LLM-guided spatio-temporal disease progression modelling
Keywords: disease progression modelling, LLM, spatio temporal modelling
Abstract: Understanding the interactions between biomarkers across brain regions during disease progression is essential for unravelling the mechanisms underlying neurodegenerative disease. For example, in Alzheimer's Disease (AD) and other neurodegenerative conditions, there are typically two kinds of methods to contract disease trajectory. Existing mechanistic models describe how variables interact with each other spatiotemporally within a dynamical system driven by an underlying biological substrate often based on brain connectivity. However, such methods typically grossly oversimplify the complex relationship between brain connectivity and brain pathology appearance and propagation. Meanwhile, pure data-driven approaches for inferring these relationships from time series face challenges with convergence, identifiability, and interpretability. We present a novel framework that bridges this gap by using Large Language Models (LLMs) as expert guides to learn disease progression from irregular longitudinal patient data. Our method simultaneously optimizes two components: 1) estimating the temporal positioning of patient data along a common disease trajectory and 2) discovering the graph structure that captures spatiotemporal relationships between brain regions. By leveraging multiple LLMs as domain experts, our approach achieves faster convergence, improved stability, and better interpretability compared to existing methods. When applied to modelling tau-pathology propagation in the brain, our framework demonstrates superior prediction accuracy while revealing additional disease-driving factors beyond traditional connectivity measures. This work represents the first application of LLM-guided graph learning for modelling neurodegenerative disease progression in the brain from cross-sectional and short longitudinal imaging data.
Primary Area: applications to neuroscience & cognitive science
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
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: 6980
Loading