Dual-Stream deep learning for multimodal feature fusion and classification of balance control in elite freestyle aerial skiers.

Balance control is a key determinant of stable landing in elite freestyle aerial skiing. Rapid and precise identification of subtle differences in athletes' balance-stability regulation is a prerequisite for targeted, evidence-based training. Conventional balance assessment typically relies on force-platform measurements of the center of pressure (COP) trajectory and subsequent time-, frequency-, and time-frequency-domain analyses. However, these indices have limited ability to capture the complex dynamics of postural control and to discriminate fine-scale differences in balance regulation among highly trained freestyle skiing aerials athletes.To address this limitation, we developed a dual-stream deep learning model that fuses time-frequency image features with COP-based statistical descriptors to classify subtle variations in balance regulation. Twenty-five elite freestyle skiing aerials athletes were recruited and performed quiet standing under two conditions: (i) bipedal stance on a stable surface with eyes open and (ii) bipedal stance on an unstable surface with eyes open. COP trajectories were recorded and their multiscale entropy computed; K-means clustering was used to stratify participants into high-, medium-, and low-stability groups. The extracted time-frequency and statistical features were then fed into the dual-stream deep learning framework for model training and validation.The proposed model achieved approximately 95% classification accuracy in distinguishing data-driven COP-based stability strata, suggesting potential utility for the sensitive assessment of balance-regulation patterns in elite freestyle skiing aerials athletes.