12 Comments
User's avatar
NanJ's avatar
4dEdited

Lay person here - I read a fair amount about AI and this is one of the best articles I’ve read because it is grounded in reality. I saw the stories about the pancreatic cancer imaging and was very impressed and excited for the possibilities. Dr Katz brings us back to reality. For the sake of patient care I hope we don’t get ahead of ourselves with AI implementation and realize in a few years we way overbought the hype.

Epaminondas's avatar

The false positive issue for any screening test, whether AI driven or not, is a huge problem. It's frequently why mass screening programs typically fail cost benefit analyses in real world environments.

Greg Katz, MD's avatar

False positives are a real challenge

Greg Katz, MD's avatar

Thanks for the kind words

William Wilson's avatar

As a physician with over 40 years of clinical experience, I'm not worried about AI taking my job. When we evaluate patients, we listen to their tone of voice and monitor their facial expressions and body language to assess what's going on. These subtle signs are often more important than the patient's history. AI simply can't do any of this. AI is helpful for charting and other routine activities, but leave the diagnosis and treatment to us.

Greg Katz, MD's avatar

You never know what will happen!

Kel Feind's avatar

I’m taking a vacation from my retirement and working as a cardiologist (my specialty) in New Zealand for a year. I’m finding Gemini to be extremely useful in interpreting Echocardiograms here. They have lots of rheumatic heart disease in NZ which is rare in the US. AI is incredibly helpful to me in describing these images. I can take a 2 sec video with my phone, upload it, and have an answer a minute later. Priceless.

Timothy Gutwald's avatar

I attend CES this year and heard about how wearables can predict seniors that will develop a UTI days before they get the UTI. The problems were similar to those you highlight about the AI finding pancreatic cancer. Are we going to give these people antibiotics while they’re otherwise asymptomatic? Do we still wait until they get sick? If so, the AI is cool but what does it make better? There were also issues with false positives. And I kept asking myself how payors are going to react? If they won’t pay for early intervention, then there won’t be any early intervention.

Greg Katz, MD's avatar

It’s hard to know what to do with the algorithms output

helmingstay's avatar

As someone who recently watch and (mostly) enjoyed The Pitt, I found this discussion of charting very informative.

While I'm not an engineer, software development has been a big part of my day job for 10+ years. I was pretty skeptical about agent-based programming until recently -- and I've been really surprised with how capable models have become. Using them, it quickly becomes clear that they're good in large part because they can test solutions in real-time and get high-quality feedback. And there just aren't that many other fields where there's such a bright line between "working" and not.

I do wonder if the LLMs might be useful for chart reviews where there are clearly defined problems -- things like spotting problem med combinations? Alternately, synthetic experiments that conduct retrospective chart reviews to hypothesize risk factors like problematic drug combos. But to be helpful there needs to be clear incentives and verifiable outcomes. For a big group like KP (or UK's NHS), one can imagine having sufficient scale to get a good signal *and* the motivation to try novel cost-reduction methods. It's hard to imagine either NIH-funded research or a biomed company moving the needle here, though.

Carl Gessner, MD's avatar

Though I understand and accept your criticism of AI scribe applications, my experience moving from Dragon dictation only to Abridge AI scribe plus Dragon has been positive. It has reduced my documentation time significantly and I have been able to adjust my interview style to get a better note. Absolutely need to proofread and edit, and using bulleted format for HPI provides a more cogent note. It is especially helpful in creating patient instructions and so is Open Evidence for more specific questions. Like everything else in this technology driven era it does require new effort and time to develop., which is a burden. We are getting ready for another Epic "upgrade" and I will have to adapt to those changes.

Appreciate your thoughtfulness.

Nick Gilpin's avatar

need more doctors meeting the moment and saying no to AI… communities all over the country are pushing back on AI and data center proliferation and meanwhile doctors are sitting back saying wishy-washy things like “AI can be a tool for good if used correctly…”

pick a reason: data privacy concerns, environmental destruction, rising energy costs, “hallucinations”