CEBRA-Enabled Latent Embeddings of Wearable Biosignals for Personalized Biorhythm Modeling

Sutashu Tomonaga, Jo Fujimori, Yi-Shan Cheng, Haruo Mizutani, Kenji Doya

Abstract: Wearable devices offer an unprecedented opportunity to monitor personal health through continuous collection of physiological signals, such as heart rate (HR). However, these observed signals are typically low-dimensional and noisy, yet result from complex, nonlinear mixtures of high-dimensional latent physiological processes. To address this, we adapt CEBRA, a contrastive learning method rooted in nonlinear ICA, to deconstruct these influences and learn meaningful representations of individual biorhythms. In an exploratory study on passive recordings, we apply our method to VitalPatch data (RRI, HR, RR) from 30 older adults. Our key innovation is using phenotypic and habitual data as auxiliary variables, enabling CEBRA to learn highly consistent latent manifolds (\(R^2 > 0.76\)) that capture shared circadian dynamics while reflecting inter-individual differences in traits and habits. These embeddings serve as a prototype for persuasive technologies, facilitating anomaly detection in biorhythm trajectories to support just-in-time interventions (e.g., nudges via visualizations of users’ positions in a “health space” relative to healthier clusters).

External IDs:dblp:conf/persuasive/TomonagaFCMD26