, enabling the model to more effectively capture complex nonlinear relationships and interaction effects among multiple variables. This property makes it particularly suitable for biomechanical regression tasks involving complex movement patterns [31]. In contrast, Random Forest (RF), although also an ensemble learning method, primarily reduces variance by averaging the predictions of multiple decision trees. While this mechanism improves robustness, it may be relatively less flexible in representing intricate interactions among variables [36]. The SVR model, on the other hand, is heavily reliant on the selection of kernel functions and parameter tuning, which not only increases the complexity of hyperparameter optimization but also renders the model more sensitive to outliers [37]. Additionally, SVR requires solving a quadratic programming problem during training, leading to higher computational costs and making it inefficient for large-scale datasets, thus limiting its practical applicability [38]. In terms of preventing overfitting, XGBoost incorporates built-in regularization strategies during the model training phase, combining Lasso (L1) and Ridge (L2) penalty terms to effectively constrain model complexity and reduce the risk of overfitting on the training data [39]. This regularization strategy, together with the staged learning process of gradient boosting, enables the model to achieve a more stable balance between bias and variance. In contrast, although RF reduces variance by aggregating multiple regression trees, it lacks explicit regularization constraints, making it more prone to overfitting, particularly when dealing with noisy data. SVR, on the other hand, is highly sensitive to the choice of kernel function and hyperparameters [40], improper configurations often result in good fitting on the training set but significantly reduced performance on the test set [41]. Overall, considering the multivariate nature, nonlinear relationships, and potential interaction effects commonly present in running-related biomechanical data, the superior performance of XGBoost in this study can be attributed not only to its predictive capability but also to its structural advantage in modeling complex nonlinear biomechanical patterns. Analysis of the relationship between running posture characteristics and tibial load The SHAP results indicate that when the vertical COM position is below 927 mm, the ankle inversion angle exceeds 2.2°, and the knee internal rotation angle surpasses 2.8°, the model predicts an upward trend in tibial load. It should be noted that these findings primarily reflect statistical associations rather than direct causal relationships. Because muscle activation was not directly measured using sEMG in the present study, the muscle-related responses discussed here should be regarded as hypothesized biomechanical explanations rather than directly observed phenomena in this cohort. From a biomechanical perspective, one possible explanation is that a lower vertical COM is often associated with a deeper lower limb flexion posture, particularly during the buffering phase. This posture causes the tibia to shift forward, bringing the line of action of the vertical GRF closer to the tibial axis, thereby increasing the axial compression component [42, 43]. Furthermore, a deeper lower limb flexion may increase ankle dorsiflexion. Previous studies have suggested that greater dorsiflexion significantly enhances the external ankle dorsiflexion moment, which forces the gastrocnemius to generate stronger plantarflexion forces to maintain stability during the support phase. At this point, the higher plantarflexion force transmits increased axial load to the distal tibia via the Achilles tendon—talus pathway, leading to an increase in tibial load [42, 43]. During this process, the patellar tendon acts as a conduit for force transmission, transferring the high-intensity, sustained quadriceps tension to the tibial tuberosity, thereby significantly increasing the axial compressive load on the proximal tibia. Previous studies have shown that increased knee flexion requires continuous contraction of the quadriceps. In this context, the patellar tendon transmits the sustained high-level muscular force from the quadriceps to the tibial tuberosity, thereby enhancing the axial tensile load on the proximal tibia [44]. However, vertical COM position was analyzed as an absolute value rather than a normalized variable. Therefore, the SHAP-derived threshold should be interpreted within the context of the present sample rather than as a universally applicable cut-off. However, it is important to note that the present study employed a cross-sectional design, which prevents determination of whether a lower COM position represents a predisposing biomechanical factor contributing to increased tibial loading or a compensatory running strategy adopted following previous pain or discomfort. Therefore, the biomechanical explanations discussed above should be interpreted as potential mechanisms rather than direct causal inferences. In summary, prolonged adoption of a low COM running posture may subject the tibia to sustained high loads and potentially increasing the risk of developing MTSS [45–47]. Pending further confirmation from prospective studies, optimizing the control of vertical COM during running may represent a potential target for the prevention and management of MTSS. During the running support phase, an increase in ankle inversion angle may alter the GRF and lower limb muscle activation patterns [48, 49]. These muscle-related interpretations remain inferential, as muscle activation was not directly measured in the present study. Studies have shown that inversion shifts the center of pressure laterally, which in turn causes the line of action of the GRF to move medially [50]. This shift in the force line modifies the distribution of forces within the lower limb, increasing the medial compressive load on the tibia and subsequently raising the risk of recurrence of MTS. Furthermore, excessive ankle inversion can lead to an increased degree of knee adduction [51]. This biomechanical coupling enhances the adduction moment at the knee, which may further exacerbate the medial compressive load on the tibia [52]. The sustained increase in medial compressive load may accelerate the accumulation of bone microdamage, thereby increasing the risk of MTSS development [47]. Additionally, ankle inversion is associated with changes in lower limb muscle activation patterns, particularly increasing the demand on the plantarflexor muscles [49]. Specifically, ankle inversion can cause foot instability [53], requiring the soleus and gastrocnemius muscles to generate greater contraction forces to maintain foot stability [54]. However, existing musculoskeletal modeling studies have shown that the contraction of these two muscles during the support phase is a significant source of axial tensile and compressive forces on the tibia [55]. Repeated stretching and compression of the tibia lead to the accumulation of microdamage in the medial tibial periosteum and cortical bone, which may contribute to MTSS recurrence [56]. The results of this study also indicate a strong correlation between larger knee internal rotation angles during running and increased tibial loading. Brunner et al. [57] reported that when the knee is flexed and accompanied by internal rotation, abnormalities in lower limb alignment are likely to occur, altering the distribution of joint contact pressure. This results in excessive loading on either the medial or lateral side of the knee joint, potentially causing localized stress concentration in the tibia. Furthermore, as knee internal rotation increases, the quadriceps must enhance its activation levels during knee flexion to counteract the internal rotation moment, in order to maintain lower limb stability [58]. In this context, the tensile force generated by the quadriceps is transmitted through the patellar tendon to the tibial tuberosity, creating a pulling force along the longitudinal axis of the tibia. This increases the tensile load on the proximal tibia, thereby elevating the overall load on the tibia [59]. Additionally, excessive knee internal rotation may trigger a tensile response in the iliotibial band and lateral collateral ligament. The strain induced by these tissues further exacerbates the compressive load in the medial compartment of the knee, thereby subjecting the tibia to greater mechanical stress [60]. Thus, it is evident that excessive knee internal rotation may lead to a sustained increase in tibial loading through the cumulative effect of multiple biomechanical mechanisms, ultimately contributing to tibial overload and increasing the risk of developing MTSS. Clinical implications This study demonstrates that in novice runners with MTSS, a lower COM position, along with increased ankle inversion and knee internal rotation angles, may enhance tibial axial tensile or compressive forces. These factors significantly increase tibial loading and elevate the risk of developing MTSS. This finding has important clinical implications for gait retraining in novice runners, as appropriate optimization of running posture may help reduce excessive tibial loading in this population. For novice runners with MTSS, interventions aimed at optimizing COM position during the support phase, particularly during the buffering phase, may be beneficial for improving vertical control of the center of mass and partially optimizing mechanical load transmission through the lower limb. However, such adjustments should be implemented cautiously. In particular, attempts to increase COM position should not be interpreted as encouraging excessive lower limb extension or an overstriding pattern, especially at initial contact, because overstriding has been associated with greater braking forces, higher loading rates, and an increased risk of tibial stress-related injury. A safer and more practical strategy may be to appropriately increase step rate (cadence), which may help optimize COM excursion and running mechanics while reducing the likelihood of excessive tibial loading. Therefore, clinical recommendations should emphasize cautious and individualized gait modification, taking into account running speed, footwear cushioning, and surface conditions, to achieve safer and more effective running posture optimization. Furthermore, strengthening the evertor muscle group, such as the fibularis longus, can help maintain ankle stability in a neutral position during the support phase, thus minimizing lower limb alignment abnormalities caused by excessive inversion. Finally, emphasis should be placed on training muscles with knee external rotation functions to inhibit excessive knee internal rotation, preventing the tibia from experiencing combined rotational load while bearing vertical compressive forces, thus avoiding localized stress concentration. In addition, the specific numerical thresholds identified in our SHAP analysis—for example, vertical COM position below 927 mm, ankle inversion angle exceeding 2.2°, and knee internal rotation angle exceeding 2.8°—can serve as early warning indicators for posture correction rather than absolute pathological values. Clinicians, rehabilitation therapists, or coaches may use these indicators to monitor runners and provide timely feedback for targeted adjustments. This integration with the online personalized prediction system developed in this study allows practitioners to identify individuals at increased tibial load risk and apply preventive interventions before the onset of MTSS. Limitations Although this study employed machine learning and the SHAP framework to elucidate associations between running posture characteristics and tibial loading in novice runners, it is important to emphasize that, due to the cross-sectional design, it cannot be determined whether a lower vertical COM position, greater ankle inversion angle, and increased knee internal rotation angle represent pre-existing risk factors for elevated tibial loading or adaptive/compensatory running strategies developed following prior pain or discomfort. Therefore, the mechanistic interpretations presented in the discussion should be regarded as potential biomechanical hypotheses rather than direct causal inferences. Additionally, the relatively small sample size, primarily drawn from runners with specific regional and athletic backgrounds, may not adequately represent the broader population across different genders, age groups, and gait patterns. Consequently, this selection bias may limit the external generalizability of the model. Because vertical COM position was not normalized to participant height or leg length, the identified threshold may be partly anthropometry-dependent and may not be directly transferable across populations. Furthermore, this study did not include an asymptomatic (healthy) novice runner control group. Future studies incorporating such a control group would enable more comprehensive exploration of the pathogenic mechanisms underlying MTSS. In addition, the relatively short laboratory runway may have slightly influenced running kinematics by inducing subtle anticipatory gait adjustments near the end of the capture area. It should be noted that the discussion attributes variations in tibial load to specific muscle actions—such as increased plantarflexor and quadriceps activity associated with low COM, ankle inversion, and knee internal rotation—but musculoskeletal modeling was not performed to directly quantify individual muscle forces. Therefore, these mechanistic interpretations are indirect and hypothesis-generating, based on the estimated tibial load and its associations with running posture characteristics revealed by our machine learning analysis (XGBoost and SHAP). Despite the lack of direct muscle-level data, the discussion provides a detailed rationale for potential mechanisms, including the effects of low vertical COM on axial compression, ankle inversion on medial tibial load, and knee internal rotation on quadriceps–patellar–tibial force transmission, as supported by prior biomechanical studies. These lines of evidence lend support to the notion that specific muscle contractions may contribute to increased axial tensile and compressive forces on the tibia, thereby elevating tibial loading, while we remain cautious in interpreting these as definitive conclusions and consider them possible mechanisms. Conclusion This study, integrating machine learning algorithms with the SHAP explanation framework, systematically explores the mechanisms by which running posture characteristics influence tibial loading in novice runners with MTSS. The results indicate that a lower COM position, along with increased ankle inversion and knee internal rotation angles, nonlinearly elevates tibial loading, with specific thresholds beyond which small changes in these variables disproportionately increase tibial load. These findings highlight the complex, nonlinear interactions among multiple running posture characteristics and demonstrate the methodological advantage of SHAP in quantifying such effects, providing deeper insight into the recurrence mechanisms of MTSS. Consequently, optimizing vertical COM control, strengthening the evertor muscle group of the ankle, and enhancing the muscles involved in knee external rotation during clinical rehabilitation for novice MTSS runners will help reduce tibial loading during running, ultimately contributing to a lower risk of MTSS development.