A new study from our lab, published in Radiology, evaluated whether large language models can transform electronic health record notes into concise, clinically useful indications for diagnostic imaging. The project was led by Adrian Serapio, BS, and Timothy Chen, MD, combining expertise in machine learning and clinical radiology to assess both the accuracy and practical value of LLM-generated summaries.
In a multireader study involving 20 radiologists, the best-performing proprietary and open-source models produced indications rated as more comprehensive and factual than those originally provided by referring clinicians. The findings suggest that carefully validated LLM tools could improve imaging protocol selection and interpretation by surfacing relevant clinical information already contained in the medical record.
Read more here: https://pubs.rsna.org/doi/10.1148/radiol.253238
Serapio A, Chen TL, Tangsombatvisit B, Fields BKK, Yu Y, Guo Y, Kim SK, Miao BY, Sushil M, Hess CP, Majumdar S, Sohn JH. Radiologically relevant clinical history summarization with large language models: a multireader performance study. Radiology. 2026;320(2):e253238. doi:10.1148/radiol.253238.