New research from @katherlab, led by @JClusmann : "Prompt Injection Attacks on Large Language Models in Oncology" arxiv.org/abs/2407.18981 We show that vision-language models in oncology can be attacked easily, creating malicious output which could harm patients. @tudresden_de @NCT_UCC_DD @NCT_HD @EKFZdigital @DKFZ

New preprint from @katherlab: large language models for end-to-end clincial trial matching. Our system parses study protocols and patient records, and matches them. We evaluate thoroughly against human experts and find >90% correspondence. Led by @Dykex6 🤩 arxiv.org/abs/2407.13463 @NCT_UCC_DD @NCT_HD @DKFZ @DataScienceDKFZ @uniklinik_hd @myESMO CC @tompowles1 @NEJM_AI @Agrawal_Mri @alexia_zz @PepPamies

Finally out in @NEJM_AI 🤖, led by @Dykex6 from @katherlab: "GPT-4 for Information Retrieval and Comparison of Medical Oncology Guidelines". We've developed a RAG pipeline using the GPT-4 API to answer medical oncology questions based on guidelines. Previous studies have claimed that "large language models (LLMs) are not suited for oncology" -- but our data debunk these claims! Using RAG, LLMs are very good at medical decision-making in oncology -- as we show in a quantitative evaluation. There is still room for improvement, but LLMs+RAG are the way to go ✨ Paper link: ai.nejm.org/doi/full/10.1056… Free-access link: ai.nejm.org/stoken/default+d… @myESMO @ASCO @AACR @uniklinik_hd @Medizin_TUD @NCT_UCC_DD @NCT_HD @DataScienceDKFZ