AI extracts cardiac events from cancer records in seconds

Physicians, residents and students spent about two hours reviewing each electronic health record (EHR) in a Thomas Jefferson University study. A large language model (LLM)—an artificial-intelligence system used here to read medical text—handled the same extraction task in 20 to 42 seconds; the comparison covered 411 breast and lung cancer patients.

The target was cardiotoxicity, heart-related damage caused by cancer treatment. Breast and lung cancer patients face increased risk because the heart, lungs and breasts are close together. The damage can raise the risk of heart attacks, heart failure and other cardiac conditions, but finding evidence in patient records can require reading through hundreds of files.

The framework did more than search for a word. Its prompts looked for relevant keywords while checking for false positives, so a note saying a patient denied a disease—or that a heart attack was “ruled out”—was not counted as a cardiac event. That distinction matters in medical records, where negative findings can be phrased in many ways.

The researchers used open-source LLMs and reported that the framework found cardiac-event data 71% to 85.5% of the time. During initial screening, DeepSeek-R1-70b, Llama-3.3 and Mistral-Large showed the strongest balance of diagnostic accuracy and formatting adherence, so the three models moved into validation cohorts.

Concretely, for oncology research teams, the system could turn a slow first pass through patient records into a much faster way to assemble cardiotoxicity data. For patients, the hoped-for next step is to identify those more likely to develop heart complications and use that information to shape radiation treatment, potentially reducing side effects. The paper presents that as a goal, not a current clinical service: further refinement is still needed, and the reported work does not show improved patient outcomes.