Leveraging LLMs to extract smoking history from clinical notes (www.nature.com)

Researchers have developed a framework that leverages large language models (LLMs) to extract and enhance smoking history documentation from clinical notes in electronic health records (EHRs). Accurate smoking data is vital for assessing patient risk and monitoring health outcomes, yet traditional documentation often suffers from inconsistencies and omissions. The study compared various generative LLMs, including Gemini-1.5-Flash, PaLM-2-Text-Bison, and GPT-4, against established BERT-based models using 1,683 manually annotated clinical notes. The generative models surpassed BERT alternatives with accuracy exceeding 96% for smoking-related variables and demonstrated robust generalizability across different healthcare systems. Deploying the Gemini model, the researchers extracted data from 79,408 notes related to 4,792 lung cancer patients, highlighting that risk models incorporating smoking history outperformed existing cancer guidelines in identifying second malignancies. This work not only showcases the potential of advanced LLMs in correcting historical data inconsistencies but also points towards significant implications for individualized cancer care and health monitoring, ultimately offering a pathway to improve patient outcomes through refined and accurate smoking history documentation.