Data augmentations and transfer learning for physiological time series

Physiological time series signals e.g., measured through wearables have received increasing interest as biomarkers for sleep disorders, stress, anxiety, and other psychiatric disorders, or health conditions. However, open source datasets are scarce making it difficult to develop strong prediction models for new application areas without extensive prior data collection. We investigate the possibilities of using existing datasets as well as different simulation strategies to create a foundational model transferable to new applications. We evaluate transferability for four different tasks (open source data) and compare the performance of transfer learning and simulated data augmentations.