MATA-Former & SIICU: Semantic Aware Temporal Alignment for High-Fidelity ICU Risk Prediction

2026-04-02

AI Summary

Existing ICU risk prediction methods rely on coarse-grained binary labels and fixed timestamps, which fail to capture the pathology-driven dynamics of clinical risk evolution. To address this limitation, this work proposes MATA-Former, a novel model that employs a semantics-aware ALiBi attention mechanism to dynamically adjust attention weights based on the semantic content of medical events, prioritizing causal relationships over mere temporal proximity. Furthermore, it introduces Plateau-Gaussian soft labels to reformulate rigid binary classification as a continuous regression task across multiple time horizons, enabling holistic risk trajectory modeling. By moving beyond the constraints of fixed timestamps and hard labels inherent in conventional temporal models, MATA-Former achieves significantly improved predictive performance on both SIICU and MIMIC-IV datasets, demonstrating strong generalization capabilities on text-rich, irregularly sampled clinical time-series data.

Technology Category

Machine Learning: Calibration & Uncertainty QuantificationReasoning under Uncertainty: Relational Probabilistic ModelsKnowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsUser Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systemsGraph Algorithms and Modeling for the Web: Algorithms and analysis for heterogeneous, signed, attributed, multi-relational, temporal, higher-order, and annotated Web-related graphs

Abstract

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.

Problem

• ICU risk prediction
• temporal alignment
• clinical time series
• semantic awareness
• binary supervision

Innovation

• MATA-Former
• Semantic-aware Temporal Alignment
• Plateau-Gaussian Soft Labeling
• ICU Risk Prediction
• Irregular Clinical Time Series