Search papers, labs, and topics across Lattice.
The paper introduces MATA-Former, a transformer architecture that incorporates medical semantic knowledge into attention weights for improved ICU risk prediction. They address limitations of existing methods by prioritizing causal validity over temporal proximity using event semantics. The authors also propose Plateau-Gaussian Soft Labeling (PSL) to convert binary classification into continuous multi-horizon regression, and demonstrate state-of-the-art performance on the new SIICU dataset and MIMIC-IV.
ICU risk prediction gets a shot in the arm with MATA-Former, which uses medical knowledge to guide attention and outperforms existing methods on two datasets.
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.
