BioSynergy: Component-Structured Representation Learning for Multimodal Physiological Signals

Multimodal physiological signals provide complementary observations of coupled bodily processes. To capture this complementarity, prior physiological representation-learning methods have used fusion to combine signals, alignment to promote cross-modal agreement, and shared and private factorization to distinguish common from modality-dependent structure. However, physiological states and events can involve time-varying relations across signal pathways and such coordination is not typically modeled as a distinct representational component. We propose BioSynergy, a self-supervised framework for component-structured representation learning of multimodal physiological signals. Motivated by Partial Information Decomposition, BioSynergy derives contextualized state tokens to construct a shared physiological state component and modality-specific components, and multi-scale local-change tokens to condition a coordination-sensitive component through cross-signal relations. Component-aware teacher-student objectives reinforce these roles through representation and state alignment and modality-level contextual reconstruction. Under linear probing across three physiological datasets and six tasks, BioSynergy remains competitive with physiological foundation models and multimodal representation-learning baselines. Intervention and temporal analyses further show task-dependent patterns consistent with established physiological knowledge.