Cross-Modality Self-Supervision for Unimodal Physiological Signals Representation Learning

Deep learning methods for physiological signals offer significant promise for medical diagnosis, health monitoring, and treatment guidance. However, their development is hampered by the reliance on large, high-quality annotated datasets. Self-supervised learning (SSL) provides a viable path forward by learning representations from unannotated data. Although existing SSL methods using token vocabularies have advanced representation learning for unannotated physiological data, their reliance on a single signal modality restricts the generality and performance of the resulting models. To overcome this limitation, we present a Cross-Modal Self-Supervision (CM-SSL) Learning framework designed to learn robust representations from single-modality physiological signals. This work specifically focuses on the prevalent modalities of Electrocardiogram (ECG) and Photoplethysmogram (PPG). The core advantage of CM-SSL is its construction of a shared token vocabulary from cross-modal signals, which enhances the quality and generality of the learned unimodal representations. Extensive evaluation across a diverse suite of downstream tasks demonstrates that the CM-SSL based unimodal model achieves superior performance over baseline models. The source code can be available at URL(https://github.com/ximosss/CM-SSL.git).

Verbatim fromOpenAlex· Abstract via OpenAlex (publisher-deposited); © the authors/publisher