Enhancing Clinical Trial Selection for Cancer Patients Using Large Language Models

Identifying appropriate clinical trials for cancer patients with specific gene mutations remains a significant challenge, largely due to limitations in current search tools like ClinicalTrials.gov, which at times return irrelevant or misleading results. This diagnostic accuracy study investigates the efficacy of 2 large language models (LLMs), GPT-4.0 and Gemini 2.0, in evaluating the eligibility of patients with specific cancer-related gene mutations for clinical trials.

The study prompts GPT 4.0 and Gemini 2.0 with trial details from ClinicalTrials.gov and a particular cancer mutation. We then assess model performance against physician-curated benchmarks across 6 gene mutations (ALK, BRAF, EGFR, ERBB2, KIT, and KRAS).

The results demonstrate good F1-scores for both LLMs-averaging 64% for GPT-4.0 and 70% for Gemini 2.0-highlighting their potential to streamline clinical trial matching. Furthermore, decision trees provided interpretability by identifying key textual indicators that LLMs use.

This work demonstrates the feasibility of using proprietary LLMs such as GPT 4.0 and Gemini 2.0 "off the shelf" with both limited LLM fine-tuning and limited patient information to evaluate clinical trial eligibility.