DUDE: Deep Unsupervised Domain adaptation using variable nEighbors for physiological time series analysis

Jeremy Levy, Noam Ben-Moshe, Uri Shalit, Joachim Behar

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Deep learning, Domain adaptation, Continuous physiological time series

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Deep learning for continuous physiological time series such as electrocardiography or oximetry has achieved remarkable success in supervised learning scenarios where training and testing data are drawn from the same distribution. However, when evaluating real-world applications, models often fail to generalize due to distribution shifts between the source domain on which the model was trained and the target domain where it is deployed. A common and particularly challenging shift often encountered in reality is where the source and target domain supports do not fully overlap. In this paper, we propose a novel framework, named Deep Unsupervised Domain adaptation using variable nEighbors (DUDE), to address this challenge. We introduce a new type of contrastive loss between the source and target domains using a dynamic neighbor selection strategy, in which the number of neighbors for each sample is adaptively determined based on the density observed in the latent space. This strategy allows us to deal with difficult real-world distribution shifts where there is a lack of common support between the source and the target. We evaluated the performance of DUDE on three distinct tasks, each corresponding to a different type of continuous physiological time series. In each case, we used multiple real-world datasets as source and target domains, with target domains that included demographics, ethnicities, geographies, and/or comorbidities that were not present in the source domain. The experimental results demonstrate the superior performance of DUDE compared to the baselines and a set of four benchmark methods, highlighting its effectiveness in handling a variety of realistic domain shifts. The source code is made open-source [upon acceptance of the manuscript].

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3875