Core Concepts
Large language models can accurately and efficiently match patients to clinical trial eligibility criteria without any fine-tuning or labeled data.
Abstract
This paper explores the use of large language models (LLMs) for zero-shot clinical trial patient matching. The key highlights are:
Zero-shot evaluation: The authors evaluate the zero-shot performance of various LLMs, including GPT-3.5 and GPT-4, on the 2018 n2c2 cohort selection benchmark. Despite not being fine-tuned, GPT-4 achieves state-of-the-art results, outperforming the previous best model by a margin of +6 Macro-F1 and +2 Micro-F1 points.
Prompt engineering and retrieval pipeline: The authors investigate different prompting strategies and find that the "All Criteria, Individual Notes" (ACIN) approach provides the best balance of performance and efficiency. They also design a two-stage retrieval pipeline that can reduce the number of tokens processed by the LLM by up to a third while retaining high performance.
Interpretability: The authors have clinicians evaluate the natural language justifications generated by GPT-4 for its eligibility decisions. They find that GPT-4 can output coherent explanations for 97% of its correct decisions and 75% of its incorrect ones, enabling human oversight and collaboration.
The results establish the feasibility of using LLMs to accelerate clinical trial operations by automating patient matching, which is a key bottleneck in advancing new drugs to market. The authors discuss the potential for LLM-based systems to be deployed as "pre-screeners" that flag eligible patients, thereby allowing clinical research coordinators to focus on the most promising candidates.
Stats
One third of clinical trials fail because they cannot enroll enough patients.
Recruitment costs an average of 32% of a trial's budget.
94% of patients are never informed by their doctors about trials for which they might qualify.
Identifying eligible patients can take up to 1 hour per patient.
Quotes
"Matching patients to clinical trials is a key unsolved challenge in bringing new drugs to market."
"94% of patients are never informed by their doctors about trials for which they might qualify."
"Identifying patients who are eligible for a trial is often highly manual and time-consuming."