Interpretable student physical performance classification in physical education using a hybrid attentive tabular network (HAT-Net)
Automated responses of the student body performance in Physical Education (PE) settings as the objective forecasting of individual levels of physiological performance and fitness through the standardised cross-sectional measures has been a critical issue because of the subjectivity, low scalability, and poor interpretability of the traditional assessment procedures. A consistent, interpretable assignment of individual physiological and fitness profiles to useful performance levels is a priori to evidence-based design of PE programs and specific instructional intervention. Although the gradient boosting decision tree models have proved high baseline on organized tabular data, they are not equipped with adaptive feature weighting and contextual inter-feature dependency modeling needed in transparent educational decision-making. In this paper, a new deep learning model is proposed, the Hybrid Attentive Tabular Network (HAT-Net) that integrates three complementary modules: a feature gating attention system that selects adaptive features instance-wise, a transformer-driven interaction system with multi-head self-attention that learns high-order contextual dependencies between features, and a complementary parallel MLP branch that learns complementary nonlinear representations. The two-way outputs are combined with thick classification layers to generate softmax prediction of the four PE performance categories. Experiments on 13,393 examples of 23 engineered physiological and fitness features show that HAT-Net performs competitively on eight evaluation metrics with the task of student body performance classification, and provides significantly better interpretability.The synergistic contribution of all the three components is confirmed by the ablation analysis. Visualization of attention indicates that muscular strength, explosive power as well as cardiovascular indicators are the dominant predictor variables and that there are unique class-conditioned attention patterns that provide educators with valuable insights of specific instructional interventions. Ablation analysis confirms the synergistic contribution of all three components. Attention-based visualization reveals that muscular strength, explosive power, and cardiovascular indicators serve as dominant predictive factors, with distinct class-conditional attention patterns offering educators actionable insights for targeted instructional interventions. The proposed framework establishes a new paradigm for interpretable, attention-driven tabular deep learning in student physiological performance classification, serving as a foundational component for broader data-driven PE evaluation pipelines. Clinical Trial Registration Our study is not a clinical trial, and thus does not require a trial registration. We will ensure this clarification is included in the manuscript.