Deep learning for learning outcome prediction and evaluation in Wushu physical education

The teaching quality assessment of college Wushu (martial arts) courses has traditionally relied on instructor-based evaluation and individual skill-test indicators, making it difficult to comprehensively characterize students’ learning outcomes across multiple dimensions such as skill, physical fitness, theoretical knowledge, and classroom behavior, and equally difficult to provide early warning for students with lagging learning outcomes. To address this, this paper proposes a multi-branch deep learning framework for quantitative physical education data, MartialSense-Net, which simultaneously performs classification prediction of students’ learning performance levels and regression prediction of their overall performance scores. The framework first partitions the original heterogeneous quantitative features – physical fitness, skill, theory, and classroom behavior – into four semantic branches, and learns low-dimensional representations for each branch independently through a Multi-Branch Feature Extraction (MBFE) module; it then introduces a Cross Feature Attention Fusion (CFAF) module, which models the interactions among the different branches based on a multi-head self-attention mechanism; finally, a multilayer perceptron encoder and multi-task prediction heads are used to output the classification and regression results. The framework is evaluated as a proof of concept on a publicly available general physical education performance benchmark containing 500 student records and 20 original attributes, from which 16 original predictor variables are selected. Because this dataset contains no Wushu-specific measures of routine precision, force expression, rhythm, or posture, the present experiment evaluates the modeling architecture and does not constitute external validation in authentic Wushu courses. The study systematically compares the framework with random forest, XGBoost, support vector machine, and multilayer perceptron baselines using 5-fold stratified cross-validation, and conducts ablation experiments targeting the attention mechanism and multi-branch structure. The reproducible five-fold rerun shows that MartialSense-Net achieves a Macro-F1 score of 0.545, exceeding the plain MLP and the NoMBFE variant but not the NoAttention variant; however, the overall classification performance of the proposed model has not yet surpassed tree-based baselines such as random forest and XGBoost, and the coefficient of determination ( R 2 ) for the regression task is negative, indicating that precise regression of the overall performance score remains challenging under the current data scale and feature conditions. This paper further analyzes the limitations of deep learning-based educational assessment under small-scale and class-imbalanced tabular data conditions (samples of “High Performers” account for only 2.4%), and honestly reports all experimental results without any exaggeration of model performance. The findings therefore support MartialSense-Net only as an initial methodological benchmark for structured physical education data. Demonstrating Wushu-specific effectiveness will require multi-institutional records from authentic Wushu courses, direct Wushu skill indicators, and external validation across instructors, course levels, and student cohorts.