Speaker
Description
Real-world biomedical time-series data, including clinical records and measurements of complex biological phenomena, contain valuable insights into disease progression and underlying mechanisms. However, such data are often noisy, irregularly sampled, and partially observed, posing significant challenges for analysis. Extracting latent temporal structures from these observations is therefore crucial for advancing our understanding of dynamic biological systems. In this talk, we introduce a framework based on deep state space models (DSSMs) with variational inference to capture complex temporal dependencies in such data. The proposed approach infers latent trajectories corresponding to unobserved disease states or physiological processes, enabling interpretable analysis of system dynamics. As a real-world example, we demonstrate the application of DSSMs to electronic health record data, highlighting their ability to uncover meaningful patient representations. Furthermore, we present recent advances in constrained state space modeling, where prior knowledge—such as stability and physiological consistency—is incorporated into latent dynamics through structured constraints. These techniques improve identifiability, robustness to noise, and consistency with known biological principles. Overall, this framework provides a powerful and principled approach for analyzing real-world biomedical time-series data.
