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

Abstract

Recent advancements in large language models (LLMs) show significant potential in medical applications but are hindered by limited specialized medical knowledge. We present Me-LLaMA, a family of open-source medical LLMs integrating extensive domain-specific knowledge with robust instruction-following capabilities. Me-LLaMA is developed through continual pretraining and instruction tuning of LLaMA2 models using diverse biomedical and clinical data sources (e.g., biomedical literature and clinical notes). We evaluated Me-LLaMA on six text analysis tasks using 12 benchmarks (e.g., PubMedQA and MIMIC-CXR) and assessed its clinical utility in complex case diagnosis through automatic and human evaluations. Me-LLaMA outperforms existing open medical LLMs in zero-shot and supervised settings and surpasses ChatGPT and GPT-4 after task-specific instruction tuning for most text analysis tasks. Its performance is also comparable to ChatGPT and GPT-4 for diagnosing complex clinical cases. Our findings highlight the importance of combining domain-specific continual pretraining with instruction tuning to enhance performance in medical LLMs.