<p>Anaerobic power is a core physical fitness indicator that substantially influences competitive performance in taekwondo; however, the Wingate 30-second test (WAnT) imposes stringent laboratory requirements and demands specialised equipment, making it impractical for widespread deployment in grassroots training settings. This study aims to investigate the feasibility of predicting the anaerobic power level (high vs. low) of taekwondo athletes from inexpensive and readily accessible anthropometric and biomechanical features, and to propose a domain-aware deep learning model tailored to the small-sample sports science scenario. Using a publicly available dataset comprising 320 elite taekwondo athletes (177 male, 143 female), 19 anthropometric and biomechanical features were collected, and binary classification labels (high/low power, 161 vs. 159) were constructed by thresholding the relative Wingate peak power (W/kg) at the gender-stratified median (female median 8.41 W/kg, male median 11.97 W/kg). We propose the anthropometric–biomechanical multi-branch attention network (ABM-Net), which explicitly models the domain knowledge structure of anthropometric–biomechanical features through intra-branch self-attention encoding across three parallel branches and bidirectional cross-branch attention mechanisms, augmented by a learnable linear residual bypass and several small-sample regularisation strategies (label-smoothed BCE, Mixup, SWA, and a 5-seed deep ensemble). Under a unified 5-fold stratified cross-validation scheme (random_state = 2024), ABM-Net was systematically compared with 11 models comprising 5 classical machine learning baselines and 5 tabular deep learning baselines, with six evaluation metrics reported: accuracy, precision, recall, F1-score, AUC-ROC, and Matthews Correlation Coefficient (MCC). SHAP and LIME were additionally applied to conduct a dual-method interpretability analysis of ABM-Net. ABM-Net obtained the best accuracy (0.8562), precision (0.8485), and MCC (0.7127) among the 11 models; however, a logistic-regression baseline reached a marginally higher F1 (0.8597 vs. 0.8589) and AUC (0.9379 vs. 0.9317), so ABM-Net’s advantage is confined to MCC and specificity (0.843) and should be read as a modest rather than across-the-board improvement. Permutation importance ranked gender (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(\Delta \textrm{AUC}=0.2799\)</EquationSource></InlineEquation>), vertical jump height (0.2207), and 30-metre sprint (0.1580) highest, whereas the additive SHAP and LIME attributions were dominated by training experience and age; rather than being mutually confirmatory, these views are complementary–under the sex-stratified labelling gender has near-zero marginal correlation with the label (<InlineEquation ID="IEq2"><EquationSource Format="TEX">\(r\approx 0\)</EquationSource></InlineEquation>) and therefore contributes chiefly as a sex-conditioning variable rather than through a direct additive effect. Through its domain-aware branch architecture and cross-branch attention mechanism, ABM-Net captures the nonlinear synergistic effects between anthropometric and biomechanical features, achieving performance competitive with strong classical baselines on small-sample sports science tabular data. The model may serve as a low-cost screening aid that complements–rather than yet replaces–laboratory-based Wingate testing, and, once validated on independent cohorts, could provide interpretable support for athlete selection screening and personalised training monitoring in taekwondo.</p>

Predicting anaerobic power status of taekwondo athletes from anthropometric and biomechanical features: a multi-branch attention deep network with SHAP and LIME interpretability

期刊论文
出版模式：

开放获取
发表日期：

2026年7月7日

查看全文

Zhigao Yan,
Huanlong Tian,
Xiaolin Zhou

摘要

Anaerobic power is a core physical fitness indicator that substantially influences competitive performance in taekwondo; however, the Wingate 30-second test (WAnT) imposes stringent laboratory requirements and demands specialised equipment, making it impractical for widespread deployment in grassroots training settings. This study aims to investigate the feasibility of predicting the anaerobic power level (high vs. low) of taekwondo athletes from inexpensive and readily accessible anthropometric and biomechanical features, and to propose a domain-aware deep learning model tailored to the small-sample sports science scenario. Using a publicly available dataset comprising 320 elite taekwondo athletes (177 male, 143 female), 19 anthropometric and biomechanical features were collected, and binary classification labels (high/low power, 161 vs. 159) were constructed by thresholding the relative Wingate peak power (W/kg) at the gender-stratified median (female median 8.41 W/kg, male median 11.97 W/kg). We propose the anthropometric–biomechanical multi-branch attention network (ABM-Net), which explicitly models the domain knowledge structure of anthropometric–biomechanical features through intra-branch self-attention encoding across three parallel branches and bidirectional cross-branch attention mechanisms, augmented by a learnable linear residual bypass and several small-sample regularisation strategies (label-smoothed BCE, Mixup, SWA, and a 5-seed deep ensemble). Under a unified 5-fold stratified cross-validation scheme (random_state = 2024), ABM-Net was systematically compared with 11 models comprising 5 classical machine learning baselines and 5 tabular deep learning baselines, with six evaluation metrics reported: accuracy, precision, recall, F1-score, AUC-ROC, and Matthews Correlation Coefficient (MCC). SHAP and LIME were additionally applied to conduct a dual-method interpretability analysis of ABM-Net. ABM-Net obtained the best accuracy (0.8562), precision (0.8485), and MCC (0.7127) among the 11 models; however, a logistic-regression baseline reached a marginally higher F1 (0.8597 vs. 0.8589) and AUC (0.9379 vs. 0.9317), so ABM-Net’s advantage is confined to MCC and specificity (0.843) and should be read as a modest rather than across-the-board improvement. Permutation importance ranked gender (\(\Delta \textrm{AUC}=0.2799\)), vertical jump height (0.2207), and 30-metre sprint (0.1580) highest, whereas the additive SHAP and LIME attributions were dominated by training experience and age; rather than being mutually confirmatory, these views are complementary–under the sex-stratified labelling gender has near-zero marginal correlation with the label (\(r\approx 0\)) and therefore contributes chiefly as a sex-conditioning variable rather than through a direct additive effect. Through its domain-aware branch architecture and cross-branch attention mechanism, ABM-Net captures the nonlinear synergistic effects between anthropometric and biomechanical features, achieving performance competitive with strong classical baselines on small-sample sports science tabular data. The model may serve as a low-cost screening aid that complements–rather than yet replaces–laboratory-based Wingate testing, and, once validated on independent cohorts, could provide interpretable support for athlete selection screening and personalised training monitoring in taekwondo.