SELPHY-TS: Self-supervised Learning for PHYsiological Time Series

• Applicant: Professor Dr. Nils Strodthoff
• Subject Area: Methods in Artificial Intelligence and Machine Learning ; Cardiology, Angiology
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 553038473

Ever since the deployment of the ChatGPT chatbot, foundational models, i.e., models that have been pretrained on large datasets and can be applied to different downstream tasks, are on everyone's lips. This proposal addresses the development of such foundational models in the domain of physiological time series data, in this case for electrocardiography and photoplethysmography data, through modern techniques from self-supervised learning. Here, we strive to approach the problem from two directions by learning representations (1) from unlabeled signals alone and (2) from combinations of signals and corresponding free-text reports based on large publicly available datasets. On the one hand, we aim to demonstrate the feasibility and the advantages of scaling such models to large datasets and the targeted improvement of methods of self-supervised learning. On the other hand, the trained representations are supposed to be analyzed from different perspectives: (1) The unsupervised identification of categories through the decomposition of self-supervised representations opens a complementary, data-driven perspective on man-made expert ontologies, both on sequence-level (such as rhythm annotations) and on sample-level (such as pathologies). (2) Leveraging techniques from explainable AI, we can identify concepts associated with certain categories beyond the level of single examples. We envision that the results of this project are both of high relevance for the further development of representation learning for (physiological) time series from a methodological perspective but also for the respective medical application domains, cardiology and acute medicine, as basis for the development of multimodal decision support systems but also for scientific discovery from a data-driven perspective.

• DFG Programme: Research Grants
• International Connection: Netherlands, United Kingdom
• Co-Investigator: Professor Dr. Wilhelm Haverkamp
• Cooperation Partners: Professor Dr. Hjalmar Bouma; Dr. Peter Charlton