Improving Patient-Clinical Trial Matching Using Convolution Neural Networks

Abstract

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.

Ähnliche Arbeiten