Learning to match patients to clinical trials using large language models.

Area of Science:

• Information Retrieval
• Natural Language Processing
• Clinical Informatics

Background:

• Patient recruitment for clinical trials (CTs) is often inefficient.
• Traditional information retrieval methods struggle with the semantic complexity of patient-trial matching.
• Large Language Models (LLMs) offer advanced semantic processing capabilities.

Purpose of the Study:

• To investigate the efficacy of LLMs in enhancing patient-to-clinical trial matching.
• To compare LLM-based approaches against traditional methods within an information retrieval pipeline.
• To analyze the trade-offs between effectiveness and computational cost.

Main Methods:

• Utilized a multi-stage retrieval pipeline incorporating BM25, Transformer-based rankers, and LLM-based methods.
• Employed TREC Clinical Trials 2021-23 track collections for evaluation.
• Compared LLM integration in query formulation, filtering, ranking, and re-ranking.

Main Results:

• LLM-based systems, especially with fine-tuned re-ranking, outperformed traditional methods in nDCG and Precision.
• Fine-tuning LLMs improved their ability to identify eligible trials.
• LLM approaches achieved competitive performance with state-of-the-art systems in TREC challenges.
• Increased computational cost and reduced efficiency were observed with LLM usage.

Conclusions:

• LLMs show significant promise for automating and improving patient-to-clinical trial matching.
• Further research is needed to optimize the balance between computational cost and retrieval effectiveness.