NEW YORK — What happened: A major academic medical center published operational metrics on an in-house retrieval-augmented generation (RAG) system designed to screen patient electronic health records (EHRs) against complex oncology trial inclusion criteria.
Why it matters: Over 80 percent of clinical trials fail to meet recruitment timelines, often because clinicians cannot manually parse hundreds of dense trial protocols against thousands of patient charts.
Deep dive: The system utilizes localized, privacy-compliant embeddings that run on hospital-owned infrastructure rather than external cloud APIs. It scans unstructured clinical notes, lab values, and genomic sequencing data to flag eligible candidates for clinical investigators.
Report check (claims vs what is verified vs still rumor): Rumors suggested the software autonomously enrolls patients without human intervention. In reality, the tool functions strictly as a recommender system; every match requires manual chart review by a certified clinical research coordinator.
Open questions: How vulnerable are these local retrieval pipelines to subtle adversarial prompt injections hidden within unstructured physician notes?