2023-11-01 NOV
Journal Club
by Peter Lee (Author), Carey Goldberg (Author), Isaac Kohane (Author)
Lots of videos and presentations --> https://www.youtube.com/results?search_query=The+AI+Revolution+in+Medicine%3A+GPT-4+and+Beyond
Notes
This article follows a practitioner-scientist form that could be traced back to Erik Topol Deep Medicine, and more recently GPT what it is by Stephen Wolfran
Chapters organized by author engagements, there is some leaking of topics throughput the book.
Physicians facing the choice between succumbing to admin overload or GPT assistance
Ontologies are not front-line tools. Checking in and out of operational standards is the way to go (recall examples with FHIR, both operation and research, and CPT codes)
Allucination approached as missing data.
Functionalizing LLMs and indexing prompts to embeddings appears to be the organizing engine the book anticipates.
Learning and Language, this has been known for a long time but we appear to have stumbled on something very important we don't quite understand. Without that investigation, it will be hard to claim it as part of a scientific approach.
Polemic, for the long run - I don't see a big difference between prompt engineers and excel experts that goes beyond the size of the paycheck ... Zack makes repeated allusions to this: medicine operates plenty of instruments that it doesn't completely understand, but that doesn't imply it doesn't value dissecting it with both exploratory and discriminant analysis approaches.
... this conversation starts in Chapt 4 but it is only in Chapt 6 that it takes place, with the Nobel prizewinner citation about humans being simple individually, but complex in the way they respond to context using associative recognition.
Sidenote from Chapt 6 - there is a persistent memory issue with GPT, aggravated by the lack of an online learning capability - a lot of (computer science) research is being directed to both practical fixes and systematic exploration.
Patient-Doctor dyad expanding to a triad with the addition of an AI assistant. The self-reflective and even apologetic style is a major strength.
GPT workflows for life-long healthcare - free your caterpillars! Recall the insistence on assistant's assistant - social gaming.
Recall the "make an App" response (Chapt 4-5) and the undefined liability
admin collapse: prior auth, CPT (>10k procedures) and ICD codes (>50k processes). Have a close look at https://www.hl7.org/fhir. It is proposed that this is where GPT will anchor cross-system operation. Yes, including data sharing. Recall penalties and liabilities of not using FHIR - see pdf Chapter 7 in the book.
From liability to equitable care delivery, researchers can use GPT to make it transparent.
This is just the beginning !!!
JSON-LD workflow engines
Hackathon
Update on functionalized language models
epiVerse caterpillar (Jonas)
https://epiverse.github.io/#url=https://episphere.github.io/caterpillar/caterpillar.js&ui=uiAnyone else :-)
...
...