Representation Learning Enables Uncovering Psychophysiological States from Wearable Data

Human physiology contains valuable information about an individual’s mental and physical states, but extracting and interpreting this information poses significant challenges. In this study, we present a preliminary approach that leverages representation learning to uncover meaningful insights from physiological data. Using the self-supervised learning technique and a large amount of unlabeled raw electrocardiogram (ECG) recordings from wearables, we developed an ECG representation model. This model generates high-dimensional embeddings of raw ECG samples, which can be employed for downstream tasks or visualized in a meaningful 2D space using dimensionality reduction techniques such as T-SNE. The resulting embeddings reveal patterns and clusters that may correspond to recurring psychophysiological states. Furthermore, the proposed approach facilitates direct comparison of physiological data both within and across individuals, offering potential applications in addressing mental- and health-related challenges, including panic or anxiety attacks, depression, and neurodegenerative diseases.