Good article and glad to have someone with clinical experience, research experience, and a passion for tech such as yourself at the table.
I recently learned of Perplexity AI, and I’ve tried it for a few clinical questions. It is still not as good as humanly doing the hard work of PubMed and reference book synthesis, but so much faster of course… and it cites references. They are real clickable references! Often from journal sources, but also from more watered down educational sites like WebMD.
Let us know if you have any experience with this one someday. It’s hard to keep up with the AI Cambrian explosion.
Extraordinary week, and extraordinary report, Eric. I don't know of any other observer/reporter in the medical AI space who's anywhere near as comprehensive.
When I spoke in December at the IHI Leadership Forum I said that medicine is having a "monolith moment," referring to "2001: A Space Odyssey." NOTHING is going to be the same now that we've touched it, and we all better fasten our seatbelts ... and LEARN about the new world we are already in. This column supports that analogy. WOW.
You should see what our friend Hugo is up to for his heart health, now that he has this incredible range of tools to work with. He recently spun off a custom GPT-4 and fed it 110MB of PDFs from his various EMRs, and now has extended conversations with it to explore different questions. "At last: autonomy!!" he laughed to me.
Also, to GPT-4, while researching new BP meds due to age, prepping for a conversation with his cardiologist: "Please compare the inotropic and chronotropic effects of verapamil and diltiazem on the human heart."
I was on a panel Wednesday at the AMIA Informatics Summit with some co-authors of a paper that's in press, addressing how to do governance in a world that's changing this fast. In drafting it, colleagues Leon Rozenblit of Yale and Amy Price (Dartmouth) spoke of the Tiger Beetle - a desert scavenger that moves so fast that it outruns its own brain (ability to process photons) so it must burst, stop, look around, and repeat. They proposed this is how AI governance will need to work!
The SyntheMol AI sounds so promising! I'm wondering whether anyone is looking to make better vaccines with a similar program. It would be great if some AI can come up with a more stable version of the mRNA vaccines that will make them available to places in the world that can't use the current versions. Or find antigenic regions to make a pan-coronavirus vaccine. Or maybe even design small vaccine molecules that are inexpensive to make and store, and to administer by the best route (nasal, oral, injection).
Historically, radiology has often been open to adopting new technologies in advance of other specialties. For example, I started using voice recognition software to dictate reports in the 90s. With AI, I think the adoption rate has been high because most of what we radiologists look at is in digital form to begin with and therefore amenable to AI assessment.
I'm most familiar with AI applications in ultrasound, specifically evaluation of thyroid nodules for risk stratification and determining whether biopsy is required. In many studies, AI has outperformed less experienced physicians in this regard, but I'm concerned that if they consistently depend on software, they'll lose or perhaps never acquire the expertise to determine when it's off base.
Good article and glad to have someone with clinical experience, research experience, and a passion for tech such as yourself at the table.
I recently learned of Perplexity AI, and I’ve tried it for a few clinical questions. It is still not as good as humanly doing the hard work of PubMed and reference book synthesis, but so much faster of course… and it cites references. They are real clickable references! Often from journal sources, but also from more watered down educational sites like WebMD.
Let us know if you have any experience with this one someday. It’s hard to keep up with the AI Cambrian explosion.
It really is an explosion. I haven't tried Perplexity but will now. Thanks!
Extraordinary week, and extraordinary report, Eric. I don't know of any other observer/reporter in the medical AI space who's anywhere near as comprehensive.
When I spoke in December at the IHI Leadership Forum I said that medicine is having a "monolith moment," referring to "2001: A Space Odyssey." NOTHING is going to be the same now that we've touched it, and we all better fasten our seatbelts ... and LEARN about the new world we are already in. This column supports that analogy. WOW.
Thanks Dave!
You should see what our friend Hugo is up to for his heart health, now that he has this incredible range of tools to work with. He recently spun off a custom GPT-4 and fed it 110MB of PDFs from his various EMRs, and now has extended conversations with it to explore different questions. "At last: autonomy!!" he laughed to me.
Also, to GPT-4, while researching new BP meds due to age, prepping for a conversation with his cardiologist: "Please compare the inotropic and chronotropic effects of verapamil and diltiazem on the human heart."
Love it!
I was on a panel Wednesday at the AMIA Informatics Summit with some co-authors of a paper that's in press, addressing how to do governance in a world that's changing this fast. In drafting it, colleagues Leon Rozenblit of Yale and Amy Price (Dartmouth) spoke of the Tiger Beetle - a desert scavenger that moves so fast that it outruns its own brain (ability to process photons) so it must burst, stop, look around, and repeat. They proposed this is how AI governance will need to work!
Love that you’ve now created a YouTube channel, too! Thanks so much for everything you do.
The SyntheMol AI sounds so promising! I'm wondering whether anyone is looking to make better vaccines with a similar program. It would be great if some AI can come up with a more stable version of the mRNA vaccines that will make them available to places in the world that can't use the current versions. Or find antigenic regions to make a pan-coronavirus vaccine. Or maybe even design small vaccine molecules that are inexpensive to make and store, and to administer by the best route (nasal, oral, injection).
Yes. All if those are being pursued with AI support!
Historically, radiology has often been open to adopting new technologies in advance of other specialties. For example, I started using voice recognition software to dictate reports in the 90s. With AI, I think the adoption rate has been high because most of what we radiologists look at is in digital form to begin with and therefore amenable to AI assessment.
I'm most familiar with AI applications in ultrasound, specifically evaluation of thyroid nodules for risk stratification and determining whether biopsy is required. In many studies, AI has outperformed less experienced physicians in this regard, but I'm concerned that if they consistently depend on software, they'll lose or perhaps never acquire the expertise to determine when it's off base.