Enhancing biomarker based oncology trial matching using large language models.

Area of Science:

• Oncology
• Bioinformatics
• Natural Language Processing

Background:

• Clinical trial enrollment relies on patient eligibility criteria often found in unstructured text.
• Genomic biomarkers are vital for precision medicine and targeted cancer therapies.
• Matching patients to clinical trials requires efficient extraction of eligibility information.

Purpose of the Study:

• To explore strategies for extracting genetic biomarkers from oncology clinical trial descriptions.
• To structure unstructured clinical trial data for improved patient matching.
• To evaluate the performance of large language models (LLMs) in this task.

Main Methods:

• Utilized open-source and closed-source large language models (LLMs) to process clinical trial study descriptions.
• Focused on extracting and structuring genomic biomarker information from eligibility criteria.
• Compared out-of-the-box LLM performance with fine-tuned models.

Main Results:

• Open-source LLMs demonstrated effectiveness in capturing complex logical expressions and structuring genomic biomarkers.
• Out-of-the-box open-source models outperformed closed-source models like GPT-4.
• Fine-tuning open-source models with additional data led to significant performance enhancements.

Conclusions:

• Open-source LLMs are a viable solution for structuring unstructured clinical trial data, specifically for genomic biomarkers.
• These models aid in identifying suitable patients for precision oncology trials.
• Further development and fine-tuning can optimize LLM performance for clinical trial data extraction.