"Irregular Multivariate Time Series"

"To evaluate the proposed framework under diverse irregular multivariate time-series settings, we conduct experiments on four publicly available real-world benchmark datasets: MIMIC-III, PhysioNet\u201912, Human Activity, and USHCN. These datasets span multiple application domains, including healthcare, biomechanics, and climate science, thereby providing a broad test bed for assessing model generalization. Specifically, MIMIC-III and PhysioNet\u201912 contain highly sparse and irregularly sampled clinical measurements, Human Activity consists of wearable sensor recordings with varying sampling patterns and sequence lengths, and USHCN includes long-term meteorological observations with substantial missing values and strong temporal dependencies. Together, these benchmarks capture the key challenges of IMTS forecasting, such as asynchrony, sparsity, missingness, and heterogeneous temporal dynamics, enabling a comprehensive and realistic evaluation of forecasting performance."