Time-dependent Sampling for Contrastive Self-supervised Learning of Longitudinal Biosignals Representations

Many chronic diseases exhibit complex and slow time courses, and in asymptomatic stages it may be possible to detect signs of disease through longitudinal monitoring with wearables.Properly accounting for temporal dependencies in the learned representations of wearable biosignals is crucial to better characterize the progression of disease and improve human health.While previous research has demonstrated that informative representations of wearables-derived biosignals offer much promise in various medical applications, the limited longitudinal scale of most existing wearables datasets has hindered the development of computational and evaluation frameworks that capture these temporal variations with appropriately fine granularity. To address this, we examine the implicit integration of biosignal timestamps in contrastive self-supervised learning when defining the positive pairs of joint-embedding architectures, enforcing physiological consistency by encouraging positive pairs to be close in time.We demonstrate that using this temporal knowledge during pre-training leads to representations more sensitive to time, as they are better able to predict the time of day and overnight binary sleep-wake stages. We also show that these time-aware representations can improve biomarker monitoring, applying them to predict changes in cardiopulmonary fitness, diabetes status, body mass index, and cardiovascular risk.Crucially, we emphasize the importance of a longitudinal within-subject evaluation rather than the more common cross-sectional across-subject evaluation.Our results suggest that time-varying representations can improve the accuracy of health monitoring using wearable-based biosignals, and open the door for future applications of more time-aware representation learning.