DBGL: Decay-aware Bipartite Graph Learning for Irregular Medical Time Series

Jian Chen, Xiaoyan Yuan, Yuxuan Hu, Jinfeng Xu, Yipeng Du, Xiangyu Zhao, Wei Wang, Edith C. H. Ngai

Keywords: Irregular Medical Time Series, Graph Representation Learning, Smart Healthcare

Abstract: Irregular Medical Time Series (IMTS) are of great importance in the healthcare domain to better understand the patient's condition. However, the inherent temporal irregularity, arising from heterogeneous sampling rates, asynchronous observations, and variable gaps, poses significant challenges for reliable modeling. Existing methods distort the **temporal sampling irregularity** and missing pattern, while failing to capture **variable decay irregularity** in the clinical domain, leading to suboptimal representation. To address these limitations, we introduce DBGL: Decay-Aware Bipartite Graph Learning for Irregular Medical Time Series. DBGL first introduces a patient–variable bipartite graph that simultaneously captures irregular sampling patterns without artificial alignment and adaptively models variable relationships for temporal sampling irregularity modeling, enhancing representation learning. To model variable decay irregularity, DBGL designs a novel node-specific temporal decay encoding mechanism that enables each variable to decay at different rates based on sampling interval, yielding a more accurate and faithful representation of irregular temporal dynamics. We evaluate the performance of DBGL on four publicly available datasets: P19, Physionet, MIMIC-III, and P12. Results show that DBGL outperforms all baselines, and our code is also available in the supplementary material.

Primary Area: learning on time series and dynamical systems

Submission Number: 6101