An interpretable, clinically-aligned AI paradigm for VTE risk prediction: an approach using LLMs and compound attention

Manual VTE risk assessment is often inconsistent, while many high-performing AI models remain difficult to audit clinically. We propose an interpretable and clinically aligned paradigm that uses locally deployed LLMs to structure EHR narratives and an attention-based model (EMAX) to predict expert-adjudicated Caprini RAM (2010) risk. From 113,836 raw encounter records, we curated 14,808 eligible records with documented expert Caprini scores for supervised learning. On the test set, EMAX achieved an AUC of 0.9513. The framework provides an end-to-end pathway for accurate and adoptable VTE risk stratification.