Abstract
Biosignals offer valuable insights into the physiological states of the humanbody. Although biosignal modalities differ in functionality, signal fidelity,sensor comfort, and cost, they are often intercorrelated, reflecting theholistic and interconnected nature of human physiology. This opens up thepossibility of performing the same tasks using alternative biosignalmodalities, thereby improving the accessibility, usability, and adaptability ofhealth monitoring systems. However, the limited availability of large labeleddatasets presents challenges for training models tailored to specific tasks andmodalities of interest. Unsupervised cross-modal knowledge transfer offers apromising solution by leveraging knowledge from an existing modality to supportmodel training for a new modality. Existing methods are typically based onknowledge distillation, which requires running a teacher model alongsidestudent model training, resulting in high computational and memory overhead.This challenge is further exacerbated by the recent development of foundationmodels that demonstrate superior performance and generalization across tasks atthe cost of large model sizes. To this end, we explore a new framework forunsupervised cross-modal knowledge transfer of biosignals by training alightweight bridge network to align the intermediate representations and enableinformation flow between foundation models and across modalities. Specifically,we introduce an efficient strategy for selecting alignment positions where thebridge should be constructed, along with a flexible prototype network as thebridge architecture. Extensive experiments across multiple biosignalmodalities, tasks, and datasets show that BioX-Bridge reduces the number oftrainable parameters by 88--99\% while maintaining or even improving transferperformance compared to state-of-the-art methods.