DBGL Applies Decay-Aware Bipartite Graph Learning to Irregular Medical Time Series
DBGL uses bipartite graphs and variable-specific decay encoding to model irregular medical time series, outperforming baselines on four datasets while addressing gaps in prior distortion of sampling and decay patterns.
Irregular medical time series play critical role in clinical domain but pose modeling challenges from heterogeneous sampling rates, asynchronous observations and variable gaps (https://arxiv.org/abs/2604.11842).
DBGL constructs patient-variable bipartite graph to capture irregular sampling patterns without artificial alignment and introduces node-specific temporal decay encoding that models each variable's decay rate based on sampling interval (https://arxiv.org/abs/2604.11842). Existing methods including GRU-D distort temporal sampling irregularity and missingness patterns while applying less granular decay (https://arxiv.org/abs/1606.01865).
DBGL outperforms all baselines on four public datasets (https://arxiv.org/abs/2604.11842). Related works on irregular sequences such as multivariate time series imputation with GANs document similar data-quality obstacles in healthcare settings (https://arxiv.org/abs/1802.03446).
AXIOM: DBGL's node-specific decay encoding directly targets inconsistent sampling intervals common in EHR data and may lift classification reliability for clinical prediction tasks where uniform-decay models fall short.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2604.11842)
- [2]GRU-D: Recurrent Neural Networks for Multivariate Time Series with Missing Values(https://arxiv.org/abs/1606.01865)
- [3]Multivariate Time Series Imputation with Generative Adversarial Networks(https://arxiv.org/abs/1802.03446)