A carregar...
Publicação
Retrieval-Augmented Generation for Biomedical Protocols: Optimising Knowledge Retrieval to Support Healthcare Technicians in Anatomical Pathology Workflows
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.72 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Healthcare technicians in Anatomical Pathology (AP) laboratories must be familiar
with extensive and frequently updated protocols and equipment manuals. Current
use of Large Language Models (LLMs) in healthcare mainly focuses on automating
administrative tasks and answering patient inquiries, leaving a gap in technical
support for laboratory staff. To address this gap, we built and optimised a
Retrieval-Augmented Generation (RAG) assistant that answers protocol-related queries
with precise, context-grounded responses. Introducing a novel corpus of AP-lab
documents and a synthetic question-answer evaluation dataset, we experimented with
different RAG configurations that varied in three core components: chunking method
(recursive vs. semantic chunking); retrieval strategy (na¨ıve similarity vs. reranking
vs. hybrid search); and embedding model (general-purpose vs. biomedical-specific).
Using a locally hosted Llama-3, each configuration was evaluated on RAGAS metrics.
The results showed that the optimal pipeline combined 512-token recursive chunks with
hybrid search retrieval and a biomedical embedding model, outperforming the baseline
across all metrics, confirming that domain-specific embeddings notably enhance retrieval
for technical terminology, and that hybrid search effectively balances conceptual and
keyword matching. This study demonstrates that an optimised RAG framework can serve
as a reliable knowledge assistant for laboratory technicians, introduces a novel corpus
of biomedical documents, and opens pathways for future work with larger biomedical
models, multimodal inputs, and more advanced RAG architectures.
Descrição
Dissertation presented as the partial requirement for obtaining a Master's degree in Data Science and Advanced Analytics, specialization in Data Science
Palavras-chave
Retrieval-Augmented Generation Large Language Models Natural Language Processing Chunking Retrieval Biomedical Anatomical Pathology SDG 3 - Good health and well-being