Abstract: Aiming at  the limitations of existing research on respiratory rate prediction  in deep joint analysis of multimodal physiological signals, as well as the challenge of balancing long-term temporal dependencies and capturing local details, we  proposed a prediction model based on a dynamic multidimensional feature fusion network. Firstly, we constructed an adaptive multi-scale fusion module to dynamically extract multi-frequency features from both electrocardiogram and photoplethysmography, respectively, to generate a single-modal feature map containing rich  multi-scale information, thereby resolving the problem of limited receptive field of a single convolutional kernel.   Secondly, the model incorporated a hybrid spatio-temporal attention mechanism. By stacking Transformer encoding blocks and integrating local, global, and spatio-temporal triple attention strategies, 
it achieved deep interaction between heterogeneous features and precise modeling of long-term temporal dependencies. Validation results based on the BIDMC and CapnoBase public datasets show  that the mean absolute errors of the model reach 1.08 beats/min and 0.76 beats/min, respectively, which is  significantly better than  existing mainstream models in terms of accuracy 
and robustness, and can provide theoretical basis  for clinical non-invasive health monitoring.

Key words:  , multimodal, respiratory rate prediction, hybrid spatio-temporal attention, adaptive multi-scale feature fusion