In order to more effectively assess the effectiveness of ankle functional rehabilitation training for patients, a model combining a convolutional neural network module with a multi-head graph attention residual module based on dynamic gating was proposed, a specialized database comprising 20 categories of ankle rehabilitation (covering typical features such as dorsiflexion and adduction) was established to address the scarcity of existing clinical image datasets, and a multimodal network architecture was adopted in which local image features extracted by a convolutional neural network were converted into graph relationship data via a dynamic graph constructor, followed by fine-grained information fusion using a multi-layer gating mechanism and a multi-head graph attention mechanism to capture different types of neighborhood relationships from multiple perspectives, while residual connections were incorporated to mitigate the vanishing gradient problem; subsequently, experimental validation was conducted using the established dataset in a Python environment, and this experimental setup was used to investigate the model’s performance. The results show that, compared to a typical CNN-ResNet model, the proposed model achieves a 3.77% increase in accuracy, a 3.13% increase in macro precision, a 3.35% increase in macro recall, and a 3.23% increase in macro F1 score. It is concluded that the model demonstrates validity and effectiveness, making it suitable for practical engineering applications.