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

Multidisciplinary teams (MDTs) are fundamental to cancer care but face increasing burdens. This pilot study evaluates a large language model (LLM) simulating a tumor board for a clinically complex cohort of non-small cell lung cancer (NSCLC) patients, using a full-context guideline injection methodology to ground its reasoning in authoritative standards.

Ten real-world NSCLC cases were presented to Google's Gemini 2.5 Pro using prompt engineering. The model was primed by providing the complete National Cancer Institute guidelines as in-context source data in a structured JSON file. AI-generated recommendations were scored against the institutional human MDT.

The LLM demonstrated high performance, achieving mean scores of 4.9/5.0 for content accuracy, 5.0/5.0 for internal consistency, and 4.4/5.0 for clinical applicability. Importantly, no safety concerns were identified in the AI's recommendations. However, the model did not generate any novel insights beyond those considered by the human MDT.

An LLM primed with comprehensive guidelines can accurately and safely replicate MDT recommendations for complex NSCLC cases. The combination of guideline injection and meticulous prompt engineering is a critical strategy for ensuring LLM reliability. This positions these models as powerful decision-support tools to augment, not replace, expert clinical workflow.