中文医疗大模型综述: 进展、评估与挑战

Summary

This survey from Zhengzhou University examines large language models (LLMs) built for Chinese-language medicine. The authors argue that Chinese clinical practice differs from Western practice in treatment choices, prescribing habits, dosages, and especially Traditional Chinese Medicine (TCM). That makes dedicated Chinese medical data and models necessary. The review groups open-source systems released since 2023 into three areas: medical consultation, medical imaging, and mental health. It also mentions commercial products, such as Alibaba's Tongyi Renxin and Baidu's YiduCore. For each model, it compares the base model, fine-tuning recipe (SFT, IFT, RLHF/RLAIF), training data, strengths and weaknesses. The paper then reviews available evaluation benchmarks and their limits. Automatic GPT-4 judging without reference answers is unconvincing for medicine, and expert human review is costly and hard to scale. Finally, it outlines open problems, including hallucination, value alignment and data security. The survey is useful as a map of the Chinese medical-LLM landscape and its methodological gaps.

Key points

• Models catalogued: about 18 open models, including:
  • Consultation: DoctorGLM, BenTsao, BianQue, HuatuoGPT, MedicalGPT, ClinicalGPT, ChatMed, MING-MOE, IvyGPT, Zhongjing and DISC-MedLLM
  • TCM: ShenNong-TCM
  • Imaging: XrayGLM and Qilin-Med-VL
  • Psychological support: SoulChat and MindChat
• Base models and tuning methods:
  • Most systems build on ChatGLM, LLaMA, Baichuan, BLOOMZ, Qwen or InternLM.
  • SFT and IFT (often with LoRA or P-Tuning v2) dominate.
  • RLHF is used less often, probably because it is unstable.
• Imaging models:
  • XrayGLM is presented as the first Chinese chest X-ray model; it was fine-tuned on VisualGLM-6B.
  • Both XrayGLM and Qilin-Med-VL depend partly on English data translated with ChatGPT, which raises accuracy concerns.
• Mental-health models:
  • Emohaa runs on GLM-4 in 9B, 32B and 130B versions, with a context of 8k+ tokens and a crisis-detection component.
  • SoulChat fully fine-tuned ChatGLM-6B on more than 150k single-turn counselling samples.
• Benchmarks:
  • BioLAMA targets masked language models rather than generative ones.
  • MultiMedBench exists only in English.
  • Chinese-specific resources include CBLUE, PromptCBLUE and CMB.