Clinical trial matching is critical for identifying the most suitable trials for patients based on their unique medical profiles. Traditionally, this process relies on manual screen-ing by medical professionals, which is labour-intensive and inefficient, especially considering the vast number of available trials. Recent advancements explored automating this process, with large language models (LLMs) emerging as a popular solution. These models extract inclusion and exclusion criteria from unstructured patient data, encode the criteria from trials, and utilize cosine similarity to rank potential matches. However, a significant limitation of this approach lies in the interpretability of the cosine similarity scores—how and why the matches are produced often remain unclear. Our method introduces a method that combines a fine-tuned LLM for criteria generation with cosine similarity-based matching and is reinforced by symbolic reasoning to validate and enhance the interpretability of trial outcomes. Integrating neural network outputs with symbolic reasoning techniques represents a step forward in neuro-symbolic AI, aiming to provide accurate and explainable trial-matching results. The potential implications of this work are significant, offering a more reliable and transparent method for clinical trial matching that could improve patient outcomes and foster greater trust in AI-driven medical applications by validating and reinforcing the decisions of LLM and cosine similarity through additional layers of symbolic reasoning.
Salman et al. (Mon,) studied this question.