Large language models for precision oncology: Clinical decision support through expert-guided learning.

• DOI: 10.1200/jco.2024.42.16_suppl.e13609
• Publication Date: 2024-05-29T17:43:22Z
• Authors: Jacqueline Lammert; Tobias F. Dreyer; Alisa M. Lörsch; Johannes Jung; Sebastian Lange; Nicole Pfarr; Anna Durner; Marion B. Kiechle; Ulrich A. Schatz; Sonja Mathes; Kristina Schwamborn; Christof Winter; Carolin Mogler; Anna Lena Illert; Maximilian Tschoc...
• Keywords: Personalized Medicine

e13609 Background: Precision oncology revolutionized cancer treatment by identifying molecular biomarkers to guide personalized care. The ever-growing body of medical literature presents a challenge for oncologists researching targeted therapies. While recent studies investigated large language models (LLMs) streamline this process, LLM reliance on general rather than knowledge limits clinical relevance and trustworthiness. To address these limitations, we developed retrieval augmented generation (RAG) system that integrates PubMed studies, trial databases oncological guidelines with LLMs support recommendations. Molecular Tumor Board (MTB) at the Center Personalized Medicine (ZPM TUM ) guided evaluated options proposed assess their applicability decision support. Methods: We used 10 publicly accessible fictional patient cases 7 tumor types 59 distinct alterations. Our MEREDITH (Medical Evidence Retrieval Data Integration Tailored Healthcare) consists Google's Gemini Pro, enhanced RAG Chain-of-Thought (CoT) prompting. establish benchmark, experts ZPM manually annotated cases. Informed MTB expert feedback, iteratively improved our from draft relying PubMed-indexed data an system, which replicated annotation processes incorporating guidelines, drug availability (ClinicalTrials.gov, QuickQueck.de). assessed credibility LLM-generated Patient-level (likely) pathogenic alterations recommended were summarized using median interquartile range (IQR). Semantic similarity between clinician responses was cosine text vector embeddings; paired t-test significance. Results: per 2.5 (IQR: 2-3). identified 2 1-3), while 4 3-5). multiple relevant suggestions, including therapies based preclinical interactions, further assessment MTB. prioritized most suitable option. mean semantic textual increased significantly 0.69 in 0.76 (p <0.001). Thus, feedback model's ability align its thought processes. Conclusions: Leveraging instruct holds promise as novel tool precision oncology.