Multi-Agent Generative Artificial Intelligence for Transfusion Reaction Feature Extraction

A multi-agent AI pipeline extracting NHSN Hemovigilance Module transfusion-reaction features from clinical notes achieved 0.98+ precision and 0.774 sensitivity (95% CI 0.745–0.800) with Kimi K3.

• Why it matters: Many transfusion reactions are not captured in hemovigilance systems, and key feature definitions are often only available in free-text notes, creating a gap for scalable signal detection.
• What they did: The team curated 181 NHSN features and used three open-weight large language models (Qwen 3.6-27B, GPT-OSS-120B, and Kimi K3) to extract feature occurrences, timing, and details from 50 notes, with blinded physician review.
• The result: Physician concordance covered 1,201 of 1,343 co-rated occurrences (89.4%; Krippendorff’s alpha 0.73), timing was supported in >92% and clinical details in >97%, with false negatives mainly due to omissions.

AI summary of the abstract · may contain errors · not medical advice · read the paper.