Abstract
Early identification of high-risk ICU patients is crucial for directinglimited medical resources. We introduce ALFIA (Adaptive Layer Fusion withIntelligent Attention), a modular, attention-based architecture that jointlytrains LoRA (Low-Rank Adaptation) adapters and an adaptive layer-weightingmechanism to fuse multi-layer semantic features from a BERT backbone. Trainedon our rigorous cw-24 (CriticalWindow-24) benchmark, ALFIA surpassesstate-of-the-art tabular classifiers in AUPRC while preserving a balancedprecision-recall profile. The embeddings produced by ALFIA's fusion module,capturing both fine-grained clinical cues and high-level concepts, enableseamless pairing with GBDTs (CatBoost/LightGBM) as ALFIA-boost, and deep neuronetworks as ALFIA-nn, yielding additional performance gains. Our experimentsconfirm ALFIA's superior early-warning performance, by operating directly onroutine clinical text, it furnishes clinicians with a convenient yet robusttool for risk stratification and timely intervention in critical-care settings.