Zero-Shot Clinical Trial Design Using Large Language Models: A Data-Driven Approach to Protocol Generation and Cohort Identification

Abstract

This project examines the innovative use of Large Language Models (LLMs) in designing zero-shot clinical trials in oncology. The method permits autocomplete of the entire trial protocol based on limited information-name of the drug, mechanism, indication, and phase- without speciality training-related tasks. The context supplied by historical Sanofi-financed clinical trial shows relative background to what the model yields. It involves the generation of protocols, the extraction of inclusion/exclusion criteria, the design of pseudo-SQL queries to find the cohort of patients, and the creation of regulatory submission documents. The research showed that LLMs are capable of independently and effectively generating quality trial documentation. This zero-shot approach has the potential to save time and cost of initiating a trial and increase inclusivity by appropriately choosing cohorts solely based on characteristics of real-world data.