构建可信医学大模型的路线图.pdf

1、R Ro oa ad dmma ap p t to o t tr ru us st tw wo or rt th hy y mme ed di ic ca al l L LL LMMs s2 20 02 24 4年年6 6月月1 16 6日日B Be en ny yo ou u WWa an ng g1YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024落地场景分诊预问诊问诊（诊断、开处方）体检报告解读随访心理疾病预警和干预医学常识咨询医学信息检索保健咨询慢病管理医疗

2、影像医疗资源优化医疗文本处理YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP20243Challenge:医疗大模型的评估困难Guiming Hardy Chen,Shunian Chen,Ruifei Zhang,Junying Chen,Xiangbo Wu,Zhiyi Zhang,Zhihong Chen,Jianquan Li,Xiang Wan,Benyou Wang.ALLaVA:Harnessing GPT4V-synthesized Data for A

3、Lite Vision-Language Model.https:/arxiv.org/abs/2402.11684华佗GPTGPT-4QwenYSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024Chat like a doctorAccuracy of professional knowledge Medical diagnosis and clinical decision makingWe are picking the low-hanging lycheeHu

4、atuoGPTHuatuoGPT-IIPass some examsChallenge:鲁棒的医疗诊断和医疗决策任重道远YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP20245Select a well-defined problemDesterata：-不与医生竞争，服务医生-评估客观简单可复现-更大的受众，可以服务更广泛的社会-achievable in a few yearsYSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNL

5、P2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP20246Standpoint:LLM for Doctors not as DoctorsWenya Xie,et.al.Benyou Wang。LLMs as for Doctors:Leveraging Medical LLMs to Assist Doctors,Not Replace Them。in ProgressYSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP20247Prob

6、lem Definition Definition：-Given a user with implicit/emplicit information-Goal：predict a label（分类任务：科室/开药/诊断）-Key-The system could actively ask followup questions to reduce uncertainty-The user could lose patience if with too many interactions科室分类是一个最容易来验证的场景！作为一个GMAI的sandboxYSSNLP2024YSSNLP2024YSS