Scholarlink

<p>Muscle injuries account for approximately 31% of all time-loss injuries in professional football, yet existing prediction models are constrained by small sample sizes, dependence on proprietary physiological data, limited interpretability, and inadequate temporal validation designs. This study develops an interpretable, strictly prospective prediction framework trained on 438,219 player-season records sourced from Transfermarkt, spanning 15 seasons (2010/11–2024/25) across 32 leagues. To ensure temporal integrity, all predictive features are constructed exclusively from prior-season data, and models are evaluated under a chronological hold-out design in which earlier seasons serve as training data and the two most recent seasons constitute the test set. We propose a multi-method stability-weighted feature selection (MSFS) strategy that cascades filter-based redundancy removal, four-method ensemble voting, and cross-validation stability verification, reducing 57 candidate features to a robust subset. Under cost-sensitive learning designed for severe class imbalance (6.2% positive rate), TabNet attains the highest discrimination among five candidate classifiers with an area under the receiver operating characteristic curve (AUC) of 0.812 (95% CI: 0.804–0.820), outperforming XGBoost (0.806), LightGBM (0.803), long short-term memory network (LSTM) (0.789), and logistic regression (0.761). Ablation experiments confirm that stability-verified feature selection and metaheuristic hyperparameter optimisation yield complementary performance gains, and direct comparison with Bayesian optimisation demonstrates that the detective behaviour algorithm provides competitive tuning across architectures. A seven-layer Shapley additive explanations (SHAP) interpretability analysis reveals that under the prospective design, injury history emerges as the most important feature category (33%), surpassing market-and-career features (29%), confirming that the temporal correction elevates genuine recurrence risk over reporting-related proxies. A two-stage risk architecture is identified in which market-and-career features establish a baseline risk stratum, whereas historical injury burden acts as a multiplicative amplifier among predisposed individuals. Individual-level decompositions further indicate that substitution-based rotation patterns are associated with attenuated age-related risk elevation. Sensitivity analyses excluding market value features quantify the extent to which predictive performance reflects genuine injury risk versus differential reporting completeness across leagues. These findings demonstrate that publicly accessible transfer-market data, combined with attention-based deep tabular learning and rigorous temporal validation, can support meaningful and interpretable injury risk stratification without proprietary monitoring systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A TabNet-SHAP framework with stability-weighted multi-method feature selection for muscle injury prediction in professional football

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月2日

查看全文
  • Wenbo Rao

摘要

Muscle injuries account for approximately 31% of all time-loss injuries in professional football, yet existing prediction models are constrained by small sample sizes, dependence on proprietary physiological data, limited interpretability, and inadequate temporal validation designs. This study develops an interpretable, strictly prospective prediction framework trained on 438,219 player-season records sourced from Transfermarkt, spanning 15 seasons (2010/11–2024/25) across 32 leagues. To ensure temporal integrity, all predictive features are constructed exclusively from prior-season data, and models are evaluated under a chronological hold-out design in which earlier seasons serve as training data and the two most recent seasons constitute the test set. We propose a multi-method stability-weighted feature selection (MSFS) strategy that cascades filter-based redundancy removal, four-method ensemble voting, and cross-validation stability verification, reducing 57 candidate features to a robust subset. Under cost-sensitive learning designed for severe class imbalance (6.2% positive rate), TabNet attains the highest discrimination among five candidate classifiers with an area under the receiver operating characteristic curve (AUC) of 0.812 (95% CI: 0.804–0.820), outperforming XGBoost (0.806), LightGBM (0.803), long short-term memory network (LSTM) (0.789), and logistic regression (0.761). Ablation experiments confirm that stability-verified feature selection and metaheuristic hyperparameter optimisation yield complementary performance gains, and direct comparison with Bayesian optimisation demonstrates that the detective behaviour algorithm provides competitive tuning across architectures. A seven-layer Shapley additive explanations (SHAP) interpretability analysis reveals that under the prospective design, injury history emerges as the most important feature category (33%), surpassing market-and-career features (29%), confirming that the temporal correction elevates genuine recurrence risk over reporting-related proxies. A two-stage risk architecture is identified in which market-and-career features establish a baseline risk stratum, whereas historical injury burden acts as a multiplicative amplifier among predisposed individuals. Individual-level decompositions further indicate that substitution-based rotation patterns are associated with attenuated age-related risk elevation. Sensitivity analyses excluding market value features quantify the extent to which predictive performance reflects genuine injury risk versus differential reporting completeness across leagues. These findings demonstrate that publicly accessible transfer-market data, combined with attention-based deep tabular learning and rigorous temporal validation, can support meaningful and interpretable injury risk stratification without proprietary monitoring systems.