DynaMamba: Multi-scale dynamic interacting Mamba network for irregular clinical time series classification.

Abstract

Irregular clinical time series, composed of patient data from various physiological indicators, are essential for clinical decision-making. Effectively modeling these sequences for tasks like disease diagnosis and mortality prediction is a significant challenge due to their multi-scale temporal variations, irregular sampling intervals, and complex inter-variable dependencies. In this paper, we propose DynaMamba, a novel multi-scale dynamic interacting Mamba network that comprehensively addresses these characteristics through three core innovations. First, a multi-view extraction mechanism explicitly separates the data into observation, missingness, and temporal interval views, capturing crucial clinical monitoring patterns. Second, a hierarchical multi-scale embedding framework captures both fine-grained fluctuations and long-term trends by progressively fusing information across different temporal resolutions. Third, a dynamic multi-sequence modeling module uses bidirectional Mamba blocks with stochastic permutation to dynamically capture inter-variable dependencies. Extensive experiments on three real-world clinical datasets demonstrate that DynaMamba achieves state-of-the-art performance, outperforming existing methods and establishing its effectiveness and robustness in handling irregular clinical time series.