Recrutamento Automatizado para Ensaios Clínicos com GenAI e graphRAG

Abstract

Patient recruitment for clinical trials remains one of the main challenges in medical research, due to the reliance on manual screening processes of clinical records, which are frequently inefficient and error-prone. To address this limitation, this dissertation proposes an automated solution for matching patients against clinical trial eligibility criteria, based on Generative Artificial Intelligence (GenAI) and GraphRAG. The architecture was developed iteratively, exploring and comparing different techniques across multiple iterations, from entity extraction to ontological linking with biomedical knowledge bases. In doing so, the system progressively overcame lexical limitations through the use of knowledge graphs to map complex relationships between entities. The matching engine adopts a hybrid approach composed of two complementary mechanisms: a structural engine based on Cypher queries over Neo4j and a semantic engine supported by embeddings and cosine similarity. Following an initial eligibility screening, an LLM analyses the full context returned by the search engine, producing the final decision, a natural language explanation, and actionable insights for clinicians on what would be required for the patient to become eligible for the trial, positioning the system as a transparent and interpretable clinical decision support tool. As such, this approach is expected to outperform conventional manual screening methods and purely generative models, establishing itself as a robust clinical decision support tool capable of significantly accelerating recruitment in medical research trials.