Predicting ICU in-hospital mortality from text-encoded structured EHR data using adaptive transformer layer fusion.

Early identification of ICU patients at high mortality risk is essential for triage and timely intervention. We present adaptive layer fusion with intelligent attention (ALFIA), a modular architecture that jointly trains low-rank adaptation (LoRA) adapters and an adaptive layer-weighting mechanism to fuse multi-layer semantic features from a pretrained transformer backbone. ALFIA operates on text-encoded representations of the structured EHR data, in which tabular clinical variables (demographics, vital signs, laboratory values, and severity scores) are converted into standardized natural-language descriptions rather than processed as free-text clinical notes. Evaluated on the CriticalWindow-24 benchmark with MIMIC-IV and eICU cohorts, ALFIA achieves strong AUPRC while maintaining a balanced precision-recall profile. The learned embeddings can be further combined with gradient boosting (ALFIA-boost) or neural networks (ALFIA-nn) for additional gains. These findings demonstrate that text-encoded structured EHR data can support practical, generalizable early-warning models for ICU mortality risk stratification.