SepsisLens: Structure-Preserving Sequence Modelling for Decomposable Early Sepsis Warning
SepsisLens models early sepsis warning as a structure-preserving sequence problem. It preserves individual clinical variables' temporal states throughout prediction, unlike models that mix data into a single patient representation. This allows for decomposing risk into variable and organ-level components, leading to better performance across multiple ICU cohorts and lower false alarms.