Interpretable Graph Learning on Irregular Clinical Time Series

Andrea Zerio, Maya Bechler-Speicher, Maor Huri, Marie Vibeke Vestergaard, Ran Gilad-Bachrach, Tine Jess, Samir Bhatt, Aleksejs Sazonovs

Keywords: Graph Learning, Graph Representations, Interpretability, Healthcare, Time Series

TL;DR: We present GMAN, a novel and interpretable graph learning model for learning from irregular, asynchronous sets of temporal medical data.

Abstract: Real-world medical data often include measurements collected at irregular, asynchronous intervals across different tests and frequencies. For example, routine blood tests are taken at different times, producing fragmented, unevenly sampled temporal data. Effectively learning from such data requires models that can handle temporally sparse, heterogeneous measurements. In this paper, we propose Graph Mixing Additive Networks (GMAN), an interpretable-by-design model for learning from irregular temporal measurements. Our method achieves strong performance on a real-world medical task: predicting the onset of Crohn’s disease (CD) from routine biomarkers, comparable to black-box models, both graph and sequential architectures. Its interpretable design also yields clinically meaningful insights, such as dysregulation patterns in the pre-diagnostic phase of complex diseases.

Submission Number: 40