Evaluating Medium Scale, Open-Source Large Language Models: Towards Decision Support in a Precision Oncology Care...

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems

Background:

• Precision oncology requires up-to-date knowledge for complex patient cases.
• Preparing cases for molecular tumor boards (MTBs) is labor-intensive.
• Large language models (LLMs) offer potential to streamline information retrieval for MTBs.

Purpose of the Study:

• To evaluate the utility of medium-scale LLMs for answering clinical questions in MTB preparation.
• To assess the performance of on-premise LLMs using consumer hardware for sensitive data handling.

Main Methods:

• Three LLMs were selected based on benchmarks and reasoning capabilities.
• Domain experts provided exemplary MTB-related questions.
• Experts evaluated LLM-generated responses for quality and correctness.

Main Results:

• Overall LLM performance was modest, with significant issues identified.
• A high percentage of responses contained outdated, incomplete, or factually erroneous information.
• Evaluator discordance and varying confidence levels were observed.

Conclusions:

• Medium-scale LLMs are currently unreliable for precision oncology applications.
• Outdated information and confident misinformation highlight a gap between benchmark and real-world performance.
• Future research should explore advanced techniques like RAG and web search, prioritizing patient safety.