MissHyper Improves Clinical Time Series Forecasting with Hypergraphs

This research introduces MissHyper, a novel hypergraph forecasting model specifically designed to improve the analysis of irregular multivariate time series data in clinical settings. These datasets are challenging because measurements are often taken at irregular intervals, and the timing of observations itself carries important information about patient state. Traditional event-centric models can lose crucial local context by treating co-timestamped measurements as isolated events too early in the processing pipeline. MissHyper addresses this "pre-propagation representation bottleneck" by restoring the co-timestamp context before the main message-passing layers begin. It augments each event with a local support-density cue, aggregates records from the same timestamp to recover patient-state context, and adaptively fuses this context with node-specific evidence using a missingness-guided gate. This approach leads to significant improvements in multi-step forecasting accuracy on major clinical benchmarks like PhysioNet 2012, MIMIC-III, and MIMIC-IV, demonstrating the critical role of better event initialization in sparse clinical forecasting.