Prognostic data extraction harnessing a privacy-preserving large language model: a clinician-AI collaborative retrospective evaluation in head and neck oncology

Description

Privacy-preserving language models that pull prognostic details from cancer records are evaluated by Zou and colleagues. Head and neck cancer notes hold staging, virus status, and comorbidity information in free text, which makes manual extraction slow and costly, while privacy rules block sending records to outside services. The team tested locally run models on more than thirteen hundred reports from 882 patients, extracting thirty prognostic variables. Three radiation oncologists checked a fifty-case subset as the reference standard.

The best local model reached very high accuracy on that reference and processed each report in under a minute. Adding model-extracted virus status, smoking history, and comorbidity scores to a survival model improved predictions of disease-free survival and local failure. Zou and colleagues present on-site models as a practical route to faster tumour-board preparation and cleaner research registries without moving patient data off site.

Catalog record

Source: https://europepmc.org/articles/PMC13635170