An interpretable adaptive edge-weighted graph convolutional networks model for physical fitness assessment of primary school student in high-altitude regions

Abstract

Physical fitness testing is a vital means of assessing the physical health of adolescents. However, existing assessment methods face limitations in processing complex multivariate fitness data, demonstrating poor generalizability and a lack of interpretability. To address these challenges, we propose an innovative Adaptive Edge Weight Graph Convolutional Neural Network (AWE-GCN) model. By combining dynamic graph convolutional neural networks with the SHAP (SHapley Additive exPlanations) interpretability method, this study achieves the first high-precision assessment of comprehensive physical fitness and analysis of key influencing indicators for primary school students in high-altitude regions. The model achieved macro-average F1 scores of 97.41% (male) and 96.37% (female) for 25,790 student records from Xining, Qinghai, China (average altitude: 3,137 m). The SHAP analysis identified the 50-meter sprint, 1-minute sit-ups, and the 50-meter × 8 shuttles run as core indicators influencing primary school students’ physical fitness. It further elucidated the corresponding physiological mechanisms from the perspective of sensitive periods in physical development and high-altitude hypoxia adaptation. Furthermore, comparative experiments against four baseline models, including SVM, CatBoost, MLP and CNN, clearly demonstrate that the AWE-GCN model achieves the best performance across all key evaluation metrics, such as Precision, Recall, and F1-Score. The ablation study confirms that SMOTE provides a balanced data foundation, upon which the adaptive edge weighting module maximizes physical fitness classification performance, representing the model’s key capability to dynamically capture the complex relationships among physiological indicators. By integrating deep learning with interpretability analysis, this study provides a high-precision framework for adolescent physical fitness assessment, with scientific significance for optimizing physical education curriculum design and formulating health policy.

Introduction

Physical fitness test scores can reflect multiple functional indicators of an individual’s body, such as cardiorespiratory function, muscular strength, and flexibility, and play a crucial role in health education and personal development. Regular physical fitness testings not only assess students’ physical health status but also enable the formulation of professional fitness plans based on the results, thereby promoting healthy development^1,2. Furthermore, research suggests that good physical fitness and health conditions correlate with self-confidence and social skills³. Additional research points to a positive correlation between physical health and academic performance⁴and students with better physical fitness often demonstrate superior academic achievement^5–7. Therefore, implementing effective physical training and testing is important for the physical and mental development of students.

Therefore, concurrent with physical fitness testing conduction the assessment and grading of physical fitness testing scores are particularly important in physical education. This not only helps guide individual exercise and health management^8,9but also provides schools with data support to optimize physical education curricula and activity settings¹⁰allowing the planning of physical activities specifically to promote improvements in physical health and psychosocial factors^11,12. Moreover, large-scale physical fitness testing score assessments and grading data can provide a basis for formulating public health and sports policies¹³. Traditional physical fitness testing assessment methods, such as cluster analysis, descriptive statistics, correlation analysis, and analysis of variance, have been widely applied in existing research^14,15. Cho et al. used the Pearson’s correlation coefficient to conduct a validity analysis of firefighter physical fitness tests, aiming to improve and develop testing methods more suited to the occupational demands of firefighters¹⁶. Lockie et al. employed independent sample t-tests, partial correlation analysis, and stepwise regression analysis to compare performance differences between male and female firefighters in Physical fitness testing, analyze the relationship between agility tests and other physical fitness testing, and identify which physical fitness testing can predict agility performance¹⁷. Keston et al. used principal component analysis to determine the relationship between physical fitness components and performance in law enforcement-specific physical ability tests (PATs), and proposed health recommendations to help meet occupational performance demands¹⁸. However, these traditional methods exhibit several drawbacks. First, they often rely on linear models, making it difficult to capture complex non-linear relationships. Second, traditional methods often fail to effectively handle high-dimensional data and missing values, which reduces the reliability of the analysis results. Third, these methods lack flexibility and specificity in designing personalized plans. Therefore, more advanced analytical techniques must be explored to enhance the effectiveness and application value of physical fitness test data analysis.

Recently, artificial intelligence (AI) has made significant progress in the field of physical fitness assessment, with an increasing number of related studies demonstrating its potential in predicting physical fitness¹⁹. Researchers have employed machine learning (ML) models such as the Multilayer Perceptron (MLP), K-Nearest Neighbor (KNN), Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM) to effectively classify student physical fitness testing scores²⁰. Gao et al. utilized Self-Organizing Map (SOM) networks to analyze students’ physical fitness test scores, emphasizing the significant impact of body weight on health²¹. Sulla-Torres et al. conducted an in-depth classification of the physical fitness status of 1,813 children using neural networks and fuzzy logic²². Sun et al. developed a physical fitness classification model using the Naive Bayes algorithm, categorizing students’ physical conditions into four levels and achieving a precision of 81.02% and recommended improving school physical education curricula for the timely detection of health issues²³. In another study, Li et al. proposed a multi-model decision fusion method based on multi-source data for the physical fitness assessment classification of cancer patients, achieving a high precision of 95.1%²⁴.

The use of AI models in the field of physical health is also gradually increasing^25,26. Cai et al. assessed the health levels of 890 adults through questionnaires and employed machine learning techniques, such as logistic regression and deep learning, to build classification models, achieving a precision exceeding 85%²⁷. Wang et al. utilized natural language processing (NLP) techniques to deconstruct the physical fitness and health assessment system for adolescents with physical disabilities, further developing rehabilitation strategies²⁸. Furthermore, AI models have unique advantages in predicting athletic performance and status. For instance, Mandorino et al. developed a fitness index using ML models to predict the physical fitness status of soccer players, enabling personalized training adjustments and injury prevention, to not only enhance the efficiency of data analysis but also provide a scientific basis for improving athletes’ physical fitness and health management²⁹. However, numerous challenges remain in the field of physical fitness analysis. First, there is often insufficient quantity and quality of data, and the data samples typically exhibit an imbalance, which can affect the training effectiveness and predictive accuracy of the models. Second, and more critically, standard machine learning models treat physical fitness indicators as independent features or only capture simple correlations between them. This approach overlooks that the human body is an integrated system in which various physiological indicators form a complex, interconnected physiological network, representing a classic example of non-Euclidean data. Therefore, severely underestimating the intrinsic complexity of student physical fitness. Third, the interpretability of traditional machine learning models is relatively low, leaving researchers without the necessary support when referencing model recommendations. Furthermore, there is a lack of integration of cross-disciplinary knowledge, especially when interpreting the AI model prediction results from a sports and health science perspective. The model’s analytical results lack interpretation from an exercise and health science perspective, hindering their translation into effective, real-world intervention strategies.

To address the aforementioned problems and challenges, this study proposes a novel and interpretable adaptive edge weight graph convolutional neural network (AWE-GCN) model to predict the comprehensive physical fitness of primary school students in high-altitude regions. At the data level, the synthetic minority over-sampling technique (SMOTE) is incorporated in the preprocessing stage to solve the samples imbalance. In order to effectively deal with the complex interactions between physiological indicators, this study transforms data into a graph structure to more accurately capture the complex interactions among physical fitness features. Furthermore, it advances beyond standard GCNs that use a fixed and unweighted graph by learning an adaptive, task-specific-weight for each connection, thereby improving the precision of the overall physical fitness assessment. To address the lack of interpretability of machine learning models, SHAP analysis is integrated to clearly quantify the contribution of various physical test indicators to the final physical fitness rating from both global and local dimensions, making the model’s decision-making process completely transparent. Finally, it explains the reasons behind the model’s predictions and SHAP analysis results from the perspective of sensitive periods in children’s physical development and high-altitude hypoxia so as to bridge the challenge of insufficient integration of interdisciplinary knowledge and make the research results more scientific and reasonable.

Materials and methods

Graph Convolutional Networks (GCNs) learn representations of graph-structured data by aggregating feature information from nodes and their neighbors (message passing), leveraging the graph’s topology to enhance feature expression³⁰. The various physiological and motor indicators in adolescent physical fitness data exhibit inherent correlations. GCNs can transform tabular data into graph structures to more accurately capture complex interactions among fitness features, thereby improving prediction performance. However, traditional GCN models rely on a fixed adjacency matrix and cannot dynamically adjust edge weights, making them sensitive to noisy edges and task-irrelevant connections, and thus difficult to capture complex non-linear relationships between features. Therefore, this study proposes an AWE-GCN, which dynamically adjusts the importance of edges through learnable parameters, enabling it to automatically strengthen relationships between key nodes, suppress irrelevant or redundant connections, and enhance the precision of feature aggregation.

Database

The AWE-GCN model proposed in this study utilizes physical health tracking data from a subset of 11-12-year-old primary school students in Xining City, Qinghai Province, China (average altitude 2261 m), collected between 2019 and 2023, comprising 13,351 males and 12,439 females. The data were obtained with authorization from the Xining Municipal Education Bureau, from 32 public primary schools within the jurisdiction. Data comprised standardized records of students’ annual physical fitness tests, including physiological and motor indicators such as BMI, cardiorespiratory function, and athletic ability. Raw data were measured using calibrated instruments, uniformly checked, and used to assess the students’ overall physical fitness. After anonymization, the dataset comprised 25,790 valid records structured as original samples Y=(X₁, X₂, X₃, X₄, X₅, X_6, X₇). The features and classification labels of the samples are shown in Fig. 1, where features X₁–X₇ represent BMI, vital capacity, 50-meter sprint, sit-and-reach, 50-meter×8 shuttle run, 1-minute sit-up, and 1-minute skip rope, respectively. The classification label Y has four categories: Class 0, 1, 2, and 3 represent ‘Unfit’, ‘Qualified’, ‘Good’, and ‘Excellent,’ respectively. All experiments and data collection involving human subjects in this study were conducted in accordance with relevant international and national guidelines and regulations, as well as the standards of the Declaration of Helsinki. The study protocol received approval from the Science and Technology Ethics Committee of Qinghai Normal University (Approval No. 2025-05). Furthermore, informed consent was obtained from all study participants and their legal guardians prior to their inclusion in the study.

Fig. 1.

Model sample features and classification labels.

Graph data construction

After normalizing the raw sample data, the synthetic minority over-sampling technique SMOTE is employed for data augmentation to address the class imbalance issue within the samples, thereby enhancing the model’s generalization ability for minority classes³¹.To construct the graph structure, the original feature matrix X∈R^N×D is first transposed to yield X^T∈R^D×N, where N and D denote the number of samples and features dimensions, respectively. Subsequently, the Pearson correlation coefficient is computed across the features, resulting in a D×D correlation matrix, C. An undirected and unweighted adjacency matrix, A, is then derived by applying a predefined threshold of τ = 0.3 to the absolute values of the correlation matrix. The element A is defined as, if |C_ij|>τ, then A_ij=1, otherwise A_ij=0. Finally, this dense adjacency matrix is converted into a sparse edge index format to obtain the graph representation G=(V, E). In this graph, the node set V={v₁,.,v_D} represents the D features of the dataset, while the edge set E⊆V×V is composed of all feature pairs whose correlation strength exceeds the threshold τ = 0.3.

Proposed AEW-GCN model

In the task of elementary school physical fitness assessment, the correlations among various test items are complex and dynamic. The fixed graph structure defined by a standard adjacency matrix in traditional GCN is incapable of capturing these subtle variations. To address this, the AWE-GCN model introduces learnable edge weight. This enables the model to dynamically adjust the weights of edges, thereby automatically learning and quantifying the strength of the correlations between different physical test items. The specific implementation of this mechanism is as follows. An adaptive-learnable parameter vector E_A associated with all edges in the graph was initialized. During the model’s forward pass, this vector is mapped into a specific weight w = σ(e_ij) for each edge using a sigmoid function σ, where e_ij∈E_A is the raw learnable parameter corresponding to the edge (j, i). This process forms an adaptive weight matrix, W_adp, which maintains the same sparsity structure as the original adjacency matrix. In the course of graph convolution, these weights are automatically optimized via gradient descent. As information propagates along an edge, it is multiplied by the corresponding learned weight. This allows the model to learn a unique weighting coefficient for each connected pair of nodes in the graph, which is continuously optimized through gradient descent during training.

The specific architecture and workflow of the proposed AWE-GCN model are illustrated in Fig. 2. After graph data construction, the model receives the node feature matrix X* as an input. This is passed to the first graph convolutional layer, outputting the following features:

1

Fig. 2.

Overall process framework of the AWE-GCN model.

Where is the learnable weight matrix of the first graph convolutional layer, is the adjacency matrix with self-loops added andis the transition matrix of the normalized adjacency matrix. Passing H₁ as the input to the second-graph convolutional layer yields the following output:

2

Where is the learnable weight matrix of the second-graph convolutional layer. Passing H₂ as the input to the third-graph convolutional layer yields the final node embeddings H₃:

3

Where is the learnable weight matrix of the third-graph convolutional layer. After passing through three graph convolutional layers, the H₃ features are fed into three fully connected layers FC₁, FC₂ and Y. The formulas for the fully connected layers are:

4

5

6

Where W_fcn are the learnable weights of the fully connected layers and b_fcn are the bias terms.

The model was trained using the Adam optimizer and the cross-entropy loss function. To address the class imbalance issue and improve the performance of the model across different classes, a learnable class weight vector w_c is introduced into the loss function, which is defined as follows:

7

where y_{i, c} is the real classification label of the sample and p_{i, c} is the model’s predicted output.

To gain deeper insights into the decision-making process of our trained AWE-GCN model, we employed the SHAP framework for interpretability analysis. The details are elaborated in the section entitled "Model Interpretability Analysis". Algorithm 1 and Fig. 3 shows the training process of the proposed method.

Algorithm: Adaptive Weighted Edge Graph Convolutional Network

Fig. 3.

The proposed algorithm flow chart.

Algorithm1

AWE-GCN for Physical Fitness Assessment.

Results and experiment

Model performance

Tables 1 and 2 display the precision, recall, and F1 scores of the model for classifying the physical fitness test set samples for males and females, respectively, where classes 0–3 correspond to the physical fitness testing results ‘Unfit’, ‘Qualified’, ‘Good ’, and’ Excellent’. The macro-average F1 score for males across all fitness classes reached 97.41%, whereas for females, it was 96.37%, indicating that the model possessed excellent overall classification stability, with the overall performance for male classification being slightly superior to that of females. Specifically, for both male and female samples, the model’s recognition ability for Class 1 and Class 3 is exceptionally high, with their F1 scores both exceeding 98.5%, whereas the classification performance for Class 1 and Class 2 is relatively weaker, especially for the female samples, where the recall for Class 1 (91.77%), the F1 score for Class 1 (93.61%), and the precision for Class 2 (93.11%) are significantly lower than those of other classes. However, their F1 scores still exceeded 93%, maintaining a high overall performance level.

Table 1.

Performance metrics for male Samples.

Metrics	Class 0	Class 1	Class 2	Class 3	Macro-average
Precision	98.48%	97.33%	95.77%	98.12%	97.43%
Recall	98.86%	94.81%	96.48%	99.52%	97.42%
F1-score	98.67%	96.05%	96.12%	98.81%	97.41%

Table 2.

Performance metrics for female Samples.

Metrics	Class 0	Class 1	Class 2	Class 3	Macro-average
Precision	97.48%	95.54%	93.11%	99.15%	96.32%
Recall	99.57%	91.77%	95.17%	99.29%	95.45%
F1-score	98.51%	93.61%	94.13%	99.22%	96.37%

Figure 4 shows the receiver operating characteristic (ROC) curves of the model and corresponding area under the curve (AUC) values. As shown in the figure, for both males and females, the ROC curves for all classes are significantly higher than baseline. Among these, the AUC for Classes 0 and 3 is the highest, being 1 for both males and females, indicating that the model has almost no misclassifications for Classes 0 and 3. The model’s classification performance for Class 2 is slightly weaker, with an AUC of 0.95 for males and 0.91 for females in this class, indicating a small number of false positives.

Fig. 4.

Roc curves of the model.

Confusion matrices were introduced to further analyze the classification performance of the model for each class. As shown in Fig. 5, it is clear from the confusion matrices that different class quantities correspond to different color intensities. The true and predicted labels are marked on the left side and bottom of the confusion matrix, respectively, facilitating observation of the prediction results for each physical fitness test class in the figure. For the male samples, as shown in Fig. 5(a), the model performed best in classifying Class 3, correctly classifying 838 samples with only 4 samples misclassified as Class 2; it also exhibited excellent performance for Class 0, with 779 samples correctly classified and nine samples misclassified as Class 1; the performance for Class 1 was slightly weaker, with 12 samples misclassified as Class 0 and 30 samples misclassified as Class 2. Contrastingly, in Fig. 5(b), the model’s classification performance for female samples is similar to that for male samples; the ratios of correctly classified samples to incorrectly classified samples for Classes 0 and 3 are 696:3 and 702:5, respectively; the number of misclassified samples in Classes 1 and 2 increases, with 682 samples correctly classified as Class 1, 18 samples misclassified as Class 0, and 46 samples misclassified as Class 2. In Class 2, 691 samples were correctly classified, 29 samples misclassified as Class 1, and six samples misclassified as Class 3. Overall, the model demonstrated good classification performance on both datasets, with a particularly stable performance in classes 0 and 3.

Fig. 5.

Confusion matrix of model classification.

Comparative experiments

Experimental setup

To comprehensively evaluate the classification performance of the proposed AWE-GCN model, five representative models were selected for comparison: Support Vector Machine (SVM), the gradient boosting framework categorical boosting (CatBoost), Multi-Layer Perceptron (MLP), and Convolutional Neural Network (CNN). Using a stratified sampling strategy, both the male and female student datasets were divided into a training set (80%), a validation set (10%), and a test set (10%) to ensure a consistent class distribution across all subsets. Finally, all features were standardized to eliminate differences in scale. The hyperparameters for each model were meticulously tuned, as detailed in Table 3.

Table 3.

Hyperparameter settings for the 5 models.

Models	Hyperparameter
SVM	Penalty coefficient:0.5, kernel: rbf, random seed: 42
CatBoost	Iterations:50, learning_rate:0.5, depth:2, random seed: 42
MLP	Optimizer: Adam, learning rate: 0.0005, weight decay: 0.0001, training epochs: 500
CNN	Optimizer: Adam, learning rate: 0.001, weight decay: 0.0001, training epochs: 300
AWE-GCN	Optimizer: Adam, learning rate: 0.0003, weight decay: 0.0001, training epochs: 500

Experimental results

The performance of the proposed AWE-GCN model and the comparative models on the male and female student datasets is illustrated in Fig. 6. The figure shows the weighted precision, recall, F1 score and macro-average F1 score of each model. On both the male and female datasets, all performance metrics of the AWE-GCN model exceeded 96%, substantially outperforming the other comparative models. Moreover, in contrast to the significant performance disparities exhibited by MLP and CatBoost in different gender datasets, AWE-GCN demonstrated exceptional stability and generalization capability on both male and female samples. This validates the effectiveness of the AWE-GCN model for this physical fitness assessment task.

Fig. 6.

Comparative experiment performance comparison.

Ablation experiment

Experimental setup

To rigorously evaluate the contribution of each core component in proposed AWE-GCN model, an ablation study is conducted on a dataset of male and female students using the following four experimental configurations: (1) Baseline, the standard graph convolutional network. (2) Baseline with SMOTE: The baseline model trained on the balanced dataset processed by SMOTE. (3) Baseline with AWE: The baseline model integrated with the adaptive edge weighting module, trained on the original imbalanced data. (4) AWE-GCN: the full model integrates both SMOTE and the adaptive edge weighting module.

All experiments were implemented within a unified framework to ensure a fair comparison. Both the male and female student datasets were split into training, validation, and test sets at a ratio of 80%/10%/10%, using a fixed random seed of 42. All shared hyperparameters were kept consistent across the experiments: optimizer (Adam), learning rate (0.0003), weight decay (0.0001), hidden dimension (128), training epochs (500), and dropout rate (0.1).

Experimental results

The weighted precision, recall, F1 score and macro average F1 score of test set are used as core evaluation indicators. As Fig. 7 shows, the experimental results demonstrated a high consistency across both the male and female student datasets. After directly introducing the adaptive edge weight module to the baseline model on the original imbalanced data, the model’s performance showed no significant improvement compared to the baseline. The weighted indicators of the baseline model without SMOTE processing and the baseline model with adaptive edge weight are both maintained at around 89%, but the macro-average F1 scores (79.06% for males and 75.38% for females) are significantly lower. This indicates that when the data distribution is severely skewed, the adaptive edge weighting mechanism is hard to learn effective patterns, and its potential is masked by the inherent data deficiency. This is because the class imbalance causes the model’s overall performance to be dominated by the majority class, while its ability to identify minority class samples remains severely inadequate. In contrast, SMOTE effectively addressed this issue, enabling the model to recognize and process all classes in a balanced manner. This led to a substantial increase in the macro F1 score, bringing it to a level nearly to the weighted F1 Score. The proposed AWE-GCN achieved the best performance in all experimental configurations, reaching weighted F1 scores of 96.06% and 96.37%, and macro F1 scores of 96.06% and 96.36% on the male and female datasets, respectively. Compared to the baseline with SMOTE, AWE-GCN achieved a stable net gain of over 3.4% across all metrics. This demonstrates the superior learning capability of the adaptive edge weighting module when operating on a high-quality, balanced data foundation. Moreover, the macro average F1 and weighted F1 scores of AWE-GCN are almost exactly the same, indicating that AWE-GCN has achieved highly consistent and excellent performance on both majority and minority categories. The multi-dimensional metric comparison clearly indicates that the superior performance of AWE-GCN does not stem from a simple superposition of individual components. Instead, it is driven by the synergistic effect between the data-level balancing strategy and the adaptive edge weighting module.

Fig. 7.

Ablation experiment performance comparison.

Model interpretability analysis

SHAP

SHAP, which is based on Shapley values from game theory, explains classification models by distributing the contribution of each feature to the model’s prediction³². The process for multi-class SHAP first involves independently calculating a baseline prediction value ϕ₀, Next, it solves for the contribution value ϕ_j of each feature j, such that the model’s output for class k can be decomposed as:

6

In the formula, M represents the total number of features, ϕ_j is the Shapley value, indicating the contribution of feature j to the prediction for class k, and its mathematical expression is:

7

Where X is the set of all features, S is a subset of features that does not contain feature j, and f(S) is the prediction output for feature subset S. This formula obtains the Shapley value for a specific class prediction by evaluating the marginal contribution of feature j across all possible feature combinations and then calculating a weighted average. Features with larger absolute SHAP values had a significant impact on the model’s classification of primary school students’ physical fitness, highlighting their importance in the decision-making processes. The SHAP method comprehensively analyzes the output results of the model through global and local analyses. At the global level, this method reveals the overall contribution of each feature to the model’s output, its dependency relationships with the model, and feature interaction effects. At the local level, SHAP can provide an in-depth interpretation of the model’s decision-making mechanism for individual samples and quantitatively assess the feature contributions in specific instances. Therefore, it can ensure that the model’s explanation satisfies the local accuracy (the prediction value equals the sum of the baseline value and SHAP values of each feature), missingness (the contribution of a missing feature is zero), and consistency (the ranking of feature importance is consistent with the model behavior).

SHAP analysis

To interpret the role of each feature in the model macroscopically, feature importance plots were generated based on the SHAP values from the AWE-GCN model. As shown in Fig. 8, the length of the horizontal bars represents the mean absolute SHAP value for each feature arranged in ascending order along the y-axis, and the bar length signifies their degree of influence on the classification of primary school students’ physical fitness. Among both males and females, X₃ (50-meter sprint) had mean values of 0.117 (male) and 0.104 (female), which were significantly higher than the other features, indicating that it had the greatest impact on students’ physical fitness, whereas X₁ had the lowest mean values of 0.003 (male) and 0.002 (female), exerting a relatively smaller influence. For males, X₆ and X₅ exhibited nearly identical feature importance, ranking second and third, respectively. For females, X₅ had a mean value of 0.048, ranking second in feature importance, whereas X₆ ranked third, with a mean value of 0.036. Since the importance of these three features surpasses that of others, and considering their consistency across different gender groups, this suggests that they possess strong universality and representativeness, which helps in accurately grasping the key factors affecting physical health. Therefore, an in-depth investigation was conducted on the top three features in terms of importance: X₃, X₆, and X₅.

Fig. 8.

Model SHAP feature importance. X₁-BMI; X₂- Vital Capacity; X₃- 50-meter Sprint; X₄- Sit and Reach; X₅- 50-meter × 8 Shuttle Run; X₆- 1-minute Sit-up; X₇- 1-minute Rope Skip.

To display the impact of each feature on the physical fitness of primary school students more intuitively, bee swarm plots were generated, as illustrated in Fig. 9, revealing the influence range of features on the prediction of the four fitness classes (Class 0-Unfit, Class 1-Qualified, Class 2-Good, and Class 3-Excellent). Each point in the plot represents the SHAP value for a specific feature, with a positive value indicating that the feature increases the prediction outcome, whereas a negative value indicating a decrease. The color bar reflects the trend of the feature values from low (blue) to high (red). As presented in Fig. 9(a), for males, X₃ was the most influential feature across all four classes and was positioned at the top of the y-axis. Notably, for the first three classes (Classes 0, 1, and 2), the red points corresponding to X₃ are mainly distributed in the positive SHAP value interval, indicating that higher X₃ values tend to increase the prediction value. Conversely, for Class 3, the red points for X₃ are mostly located in the negative interval, suggesting that high X₃ values decrease the prediction value. X₅ exhibited a similar pattern, but its importance ranked lower than X₃. High values of X₆ had a negative impact on the predictions for Classes 1 and 2, but a positive impact on the predictions for Class 3. As shown in Fig. 9(b), for females, the feature importance rankings for Class 0 and Class 1 were X₃, X₇, X₁, and X₃, X₅, X₆, respectively. The behavior of X₃ across all fitness classes was similar to that observed in the males. High values of X₆ have a negative impact on the predictions for Class 2, but a positive impact on the predictions for Classes 1 and 3; high values of X₅, however, exert a positive effect on predictions for Classes 0, 1, and 2, and a negative effect on predictions for Class 3.

Fig. 9.

SHAP bee swarm plots for each class of males and females.

Discussion

The AWE-GCN model proposed in this study investigated the physical fitness test data of 11- to 12-year-old primary school students in Xining City, Qinghai Province, China (average altitude 3137 m) and found that the 50-meter sprint, 1-minute sit-ups, and 50-meter × 8 shuttle run are key factors influencing the physical health of children in high-altitude regions. This finding is closely related to the synergistic effects of sensitive periods of physical development and physiological adaptation mechanisms to high-altitude hypoxia.

Children’s physical development exhibits the characteristics of age-specific sensitive periods; reliance on anaerobic versus aerobic capacities differs significantly across stages, thereby influencing the ranking of feature importance. The prepubertal growth spurt typically occurs between 11 and 14 years of age. During this period, bones and muscles develop rapidly, increasing the demand for energy metabolism dominated by short bursts and making anaerobic capacity particularly important. The 50-meter sprint, as a short-duration, high-intensity exercise, relies on the phosphagen system for energy supply. Its energy output is highly efficient but short-lived (approximately 10 s), closely aligned with the demands for neuromuscular coordination and rapid strength development in children at this stage. Conversely, the 1-minute sit-up is an exercise primarily fueled by the glycolytic system; this item reflects the anaerobic endurance and strength of the abdominal muscles, while also testing the body’s tolerance to lactic acid. Both are important indicators for assessing the athletic ability of prepubertal children and possess high feature importance. After entering puberty (approximately 14 years and older), as various organs gradually mature and the cardiorespiratory function strengthens, the importance of aerobic capacity progressively surpasses that of anaerobic capacity. The 50-meter × 8 shuttle run, a key indicator of mixed anaerobic and aerobic capacity, relies on long-term, regular exercise stimuli for improvement, which requires a longer growth and development period than simply developing anaerobic capacity; this is reflected in the differing importance of this indicator between males and females.

The location where the samples were collected, Xining, is a typical high-altitude, low-oxygen environment, and so significantly affects the athletic performance of adolescents. The 50-meter sprint is a short-duration, high-intensity exercise primarily fueled by the phosphagen system, which reflects students’ anaerobic metabolic capacity. In high-altitude regions, children’s anaerobic metabolic capacity shows adaptive enhancement before puberty, glycolytic system activity is upregulated in hypoxic environments, and short-distance sprinting ability is positively correlated with anaerobic metabolism. Furthermore, the activity of lactate dehydrogenase (LDH) in the skeletal muscles of high-altitude children is approximately 18%−22% higher than that of children in plain regions, indicating that enzymes related to anaerobic metabolism are enhanced during metabolic adaptation. By optimizing anaerobic metabolism, the body can enhance exercise endurance and survival during the initial period at high altitudes, laying the foundation for subsequent adaptation processes, which is acritical initial step in high-altitude adaptation. The 1-minute sit-up is an exercise primarily powered by the glycolytic system; this item reflects the anaerobic endurance and strength of the abdominal muscles, while also testing the body’s tolerance to lactic acid. In a high-altitude hypoxic environment, the synergistic contraction efficiency of the diaphragm and abdominal muscles directly affects breathing depth and oxygen uptake efficiency. Children with insufficient core muscle strength experience exacerbated cellular hypoxia due to limited respiratory compensation, leading to a reduced rate of post-exercise lactate clearance. Therefore, core muscle function is crucial for high-altitude adaptation. Moreover, abdominal endurance in high-altitude children is significantly correlated with hypoxia tolerance, and an increase in diaphragmatic contraction efficiency can reduce resting oxygen consumption by 5–8%. Consequently, enhancing core muscle strength can improve pulmonary ventilation efficiency, thereby improving oxygen uptake, which can partially alleviate the damage caused by hypoxia and maintain normal physiological function. The 50-meter × 8 shuttle run is a composite endurance test item, with its metabolic pattern falling between anaerobic burst and aerobic endurance. Although a single short-distance sprint can maintain high-intensity output through adaptive enhancement of the glycolytic system, the multiple-turn characteristic of the shuttle run places higher demands on the body’s aerobic metabolic capacity. In a high-altitude environment, even if respiratory compensation mechanisms can increase oxygen uptake, the rightward shift in the hemoglobin-oxygen dissociation curve reduces tissue oxygen utilization, thereby limiting aerobic endurance performance. Therefore, children in high-altitude regions require a long period of adaptation to achieve relatively good physiological balance and functional maintenance in a special high-altitude environment.

In summary, the interaction between the sensitive periods of children’s pubertal growth and development and hypoxic stress influences metabolic priorities and causes the aforementioned three motor ability indicators to have a significantly higher impact than other indicators on the physical fitness training of primary school students.

However, this study also has some limitations. Although the dataset in this study covers a wide range of altitudes, it is sourced solely from the administrative area of Xining City. Future research will expand the data collection area for primary school students’ physical fitness tests to further validate the model’s generalizability. Furthermore, to enhance the assessment performance, more advanced graph-based techniques, such as Multiview Graph Convolution encoders³³ and Link-Based Attributed Graph Clustering³⁴will be incorporated into our research plan.

Conclusion

This study proposed an interpretable deep learning model based on AWE-GCN to assess the physical fitness of primary school students. The model was applied to the physical fitness testing assessment of 11–12 years old primary school students in a high-altitude region (average altitude: 3137 m). Concurrently, the SHAP method was used to conduct an interpretability analysis of the prediction results of the model. This study systematically elucidated the physiological mechanisms underlying key factors influencing children’s physical health in high-altitude environments. The main conclusions are as follows:

(1) The AWE-GCN model, through its adaptive edge-weight mechanism, effectively captures the nonlinear associations between physical fitness testing indicators, significantly enhancing the physical fitness classification performance. The model demonstrated a high prediction accuracy in physical fitness assessment tasks for both males and females. For males, the predicted macro-average precision, recall, and F1 score reached 97.43%, 97.42%, and 97.41%, respectively; for females, the predicted macro-average precision, recall, and F1 score reached 96.32%, 95.45%, and 96.37%, respectively.

(2) Comparative experiments against four baseline models, including SVM, CatBoost, MLP and CNN, clearly demonstrate that the AWE-GCN model achieves the best performance across all key evaluation metrics, such as Precision, Recall, and F1-Score. The ablation study confirms that SMOTE provides a balanced data foundation, upon which the adaptive edge weighting module maximizes physical fitness classification performance, representing the model’s key capability to dynamically capture the complex relationships among physiological indicators.

(3) SHAP analysis indicated that the 50-meter sprint (anaerobic capacity), 1-minute sit-ups (core muscle endurance), and the 50-meter × 8 shuttle run (mixed capacity) are core indicators influencing the physical health of children in high-altitude regions. Among these, the feature importance of the 50-meter sprint was significantly higher than that of the other indicators, reflecting the high dependence of prepubertal children on their anaerobic metabolic capacity.

(4) The feature importance of the aforementioned indicators in the model was shaped by the interplay between the differential reliance on anaerobic and aerobic capacities during children’s sensitive periods of growth and development and the metabolic adaptations of children in high-altitude, low-oxygen environments.

Author contributions

BW contributed to methodology, data curation, and writing - review & editing. YX contributed to data curation. XY handled project administration. XZ was responsible for conceptualization and methodology.

Data availability

The datasets used and analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval

All experiments and data collection involving human subjects in this study were conducted in accordance with relevant international and national guidelines and regulations, as well as the standards of the Declaration of Helsinki. The study protocol received approval from the Science and Technology Ethics Committee of Qinghai Normal University (Approval No. 2025-05). Furthermore, informed consent was obtained from all study participants and their legal guardians prior to their inclusion in the study.