Large language models for extracting histopathologic diagnoses of colorectal cancer and dysplasia from electronic health records

Accurate data resources are essential for impactful medical research, but available structured datasets are often incomplete or inaccurate. Recent advances in open-weight large language models (LLMs) enable more accurate data extraction from unstructured text in electronic health records (EHRs), however, thorough validation of such approaches is lacking. Our objective was to create a validated approach using LLMs for identifying histopathologic diagnoses in pathology reports from the nationwide Veterans Health Administration (VHA) database, including patients with genotype data within the Million Veteran Program (MVP) biobank.

Methods

Our approach utilises search term filtering followed by simple ‘yes/no’ question prompts for the following phenotypes of interest: any colorectal dysplasia, high-grade dysplasia and/or colorectal adenocarcinoma (HGD/CRC) and invasive CRC. We first developed the LLM prompts using example reports from patients with inflammatory bowel disease (IBD). We then validated the approach in IBD and non-IBD by applying the fixed prompts to a separate corpus of 116 373 pathology reports generated in the VHA between 1999 and 2024. We compared model outputs to blinded manual chart review of 200–300 pathology reports for each patient cohort and diagnostic task, totalling 3816 reviewed reports, and calculated F1 scores as a balanced accuracy measure.

Results

In patients with IBD in MVP, we achieved F1-scores of 96.9% (95% CI 94.0% to 99.6%) for identifying dysplasia, 93.7% (88.2%–98.4%) for identifying HGD/CRC and 98% (96.3%–99.4%) for identifying CRC. In patients without IBD in MVP, we achieved F1-scores of 99.2% (98.2%–100%) for identifying any colorectal dysplasia, 96.5% (93.0%–99.2%) for identifying HGD/CRC and 95% (92.8%–97.2%) for identifying CRC using LLM Gemma-2.

Conclusion

LLMs provided excellent accuracy in extracting the diagnoses of interest from EHRs. Our validated methods generalised to unstructured pathology notes, even withstanding challenges of resource-limited computing environments. This may, therefore, be a promising approach for other clinical phenotypes given the minimal human-led development required.