MATA-Former & SIICU: Semantic Aware Temporal Alignment for High-Fidelity ICU Risk Prediction
This work addresses the challenge of high-fidelity risk prediction in ICU care, which is critical for patient outcomes, though it appears incremental in its methodological improvements.
The paper tackles the problem of forecasting clinical risks in ICU settings by addressing limitations of current methods that rely on chronological proximity and binary supervision, proposing a semantic-aware temporal alignment approach that achieves superior efficacy and robust generalization on both a newly constructed dataset (SIICU) and MIMIC-IV.
Forecasting evolving clinical risks relies on intrinsic pathological dependencies rather than mere chronological proximity, yet current methods struggle with coarse binary supervision and physical timestamps. To align predictive modeling with clinical logic, we propose the Medical-semantics Aware Time-ALiBi Transformer (MATA-Former), utilizing event semantics to dynamically parameterize attention weights to prioritize causal validity over time lags. Furthermore, we introduce Plateau-Gaussian Soft Labeling (PSL), reformulating binary classification into continuous multi-horizon regression for full-trajectory risk modeling. Evaluated on SIICU -- a newly constructed dataset featuring over 506k events with rigorous expert-verified, fine-grained annotations -- and the MIMIC-IV dataset, our framework demonstrates superior efficacy and robust generalization in capturing risks from text-intensive, irregular clinical time series.