Vast quantities of person-generated health data (wearables) are collected but\nthe process of annotating to feed to machine learning models is impractical.\nThis paper discusses ways in which self-supervised approaches that use\ncontrastive losses, such as SimCLR and BYOL, previously applied to the vision\ndomain, can be applied to high-dimensional health signals for downstream\nclassification tasks of various diseases spanning sleep, heart, and metabolic\nconditions. To this end, we adapt the data augmentation step and the overall\narchitecture to suit the temporal nature of the data (wearable traces) and\nevaluate on 5 downstream tasks by comparing other state-of-the-art methods\nincluding supervised learning and an adversarial unsupervised representation\nlearning method. We show that SimCLR outperforms the adversarial method and a\nfully-supervised method in the majority of the downstream evaluation tasks, and\nthat all self-supervised methods outperform the fully-supervised methods. This\nwork provides a comprehensive benchmark for contrastive methods applied to the\nwearable time-series domain, showing the promise of task-agnostic\nrepresentations for downstream clinical outcomes.\n