Abstract
<title>Abstract</title> <p>This study investigated whether deep learning can classify skiing proficiency from plantar pressure distributions measured during simulator-based training and whether the resulting predictions reflect longitudinal changes in skill acquisition. Weight-shift images obtained from 27 participants were classified into beginner, novice, and intermediate proficiency levels using a YOLO11-based image classifier under a leave-one-participant-out evaluation. Longitudinal changes in classification outcomes were then analyzed using linear and spline-based mixed-effects models. To interpret the learned features, Eigen-CAM visualizations and quantitative analyses of attention regions were performed. The proposed classifier demonstrated reliable discrimination between beginner and novice pressure patterns but showed limited generalization to the minority intermediate class. Longitudinal analyses suggested that the probability of non-beginner classifications generally increased during practice, while the spline-based mixed-effects model indicated a nonlinear learning trajectory characterized by rapid early improvement followed by a plateau. Furthermore, repeated-measures correlation demonstrated a significant association between forefoot-focused attention and the probability of non-beginner classification, suggesting that the model captured biomechanically meaningful movement characteristics. These findings indicate that deep learning can provide an objective measure of skill acquisition in simulator-based ski training while offering interpretable visual explanations of the learned movement features.</p>