Simulating a virtual tumor board with large language models: a pilot study in NSCLC patients receiving immunotherapy

Area of Science:

• Artificial Intelligence in Medicine
• Oncology Decision Support

Background:

• Multidisciplinary teams (MDTs) are crucial for cancer care but face significant burdens.
• Evaluating AI tools to support these teams is essential.

Purpose of the Study:

• To assess a large language model (LLM) simulating a tumor board for non-small cell lung cancer (NSCLC) cases.
• To evaluate the LLM's performance using guideline injection for clinical decision support.

Main Methods:

• Ten complex NSCLC cases were analyzed using Google's Gemini 2.5 Pro.
• The LLM was provided with National Cancer Institute guidelines via prompt engineering.
• AI recommendations were compared against human MDT decisions.

Main Results:

• The LLM achieved high scores for accuracy (4.9/5.0), consistency (5.0/5.0), and applicability (4.4/5.0).
• No safety concerns were identified in the AI-generated recommendations.
• The LLM did not produce novel insights beyond the human MDT's considerations.

Conclusions:

• LLMs primed with guidelines can reliably replicate MDT recommendations for complex NSCLC.
• Guideline injection and prompt engineering are key for LLM reliability in clinical settings.
• LLMs can serve as valuable decision-support tools, augmenting rather than replacing expert judgment.