GutGPT: A multidimensional knowledge-enhanced large language model for gastrointestinal medicine

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing for Healthcare
• Gastroenterology Informatics

Background:

• Gastrointestinal (GI) diseases are prevalent, chronic conditions necessitating personalized, long-term patient management.
• Traditional healthcare systems face significant burdens managing these complex GI diseases.
• Existing large language models (LLMs) lack specialized GI knowledge and face challenges like slow convergence and overfitting.

Purpose of the Study:

• To develop a domain-specific LLM for improved GI disease management.
• To enhance patient care through personalized and empathetic AI-driven guidance.
• To overcome limitations of general LLMs in handling complex GI conditions.

Main Methods:

• Construction of a 191,615-entry GI disease Question Answering (QA) dataset from diverse sources.
• Fine-tuning the Baichuan-13B-Chat LLM using Low-Rank Adaptation (LoRA) and self-attention parameter sharing to create GutGPT.
• Performance evaluation using expert assessments and public benchmark datasets (CMB, CMExam).

Main Results:

• GutGPT demonstrated superior performance, outperforming 16 existing methods and achieving state-of-the-art results.
• Expert evaluations showed a 9.59% improvement in diagnostic accuracy for GutGPT compared to baselines.
• On public QA datasets, GutGPT achieved an average accuracy improvement of 22.47%.

Conclusions:

• GutGPT exhibits high accuracy and strong empathy in managing GI disease patients.
• The model serves as a valuable auxiliary tool for physicians and patients in GI disease management.
• GutGPT represents a significant advancement in AI-assisted healthcare for gastroenterology.