Duke Center for Health Informatics: Self-Supervised Deep Learning for Medical Time Series Analysis
Deep learning has achieved remarkable success in the field of medical time series analysis. However, the effectiveness of advanced models in this domain heavily relies on the availability of high-quality labels, which are not only labor-intensive to obtain but also specific to the medical domain and often in short supply. To address these challenges, self-supervised contrastive learning has emerged as a promising approach, capitalizing on the inherent consistency present within unlabeled data.
Adapting self-supervised contrastive learning, originally proposed for image processing, to the complex realm of medical time series analysis presents unique challenges. Unlike other data formats, medical time series data exhibit a hierarchical structure that encompasses information at multiple levels, including patients, sessions, trials, samples, and observations.
This seminar introduces COMET, an innovative hierarchical framework designed to exploit data consistencies at all these inherent levels within medical time series data. COMET systematically captures data consistency from four distinct data levels, thus maximizing the utilization of information in a self-supervised manner.
Dr. Xiang Zhang is an Assistant Professor at the Department of Computer Science at UNC Charlotte. He is serving as the director of Charlotte Machine Learning Lab (CharMLab) since 2023. Before joining UNC Charlotte, he was a postdoctoral fellow at Harvard University working on deep learning for medical data analysis. Xiang received his Ph.D. degree (in 2020) in Computer Science from the University of New South Wales (UNSW). His research interests lie in data mining and machine learning with applications in pervasive healthcare, medical time series, and Brain-Computer Interfaces (BCIs). Xiang's research outcomes have been published in prestigious conferences (such as ICLR, NeurIPS, and KDD) and journals (like Nature Computational Science).
Zoom Meeting link:
Meeting number (access code) 920 0552 7024, Meeting password: 128631