Medical LLMs: Fine-Tuning vs. Retrieval-Augmented Generation

Large language models (LLMs) are trained on huge datasets, which allow them to answer questions from various domains. However, their expertise is confined to the data that they were trained on. In order to specialize LLMs in niche domains like healthcare, various training methods can be employed. Tw...

Full description

Saved in:
Bibliographic Details
Main Authors: Bhagyajit Pingua, Adyakanta Sahoo, Meenakshi Kandpal, Deepak Murmu, Jyotirmayee Rautaray, Rabindra Kumar Barik, Manob Jyoti Saikia
Format: Article
Language:English
Published: MDPI AG 2025-06-01
Series:Bioengineering
Subjects:
Online Access:https://www.mdpi.com/2306-5354/12/7/687
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849419202317778944
author Bhagyajit Pingua
Adyakanta Sahoo
Meenakshi Kandpal
Deepak Murmu
Jyotirmayee Rautaray
Rabindra Kumar Barik
Manob Jyoti Saikia
author_facet Bhagyajit Pingua
Adyakanta Sahoo
Meenakshi Kandpal
Deepak Murmu
Jyotirmayee Rautaray
Rabindra Kumar Barik
Manob Jyoti Saikia
author_sort Bhagyajit Pingua
collection DOAJ
description Large language models (LLMs) are trained on huge datasets, which allow them to answer questions from various domains. However, their expertise is confined to the data that they were trained on. In order to specialize LLMs in niche domains like healthcare, various training methods can be employed. Two of these commonly known approaches are retrieval-augmented Generation and model fine-tuning. Five models—Llama-3.1-8B, Gemma-2-9B, Mistral-7B-Instruct, Qwen2.5-7B, and Phi-3.5-Mini-Instruct—were fine-tuned on healthcare data. These models were trained using three distinct approaches: retrieval-augmented generation (RAG) alone, fine-tuning (FT) alone, and a combination of both (FT+RAG) on the MedQuAD dataset, which covers a wide range of medical topics including disease symptoms, treatments, medications, and more. Our findings revealed that RAG and FT+RAG consistently outperformed FT alone across most models, particularly LLAMA and PHI. LLAMA and PHI excelled across multiple metrics, with LLAMA showing superior overall performance and PHI demonstrating strong RAG/FT+RAG capabilities. QWEN lagged behind in most metrics, while GEMMA and MISTRAL showed mixed results.
format Article
id doaj-art-dea863ab480f42748158c8fca92a1178
institution Kabale University
issn 2306-5354
language English
publishDate 2025-06-01
publisher MDPI AG
record_format Article
series Bioengineering
spelling doaj-art-dea863ab480f42748158c8fca92a11782025-08-20T03:32:12ZengMDPI AGBioengineering2306-53542025-06-0112768710.3390/bioengineering12070687Medical LLMs: Fine-Tuning vs. Retrieval-Augmented GenerationBhagyajit Pingua0Adyakanta Sahoo1Meenakshi Kandpal2Deepak Murmu3Jyotirmayee Rautaray4Rabindra Kumar Barik5Manob Jyoti Saikia6Biomedical Sensors & Systems Lab, University of Memphis, Memphis, TN 38152, USABiomedical Sensors & Systems Lab, University of Memphis, Memphis, TN 38152, USADepartment of Computer Science and Engineering, Odisha University of Technology and Research, Bhubaneswar 751003, IndiaDepartment of Computer Science and Engineering, Odisha University of Technology and Research, Bhubaneswar 751003, IndiaDepartment of Computer Science and Engineering, Odisha University of Technology and Research, Bhubaneswar 751003, IndiaSchool of Computer Applications, KIIT Deemed to be University, Bhubaneswar 751003, IndiaBiomedical Sensors & Systems Lab, University of Memphis, Memphis, TN 38152, USALarge language models (LLMs) are trained on huge datasets, which allow them to answer questions from various domains. However, their expertise is confined to the data that they were trained on. In order to specialize LLMs in niche domains like healthcare, various training methods can be employed. Two of these commonly known approaches are retrieval-augmented Generation and model fine-tuning. Five models—Llama-3.1-8B, Gemma-2-9B, Mistral-7B-Instruct, Qwen2.5-7B, and Phi-3.5-Mini-Instruct—were fine-tuned on healthcare data. These models were trained using three distinct approaches: retrieval-augmented generation (RAG) alone, fine-tuning (FT) alone, and a combination of both (FT+RAG) on the MedQuAD dataset, which covers a wide range of medical topics including disease symptoms, treatments, medications, and more. Our findings revealed that RAG and FT+RAG consistently outperformed FT alone across most models, particularly LLAMA and PHI. LLAMA and PHI excelled across multiple metrics, with LLAMA showing superior overall performance and PHI demonstrating strong RAG/FT+RAG capabilities. QWEN lagged behind in most metrics, while GEMMA and MISTRAL showed mixed results.https://www.mdpi.com/2306-5354/12/7/687large language modelshealthcaremedicalretrieval-augmented generationfine-tuning
spellingShingle Bhagyajit Pingua
Adyakanta Sahoo
Meenakshi Kandpal
Deepak Murmu
Jyotirmayee Rautaray
Rabindra Kumar Barik
Manob Jyoti Saikia
Medical LLMs: Fine-Tuning vs. Retrieval-Augmented Generation
Bioengineering
large language models
healthcare
medical
retrieval-augmented generation
fine-tuning
title Medical LLMs: Fine-Tuning vs. Retrieval-Augmented Generation
title_full Medical LLMs: Fine-Tuning vs. Retrieval-Augmented Generation
title_fullStr Medical LLMs: Fine-Tuning vs. Retrieval-Augmented Generation
title_full_unstemmed Medical LLMs: Fine-Tuning vs. Retrieval-Augmented Generation
title_short Medical LLMs: Fine-Tuning vs. Retrieval-Augmented Generation
title_sort medical llms fine tuning vs retrieval augmented generation
topic large language models
healthcare
medical
retrieval-augmented generation
fine-tuning
url https://www.mdpi.com/2306-5354/12/7/687
work_keys_str_mv AT bhagyajitpingua medicalllmsfinetuningvsretrievalaugmentedgeneration
AT adyakantasahoo medicalllmsfinetuningvsretrievalaugmentedgeneration
AT meenakshikandpal medicalllmsfinetuningvsretrievalaugmentedgeneration
AT deepakmurmu medicalllmsfinetuningvsretrievalaugmentedgeneration
AT jyotirmayeerautaray medicalllmsfinetuningvsretrievalaugmentedgeneration
AT rabindrakumarbarik medicalllmsfinetuningvsretrievalaugmentedgeneration
AT manobjyotisaikia medicalllmsfinetuningvsretrievalaugmentedgeneration