AI in Healthcare By the Numbers

Welcome back to Generative AI 101, the podcast where we read the study instead of the press release Which feels especially urgent when the press release involves your body. All right. I'm Emily Laird your second opinion on the second opinion Today, we're gonna explore the reality behind the AI hospital So not the one in the concept video the one being built right now inside the workflows This is gonna be the story of the note that writes itself the scan that gets a co-reader And why the biggest shift in American medicine is happening in the paperwork Buckle up because we're about to find out that AI did not come for the diagnosis It came for the documentation. Welcome to Generative AI 101 All right, let's get into it some internet has spent this entire year insisting that 2026 is the new 2016 But I love that for me. Let's test that in medicine where the 2016 question was extremely specific And I'd say extremely dramatic. Will the machine beat the radiologist 10 years later We have data and the data says we asked the wrong question with tremendous confidence the American Medical Association Surveyed nearly 1700 physicians this past year 81 percent reported using AI professionally That is more than double the number from 2023 as I told you yesterday the adoption here is real But look at what they're doing with it, okay? So summarizing research writing discharge instruction documenting visits summarizing patient charts drafting patient messages The single most common use coming in at 39 percent isn't actually summarizing Medical literature and standards of care assistive diagnosis 17 percent so much lower 17 like right the thing we built a decade of anxiety around is actually the smallest slice on this board Everything else is the work around the medicine which brings me to my one little jargon pivot piece of vocabulary worth unpacking today And that's ambient clinical documentation. So you have been in a meeting room or somebody was assigned to take minutes You talked they typed and at the end you got a document you didn't write but had to approve and if they got a detail wrong Congratulations, it is your problem now because your name is on the bottom This is ambient documentation. So doctor and patient talk like two people in a room a system listens Turns the conversation into structured clinical information and then it's going to hand back a drafted note Sometimes with coding sometimes with the patients Instructions already prepared nobody dictated it nobody typed through the visit and nobody is off the hook because a human still signs I have done work with integration and piloting this in healthcare spaces And it's always so fun to me when I go into do my checkups or whatever it is and they're like hey Do you care if I use this and I'm like yeah, ambient recording device for patient notes and they're like how do you know? Oh, I'm like Oh, I'm not gonna tell you because I sign on NDAs and in hospitals and you do a it's a lot of privacy But I always think it's cute when they think it's cute that I know what it is So does it work? So here the evidence is good and more interesting than they think the vendors would like 2025 study followed 263 clinicians across six health systems So after 30 days with an ambient scribe reported burnout dropped by almost 52 percent to 38.8 percent 13 points in burnout research. That's that's not a rounding error. That's a solid lift I love it a randomized trial then put 238 outpatient physicians across 14 specialties on two different ambient systems one cut note writing by 9.5 percent the other one did not Significantly cut it at all same category of product, but a different outcome Physicians also flag occasional clinically significant inaccuracies Which is the entire reason a human still signs and then back in my state a randomized trial through the University of Wisconsin Hospitals and clinics track 66 practitioners and more than 71,000 notes so Documentation time in this study fell by about 0.36 hours a day work exhaustion dropped note quality held billing accuracy improved and From the folks that I've worked with in pilots and integration work like this The one thing they say that it comes challenging is the AIs and I don't know why this is weird They're really good with a lot of the medical terms and stuff. They stay very accurate It's the names that they're not always good at so sometimes physicians have unique names And then they if it's the physicians assistant or nursing assistant that will spend time reviewing the notes for accuracy based on Sitting in on the the patient physician appointment They're like why can't it just get the doctor's name right? I just think that's so cute I don't know, but that's I mean that's been one of the main things. I was brought to my attention So here's the part I'm gonna want you to kind of sit with today those clinicians did not turn around and see meaning fully more patients The times saved did not get converted into throughput. So the honest read is this ambine AI is not a productivity Technology it's an attention technology. It gives the clinician back the middle of the appointment and the end of the evening So the entire premise is that the undivided attention is now a scarce enough resource that you have to engineer the room to get it Medicine just spent a decade doing the opposite by accident wedging a keyboard between two people The scribe is not the miracle, but getting the eye contact back sort of is now the part of this episode I actually want you to leave with in another part I'm incredibly familiar with in the space of AI integration to radiology or as I think about it the original Battlefield roughly 84% of FDA authorized AI medical devices take images as their primary input and Radiology dominates that field. So with AI plus doctor is Ever going to be a clean win. It should be a clean win in this space in a test ran by nature medicine 140 radiologist 15 chest x-ray diagnostic tasks and the finding was not AI helps The finding was that it depends enormously on which radiologist you happen to get when the AI was right It could lift performance when the AI was wrong. It pulled physicians towards the wrong answer See they start to question themselves and here's the kicker years of experience did not predict who resisted the bad suggestion Neither did subsequent reality Neither did familiarity with AI nobody could tell you in advance who the machine would help and who it would mislead and that Should probably permanently retire the phrase doctor plus super computer Sometimes you get a useful second opinion Sometimes you get a confident wrong one delivered by something that never sounds unsure Which is sort of the problem we've talked about at length with gender to be eye So when we've done this in piloting spaces. I've worked in it's a two to one two physicians blind of each other's results We'll diagnose and then the AI diagnosis and you two to one So there's a rare anomaly when the AI is a differing propinion in the physicians and that is gonna help Right because you're not going one-to-one so you're not having a physician really mislead you have two that can work together Then collectively once they've diagnosed so same lesson different departments Stamford embedded AI drafts for patient messages with 162 clinicians about 20% Avaligable drafts got used Task load and exhaustion went down actual time to answer messages. No significant change UC health generated more than 21,000 drafts across nine clinics Nurses generally liked the tone help 12% of the drafts got used 12% that gap between what the technology can produce and what a working Clinician will actually accept is honestly the most honest number in health carry eye But the trajectory is legible generate information Interpret information organized work-around information then act on it which raises the thing that keeps regulators awake When one physician makes one mistake one patient is affected Okay, when one widely deployed algorithm makes a mistake Thousands of clinicians inherent the same mistake on the very same afternoon and physicians are already kind of waking up to this Right, so 88% report at least some concern about erosion of their own clinical skills 85% say doctors should be involved in deciding which of these systems get adopted at all and in the state of Utah They've started to allow AI to give prescriptions, but it's like if you dig under the hood of that It is very much the AI can Re-up a prescription that has been a known prescription of that individual But before that prescription even goes to the pharmacy the actual doctor has to give their stamp of approval So it's like it's a very minimal Integration, but I know it made headlines and it was a big whoa moment for a lot of people But it's pretty tame so somewhere tonight Clinician is closing a laptop at nine instead of 11 which if you have spent time around people that work in hospital settings primarily physicians. They work late. Yeah, you get paid a lot. You work a ton. So they're closing their laptop and nine instead of eight. The note is already drafted and the scan already has three regions circled and none of it decided anything. It just arrived earlier and better organized. The hospital is not becoming autonomous, no matter what everyone seems to be saying. It is becoming human and AI operated one signature at a time and the question is no longer whether the machine belongs in the building. It's which parts of medicine we hand over which parts stay ours and whose name is on the bottom of the page when the two of them disagree. That's all for today's exploration of the AI hospital. If you've got questions, thoughts or just want to connect you can find me on LinkedIn. Just look up Emily Laird, L.A.R. Dean. Let's keep the conversation going. Until next time, stay curious, stay adventurous and remember, you know more about the AI hospital now than you did when you were alive.