Very interesting - I would like to see the link to the screen captures with the prompts used, results, etc. But the link you provide doesn't work - could you double check it? Thanks!
@Katie Rowley, I'm very sorry that I missed your question here! I'll need to check my notification settings.
I had no idea the link wasn't working, but you were right. Oddly, when I went in to edit it just now, it *was* working as coded ... but it did not work in the post's distribution email ... and when I saved it just now, it *does* work now (without me changing anything).
I blame Substack. Anyway, please let me know if it works now.
This differential is very good, including the ranking. It did though miss a critical Ddx that every dermatologist would have included, a drug rash. Drug would not be my choice based on the photos, but a drug hypersensitivity could have grave consequences if missed. Drug would certainly be more likely than CTCL which it listed. Such omissions point to the limitation of AI like ChatGPT to ask relevant history and to consider the base rate of different diagnoses. I'm sure that will improve with better model reasoning (test time compute, etc). I'm sorry he had to wait months for an appt and glad he got the help he needed.
Re accuracy, there's a reason why "Hugo's Law" is named after this Hugo - he says "I don't use AI to give me answers - I use it to help me think." And you can bet your bottom dollar that as long as the US health system can't produce a doctor for us to SEE, anyone with any empowerment at all (or compassion for the sufferer) is damn well GOING to use what we can get.
In other words, the fact that every dermatologist would have included a drug reaction is of no damned use if WE THE PATIENTS can't GET to someone. I know that you folks working in the dysfunctional health system have no more ability to change its structure than we sufferers do, but we're not sitting around when we have an option.
Meanwhile, as we grant the limitations of genAI, consider the limitations of human docs - see other posts on this blog, e.g. about Courtney Hoffman (17 docs over 3 years couldn't get the dx) or Sue Sheridan (ER doc missed her Bell's Palsy). See also the post about that, in November: Is #PatientsUseAI safe and okay? Compare it to what we HAVE.
This is beautiful to read!
I've always been a big fan of empowering patients, and this AI case study just confirmed it's totally possible!
Very interesting - I would like to see the link to the screen captures with the prompts used, results, etc. But the link you provide doesn't work - could you double check it? Thanks!
@Katie Rowley, I'm very sorry that I missed your question here! I'll need to check my notification settings.
I had no idea the link wasn't working, but you were right. Oddly, when I went in to edit it just now, it *was* working as coded ... but it did not work in the post's distribution email ... and when I saved it just now, it *does* work now (without me changing anything).
I blame Substack. Anyway, please let me know if it works now.
In case it doesn't, here's the direct link:
https://docs.google.com/document/u/2/d/e/2PACX-1vSf-qwq-frScQzSHjhB9Xr-7FHEQwkljn9h5WJJfz2inb1ct70UBp9_2tOL84FewV6iO3YCM-iPQnLo/pub
Do let me know. Sorry again for missing your appeal.
No worries and thank you for checking ! Yes it works now for and I’ll take a look.
This differential is very good, including the ranking. It did though miss a critical Ddx that every dermatologist would have included, a drug rash. Drug would not be my choice based on the photos, but a drug hypersensitivity could have grave consequences if missed. Drug would certainly be more likely than CTCL which it listed. Such omissions point to the limitation of AI like ChatGPT to ask relevant history and to consider the base rate of different diagnoses. I'm sure that will improve with better model reasoning (test time compute, etc). I'm sorry he had to wait months for an appt and glad he got the help he needed.
Hi, Dr. Jeff - long time!
Re accuracy, there's a reason why "Hugo's Law" is named after this Hugo - he says "I don't use AI to give me answers - I use it to help me think." And you can bet your bottom dollar that as long as the US health system can't produce a doctor for us to SEE, anyone with any empowerment at all (or compassion for the sufferer) is damn well GOING to use what we can get.
In other words, the fact that every dermatologist would have included a drug reaction is of no damned use if WE THE PATIENTS can't GET to someone. I know that you folks working in the dysfunctional health system have no more ability to change its structure than we sufferers do, but we're not sitting around when we have an option.
Meanwhile, as we grant the limitations of genAI, consider the limitations of human docs - see other posts on this blog, e.g. about Courtney Hoffman (17 docs over 3 years couldn't get the dx) or Sue Sheridan (ER doc missed her Bell's Palsy). See also the post about that, in November: Is #PatientsUseAI safe and okay? Compare it to what we HAVE.
https://patientsuseai.substack.com/p/is-patientsuseai-safe-and-okay-compare
Having said all that, it's great to be talking with you again!