DispFormer: A dual attention transformer with denoising for biomedical irregular time series classification

Electronic Health Records (EHRs) collected in intensive care units (ICUs) consist of irregular multivariate time series (IMTS), where measurements are recorded asynchronously and missingness is pervasive. These characteristics pose significant challenges for reliable clinical prediction. Existing methods typically adopt a sequential modeling paradigm, where temporal dependencies are first learned within individual variables, followed by inter-variable interactions at aligned time steps. However, this design restricts cross-variable interactions to identical timestamps, limiting the ability to capture dependencies spanning different variables and time periods. In addition, most approaches rely on vectorized representations with padding and zero imputation, which introduce artificial discontinuities and amplify noise. To address these challenges, we propose DispFormer, a novel framework for biomedical IMTS classification that jointly models global temporal-variable dependencies while mitigating imputation-induced noise. Specifically, DispFormer introduces a dispatcher dual attention mechanism, in which learnable dispatcher tokens serve as global communicators to mediate information flow across both time and variables. This mechanism operates through three stages-Gathering, Interaction, and Dispatching-enabling holistic dependency modeling beyond conventional stepwise dual-attention designs. Furthermore, we incorporate a lightweight information bottleneck-based denoising module to suppress redundant and noisy representations introduced by imputation. Extensive experiments on three benchmark datasets, including ICU clinical prediction tasks, demonstrate that DispFormer consistently outperforms state-of-the-art methods, particularly under high-missingness conditions. Source code is available at https://github.com/junjzhang7/DispFormer.