This case study reviews the objectives, hurdles, and outcomes of a project designed to use Artificial Intelligence (AI) to assist the clinical trial review process for pancreatic cancer patients by integrating the clinicaltrials.gov database, commercial Large Language Models (LLMs), and Structured Query Language (SQL) databases. The case study brought to light a significant number of challenges and limitations with both the clinicaltrials.gov database and the LLMs, but by disassembling the screening process into small, verifiable components, optimizing the required LLM prompts with hundreds of iterations to produce effective Python scripts, and then using MySQL to both scale and reassemble the component decisions, a fully functioning and secure production environment was constructed in the cloud and is available for use. 

This approach differed substantially from traditional AI approaches in that LLMs were used solely in the first step of the process to aid the researcher in writing many separate Python scripts, each of which focused on one trial dimension and whose efficacy could be effectively vetted by a human researcher. Once the Python scripts were refined over many iterations, the cord to AI was cut, with only the scripts and MySQL queries taking over the entire process. This approach significantly enhanced outcomes by quarantining patient data from AI, ensuring deterministic results (the same question yields the same answer every time, often not true in AI), minimizing resource use, and maximizing speed. Our approach used a combination of filtering and a scoring rubric not only to reduce the size of the potential trial set, but also “promote” the trials that were best suited for each specific patient.