LLMs Extract Cardiac Data From Cancer Patients' Electronic Records

Researchers at Thomas Jefferson University have developed a large language model-based algorithm capable of rapidly extracting cardiac event data from the electronic health records of cancer patients, addressing a critical bottleneck in oncology research. Published in the International Journal of Radiation Oncology, Biology, Physics in 2026, the study highlights a significant efficiency gain in monitoring cardiotoxicity, a known complication of treatments for breast and lung cancers due to the anatomical proximity of these organs to the heart. Traditional manual review of electronic health records is highly time-consuming, requiring physicians to sift through hundreds of clinical notes to identify heart-related complications. The Jefferson team addressed this by deploying a structured prompting framework on open-source general-purpose reasoning models, specifically DeepSeek-R1-70b, Llama-3.3, and Mistral-Large. These models were evaluated for their diagnostic accuracy and formatting adherence during a development phase before advancing to validation cohorts. In testing across 411 breast and lung cancer patient records, the algorithm successfully identified cardiac events in 71 to 85.5 percent of cases. Crucially, the system demonstrated a strong ability to contextualize negative clinical statements, preventing false positives when patients were recorded as having denied or ruled out cardiac conditions. The operational efficiency of the system was substantial. While human reviewers required approximately two hours to thoroughly examine each electronic health record, the large language models completed the extraction process in between 20 and 42 seconds per record. The research was led by first author Wenchao Cao, PhD, alongside oncology resident Nilanjan Halder, MD, and a cohort of participating medical students and an undergraduate researcher. The successful integration of large language models into clinical data extraction workflows offers a scalable solution for oncology research. By automating the identification of cardiac complications, the system frees medical professionals to focus on direct patient care and complex clinical decision-making. Researchers project that further refinement of this algorithmic framework will enable more precise risk stratification for cardiotoxicity. Ultimately, the technology aims to support the development of personalized radiation therapy protocols, allowing clinicians to mitigate cardiac side effects and improve long-term survival outcomes for patients undergoing cancer treatment.