Medical foundation large language models for comprehensive text analysis and beyond.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Natural Language Processing

Background:

• Large language models (LLMs) show promise in medicine but lack specialized knowledge.
• Existing models struggle with complex medical data and tasks.

Purpose of the Study:

• Introduce Me-LLaMA, a family of open-source medical LLMs.
• Enhance LLMs with domain-specific medical knowledge and instruction-following abilities.

Main Methods:

• Continual pretraining and instruction tuning of LLaMA2 models.
• Utilized diverse biomedical literature and clinical notes for training.
• Evaluated on six text analysis tasks and complex case diagnosis.

Main Results:

• Me-LLaMA outperforms other open-source medical LLMs.
• Achieved performance comparable to or exceeding ChatGPT and GPT-4 on various benchmarks.
• Demonstrated clinical utility in complex case diagnosis.

Conclusions:

• Combining domain-specific pretraining with instruction tuning is crucial for medical LLMs.
• Me-LLaMA offers a powerful, open-source solution for medical AI applications.