Skip to content
- Overall: 前几章貌似有干货，但微软的作者们似乎对这个研究的底层机制理解有点偏颇(特别在后面几张中)，更像是表层用户.
- Chap 1
- 作者认为GPT-4 interaction 拟人化很好，并且政治正确很好 – 在医患沟通应该大有可为
- 作者号称GPT-4能答对90%+ USMLE 问题
- Chap 2
- 对医疗人员，杂事耗时49%
- clinical paperwork – Msft DAX – 可以通过病人医生对话声音来自动写Notes
- clinical doc / summarization 时hallucination是危险的
- 一个AI写文章，另一个来批判
- AI assistant + AI verifier
- AI verifier 不仅仅检查AI结果，还检查医生结果
- 标准化输出(JSON) – e.g. HL7 FHIR 标准
- GPT-4 as universal translator – 给consumer 解释explanation of benefits (from medical bill)
- GPT-4 summarize medical paper
- Chap 3
- GPT-4 seems better at reviewing poetry / text than writing it
- GPT-4 limitations (numerical calculation error – e.g. correlation)
- Chap 4
- 作者的例子用differential diagnosis – 诊断一个十万分之一的先天性基因异常
- Chap 5 – Not Much
- Chap 6
- Math in treatment – 例如, 护士看滴液多少滴的时候需要解一个小学生应用题
- “Copilot” in drug administration
- One method to reduce chance of error in GPT is “show your work”
- Sudoku example
- Potentially remove GPT’s problem in logic / math to allow it to use tools – example: something called SAT solver
- Chap 7 – Paperwork
- Patient intake form
- GPT-4 identify billing codes.
- Prior Authorization – 避免医疗人员乱收费，需要先authorize 某种检查再实施
- 以医疗backoffice为GPT-4起点的可能性
- Chap 8
- Chap 9
- low risk vs. high risk app for AI in med
- Chap 10