Self-Supervised Learning for Clinical Time Series via Robust Optimization and Adaptive Data Augmentation

Time-series clinical data are crucial for advancing healthcare analytics in areas like predictive modeling for patient outcomes, monitoring and managing chronic diseases, personalized treatment planning, and real-time alerting systems for critical care. Current studies focused on leveraging self-supervised learning methods to fully exploit the potential of such data without requiring a large number of labels. However, current methods use non-adaptive, hand-configured intensities, provide no mechanism for automatically discovering hard augmentations, and seldom integrate magnitude and temporal domains. To fill this gap, our study introduces robust self-supervised learning for unlabeled time-series data. Specifically, we propose a novel parameterized augmentor that emphasizes training on challenging samples to enhance learning effectiveness. Additionally, our temporal augmentation strategy is designed to add diversity to the temporal aspect of the data by time stretching and compression. Furthermore, alternating training framework is designed to integrate the augmentor with other model components for the min−max optimization, while using three unique loss functions to capture a broad spectrum of both global and predictive features effectively. Our experimental results demonstrate that this approach significantly outperforms existing methods, especially in scenarios with limited labeled data, and shows remarkable proficiency in cross-domain feature extraction.