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

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Biomedical Data Science

Background:

• Early identification of high-mortality risk patients in the Intensive Care Unit (ICU) is critical for effective triage and timely interventions.
• Existing methods may not fully leverage the rich semantic information within structured Electronic Health Record (EHR) data.

Purpose of the Study:

• To develop and evaluate a novel AI architecture, Adaptive Layer Fusion with Intelligent Attention (ALFIA), for predicting ICU patient mortality risk.
• To assess the efficacy of using text-encoded structured EHR data for generalizable early-warning models.

Main Methods:

• ALFIA architecture jointly trains Low-Rank Adaptation (LoRA) adapters and an adaptive layer-weighting mechanism.
• Multi-layer semantic features are fused from a pretrained transformer backbone using text-encoded structured EHR data.
• Tabular clinical variables are converted into standardized natural-language descriptions.

Main Results:

• ALFIA demonstrated strong performance on the CriticalWindow-24 benchmark using MIMIC-IV and eICU cohorts, achieving a high Area Under the Precision-Recall Curve (AUPRC).
• The model maintained a balanced precision-recall profile, indicating reliable risk stratification.
• Further gains were observed when combining ALFIA embeddings with gradient boosting (ALFIA-boost) or neural networks (ALFIA-nn).

Conclusions:

• Text-encoded structured EHR data can effectively support the development of practical and generalizable early-warning models for ICU mortality risk.
• ALFIA provides a robust and adaptable framework for leveraging multi-layer semantic features in EHR data.
• This approach holds promise for improving clinical decision-making and patient care in critical care settings.