Bridging Unpaired Modalities with Unified Physiological Code Learning for Multimodal rPPG Measurement

Multimodal remote photoplethysmography (rPPG) improves physiological measurement by integrating complementary sensing modalities, but existing methods typically require synchronized multimodal recordings. This requirement limits the reuse of unpaired physiological data and complicates the addition of new sensing modalities. In this work, we present the first study of unpaired multimodal rPPG learning, which learns from modality-specific recordings without cross-modal correspondence during training. The main challenge lies in learning a shared physiological representation across heterogeneous modalities in the presence of mismatched physiological states and modality-specific sensing characteristics. To address this challenge, we propose PhysioBridge, which connects unpaired modalities through a unified physiological code space. Specifically, Physiological Code Space Construction learns complementary modality-private and shared codes, enabling physiological information to be represented in a modality-invariant manner while preserving modality-specific characteristics. Codeword Distribution Alignment then matches codeword usage across modalities under comparable physiological conditions, encouraging representations from different modalities to utilize a consistent shared code space. At inference, synchronized modalities are adaptively fused based on the relevance of their representations to the shared physiological code space. Experiments on public and self-collected datasets show that PhysioBridge achieves competitive performance without paired training data and generalizes across diverse sensing configurations.