Ep 37 A What, Not a Who - AI for Healthcare, Part 2

Why AI Looks Different For Doctors

SPEAKER_00 0:04

This is Dr. Patient, a podcast that examines all the aspects of the patient-provider relationship. I'm your host, Heather Johnston, MD, a real-life doctor and patient. A couple of weeks ago, in the last episode, I gave you the patient side of AI in healthcare, how you might be using it to research a symptom, translate a lab result, or like in my case, plan a workout around a bad hip. Today I want to flip the table and show you what's happening on the other side of it, the doctor side. Not like a weird sci-fi version where a robot is examining you and not the version that Ty Evelyn showed us in the last episode where an AI doctor bot doesn't know what's going on. I'm talking about real life stuff that's actually in use by providers. A piece of software drafting a note here, flagging a scan there, summarizing two years of research in under a minute somewhere else. I noticed an effect AI is having in the exam room as a patient myself recently at one of my own doctor's appointments. She set a phone on the counter, asked me if it would be okay to use it to listen and record our conversation, and that it would draft her note afterwards. I said yes because I was so curious, and for the rest of the visit, she just talked to me face to face. No typing, no turning away to click through screens. I didn't think much of it in the moment, but driving home I realized that it's a rare visit where I have a doctor's full attention the entire

Ambient Scribes And Full Attention

SPEAKER_00 1:28

time. And that's the technology that I want to start with. One of the biggest shifts in day-to-day AI use for doctors right now is something called an ambient scribe. Here's how it works. During your visit, your doctor might ask, like minded, if it's okay to record the conversation that you have. Currently 12 states in the U.S. require someone to get permission to record you, by the way, and Illinois, where I am, is one of them. If you say yes, then an app on the doctor's phone or laptop listens to the conversation, and afterwards it generates a full clinical note for the doctor, organized the way a medical chart is supposed to be organized, with symptoms, exam findings, what their assessment and plan is, etc. The doctor reviews it sometime later, edits it, and signs off. The big names doing this right now are companies like A Bridge, Ambience Healthcare, Suki, Nobla, and Microsoft's Dragon Copilot. If you've by chance heard of any of those, or if you're looking for a stock tip. I think this is a good thing, at least for patients. I mean, for years, doctors have had to split their attention between you and a keyboard during a visit, usually typing while you're talking. I actually once talked to a doctor who fell off her swivel stool during a patient encounter because she was trying to swivel so much between her computer and the patient behind her. I love that story. Other providers would choose not to type and talk, but then they would turn around after you left to spend another five or ten minutes finishing the note. Or they would try to write it that night or the next day from memory. Ambient scribes will help prevent exactly these scenarios. And adoption of this technology has moved fast. At UCSF, 70% of physicians in the health system are now using an AI scribe in their visits. At Kaiser Permanente, over 2,000 docs are using it. And of those, 84% said the tool improved their ability to connect with patients. Makes sense. And 82% reported greater job satisfaction from the reduced documentation load. This may not sound like a big deal to you, but it's important that you know that doctors spend a huge part of their day documenting everything that they do and say. So anything that diminishes that time means hopefully more time in a patient visit and a more patient-feeling personal visit overall. This, of course, has even been studied. The biggest, most rigorous study on this came out of JAMA, the journal of the American Medical Association, this past April. There was a collaboration between Mass General and UCSF that tracked over 1,800 clinicians using scribes versus 6,700 or so control clinicians across five different academic medical centers. The study tracked their progress for more than two years. The scribe users saved about 16 minutes of documentation time for every eight hours of patient care. That's honestly not as much as I thought it would be, but it does add up. That's almost an hour and a half of save time in a week. I mean, not a huge amount, but some providers received more training on the tool than others, and use of the tool wasn't consistent in the user group, so it's possible it would add up to more minutes saved in real practice. For example, in that study, the doctors who used the tool for more than half of their visits saw twice the reduction in total time spent in the charting software. So now we're saving up to almost three hours a week. But only about a third of them actually used it that consistently. Here's a real life scenario of how this might look to you, the patient. You go in for a visit about ongoing knee pain. Your doctor asks you questions, listens, examines your knee, except now there's a small microphone or phone on the desk instead of a laptop screen that they're looking at the whole time. After you leave, the AI scribe has already drafted a note capturing what you said about when the pain started, what makes it worse, and the plan that you both agreed on. Maybe it's physical therapy for six weeks before considering imaging. Your doctor skims the note sometime later, fixes a couple of details, and signs it. So overall it will feel about the same or better to a patient because you would now have more of their attention.

Portal Messages And The Portal Paradox

SPEAKER_00 5:45

The second big area where AI is coming into healthcare on the provider end is AI drafting the actual messages the provider send back to you through the electronic patient portal. That's that same software where you message your doctor about a rash or medication question. This whole messaging system with EMR's electronic medical record systems has exploded lately. A JAMA published analysis found patient portal messages rose about 153% between 2020 and 2025, and that growth landed hardest on primary care providers. But they're often replying to these messages outside of any scheduled or paid time and frequently at night, which is a real driver of doctor burnout on its own. As a patient, it's just so easy to click send on that message and not think about what happens next. But if we all do this, then there has to be some kind of staff on the other side answering all of these additional messages. But there really isn't in many offices, at least not yet. Most offices haven't hired more people to deal with the rise in electronic messages, so the employees, maybe providers, maybe not, are simply fitting more into their day, likely without any increase in compensation for it. I've brought this up to friends and family, and some have grumbled that it seems like part of their job, so what's the problem? But I'd like to know if you offer to do an extra hour of work at your job for free every day. If not, then let's stop thinking that healthcare providers should do that. As a response to the staffing problem, health systems are now testing AI tools to draft these replies. I'm assuming that a provider or someone with some kind of training reviews and sends it on, though, based on some messages that I've gotten back and heard about, I'm not sure how much effort is consistently put into that review and edit part. But a Dartmouth-led study, the first large-scale evaluation of its kind, looked at over 146,000 real portal conversations between more than 10,000 patients and their primary care doctors at a rural health system, and the study was testing drafts from several different AI models, including Claude, Gemini, and ChatGPT. What they found is something that people in the field have started calling the portal paradox. This is when the AI drafts run long or miss a needed follow-up question or include a detail that wasn't quite accurate for that patient, resulting in doctors spending real time fixing those drafts, sometimes close to as much time as just writing the reply themselves would have taken. So the tool that was supposed to save time on your inbox message can, in practice, just move the time for the provider from typing to editing. And this extra time may just cancel out the time saved from the last section we talked about, that hour and a half per week saved on scribe use. So we're kind of adding even so far here just in terms of time saved. The

Radiology AI That Reorders Urgency

SPEAKER_00 8:42

third area I'm going to touch on is the use of AI to read radiology studies, X-rays, CT scans, MRIs, mammograms. This is to date the most heavily regulated corner of medical AI and also the most mature. As of this spring, the FDA has cleared over 1,500 AI-based medical algorithms in healthcare, and about three-quarters of these are for radiology specifically. Earlier this year, a company called AI Doc, AI D O C got clearance for a tool that can triage 14 different urgent findings from a single abdominal CT scan, things like a bowel obstruction, appendicitis, or a spleen injury with pretty high sensitivity. That means how often a positive test result is correct. That was around 97%. And the specificity, that's how often a negative test result is correct, was around 98% in their studies. For you non-science people, those numbers are fantastic. That tool now runs in close to 2,000 hospitals and processes about 60 million cases a year. A competitor, query.ai, q-u-re-ai, has racked up 26 separate FDA cleared indications across nine different products, everything from flagging a suspected brain bleed on a head CT to spotting signs of tuberculosis on a chest x-ray. So this is not one tool doing one job. It's dozens and dozens of narrow, specific tools, each cleared for a specific finding on a specific kind of scan. Here's a scenario that shows how this could be really helpful. Someone comes into the ER at 2 in the morning with vague abdominal pain, gets a CT scan, and the AI tool flags a subtle finding that's consistent with a bowel obstruction, which is a true emergency, before the overnight radiologist has even opened the images. It automatically moves that scan to the top of the radiologist's queue instead of having it wait in a general line and taking longer for the radiologist to get to it. That kind of prioritization, not replacing the radiologist's reading, but instead reordering which scans get looked at first is an example of where a real benefit is with these tools. And yet, even with all these FDA approvals, actually working this into daily routine is still a work in progress. More than a quarter of physicians in a recent AMA survey said they'd received no training at all on the AI tools available to them. And only 11% said they'd gotten substantial training. So getting something cleared by the FDA and actually using it routinely and well in a medical center are two very different things, and there's sort of a gap between those right now. Another thing to mention about this idea of using the AI in radiology is the impressive accuracy. For example, AI is now a pretty meaningful part of mammography reading worldwide, with the strongest evidence for this coming from Sweden's Maasai trial. It was the first randomized control trial of AI and breast cancer screening. The results of it were published in the Lancet in January 2026. The study followed nearly 106,000 women who were randomized to either an AI-supported screening, where both an AI tool and a radiologist read their mammogram, or they went the standard route and had their mammogram read by two radiologists. The AI-involved group increased the cancer detection rate by 29%, and it cut radiologists' reading time by about 44%. On the regulatory side, adoption has moved well beyond one country. As of mid-2024, the FDA, at least here, had cleared more than 20 AI-based products for mammography in the US alone, spanning cancer detection, breast density assessment, and risk prediction, with clearances continuing into 2026, including tools like Deep Health's AI-powered breast ultrasound, that was cleared in July of 2026. Together, these point to AI functioning less as a replacement for radiologists and more as a second reader that catches more cancers while reducing the workload. And that idea of a human working alongside AI is something I'll keep coming back to because it does seem to be the sweet spot currently.

Bias Risks In Medical Algorithms

SPEAKER_00 13:12

This does seem like a good time to bring up the issue of algorithmic bias. Remember that AI learns from data, and if that data doesn't represent everybody equally, the tool won't work equally for everybody. For an example, on this topic of mammograms, the accuracy of the algorithms to pick up breast cancer will be based on what type of breast tissue the tool was trained on. For example, dense or not dense breasts. And then it will also depend on the patient who it's reviewing. An AI tool that was trained mostly on non-dense breast tissue will perform worse when reviewing mammograms of women with dense breast tissue. Another example is with dermatology, that's the specialty of the skin. When I was in practice at the University of Chicago, the vast majority of my patients were African American, but the dermatology or skin health book that I had been assigned in medical school and often referenced contained mainly photos of Caucasian skin. Well, things don't look the same on white versus black or other color skin, and I had to get a collection of dermatology books specifically geared towards the population that I was seeing. There have been issues with this very scenario with dermatology-based AI algorithms, which are sometimes trained more extensively on lighter skin tones, so are worse at catching skin cancer, for example, on darker skin. This issue should get more and more minor over time as AI tools are fed more and more and hopefully varied data. Mammograms and other scans aren't the only images AI is reading right now. There's another field of medicine called pathology, which is when doctors look at tissue samples under a microscope to diagnose things like cancer. Pathology got its own FDA authorized AI tool a few years back. It was a product called Page Prostate, P-A-I-G-E. It flags areas on a digitized biopsy slide that looks suspicious for cancer before the pathologist starts their own viewing and reading. In the study that got it authorized, pathologists using that tool alongside their own judgment caught meaningfully more actual cancers. Again, the pathologist is still the one signing the diagnosis. The AI is just helping prioritize certain slides first and is also making sure fewer subtle findings are missed.

Sepsis Alerts That Can Save Lives

SPEAKER_00 15:33

There's a fourth category of AI working in healthcare worth naming because it's just so interesting, and that is AI that watches your vital signs and lab trends in the background and raises a flag before a human might necessarily catch it. It's currently most used for sepsis, which is when your body starts shutting down when you have a life-threatening infection in your bloodstream. The condition can be fatal, and every hour of delayed treatment raises the risk of death. Duke Health built one of the earliest versions of this type of AI called Sepsis Watch, which has been used since November 2018, and deaths from sepsis at Duke in their published study dropped 27% after it rolled out. That's a crazy decrease. And that decrease was also seen with a newer system called TrueS, T-R-E-W-S, built at Johns Hopkins, where when a clinician confirmed the alert from the AI tool within three hours, the in-hospital death rate dropped by 18%, along with shorter hospital stays and less organ failure. This sounds hopeful, and there are other even bigger tools out there in use for this purpose. However, when one of the bigger tools that Epic built for this was studied outside of the hospital system, it actually performed much worse, missing roughly two-thirds of sepsis cases in one study out of Michigan. So the exact same category of tool, watching vitals and predicting sepsis, can either save lots of lives or can get it wrong a lot, depending on how well the tool was built and validated before a hospital system trusts it enough to use it. Jury is still out, I guess, on this category, though something to watch because it's really very interesting.

Pharmacy Warnings And Alert Fatigue

SPEAKER_00 17:17

A fifth area of AI use for providers is in the pharmacy world. Most of the time, when a doctor orders you a new prescription through the electronic health record, there's almost certainly software that's quietly checking that order against every other medication you're on, as well as your allergies and your kidney and liver function if it has that data, and it pops up a warning if something looks dangerous. This kind of clinical decision support has been shown to cut the risk of a medication error or a bad drug reaction roughly in half. On the downside, these softwares tend to be a little trigger happy and override rates when a provider simply clicks to ignore a warning on these drug interaction alerts routinely run above 90%, meaning doctors click through the vast majority of them. Research suggests that's not simply doctors being careless or lazy. The thing is a lot of the alerts really are low value and not tailored to the specific patient, so they create alerts on drug interactions that are clinically unimportant for that patient, and the doctors correctly wave them off. But that volume of notifications causes something called alert fatigue when a system throws so many warnings, including a lot of genuinely minor and unimportant ones, that the human being on the other end just stops actually reading closely. And studies show sometimes that even the alerts that doctors think are important still get overridden. So it's a real life, likely unintended design flaw to the tool that has to be acknowledged in future designs.

Prior Authorization Turns Into AI War

SPEAKER_00 18:51

While I'm on the subject of AI sometimes creating extra work instead of saving it, there's a strange AI race happening in the prior authorization world right now. For those of you who are lucky enough to not know what that is yet, prior authorizations are a process undertaken by health insurance companies when they want to review a suggested treatment or test recommended by a provider before they just say it's okay. Typically, the provider starts the whole thing off by saying something like, Hey, you need an MRI of your back to figure out why your leg is numb. The doctor's office already knows that your insurance will require a prior authorization and will sometimes use AI to complete a prior auth request. The insurance company very likely is also using an AI tool to review that request and either denies or approves it. If it gets denied and someone has to appeal the decision, that same medical office might again use AI to submit a claim appeal. It's literally becoming some sort of a robot war. Startups like Claimable and Counter Force Health build AI tools that draft and file prior authorization appeals in minutes instead of hours, and providers are widely adopting them. One physical therapy chain told a reporter that a tool called SPRI, S-P-R-Y, cut their appeal writing time from hours down to minutes, and that without it, they'd have to start turning patients away just to keep up. And it's working. Claimable reports reversing roughly three out of four denials at appeals and counterforce reports about seven and ten, although that success rate may be due to the program, but there's also lots of emerging evidence that prior authorization denials that are appealed are usually overturned. So there's that. All of this is related to today's episode because the average physician's office reportedly spends about 13 hours a week just on prior authorization paperwork, according to a 2024 AMA survey. And the healthcare industry spent nearly $1.3 billion on the process in a single recent year. A former chief data officer at Kaiser Permanente, United Healthcare, and Optum put it more bluntly. He predicts we're heading toward what he called, quote, agent war's galore. It's almost comical until you remember that real people's real medical care is sitting in the middle of that fight. Zooming out to the big picture for a minute and looking at numbers on overall AI use by providers, adoption is really taking off. The AMA's most recent survey from this past winter found that 81% of physicians now use AI in some form in their practice, more than double the 38% who said the same back in 2023. The single most common use is doctors using AI to look something up, like you do. There's a tool called open evidence that's become the dominant player here. About 65% of U.S. physicians, including me, now use it, and it logged over 1 million. Physician consultations in a single day this past March. It works like a search engine that you're using, but it's built specifically for medical evidence. So a doctor types in a clinical question and gets back a summary of the actual research, including sources and footnotes. And the replies are in heavy medical language meant only for physicians. An issue tied to this is something I brought up in my episode on second opinions, which is that medical research is coming out so fast now that no individual doctor can read all of it. And that's actually where use of this tool comes in. A cardiologist with a patient on an unusual combination of medications can ask a tool like this to summarize the last two years of research on that specific combination in under a minute, instead of spending a whole evening digging through research databases themselves. The tool isn't making the treatment decision, although I'm sure that some providers are asking it to. It's compressing the research legwork, saving a ton of time. Note again the pattern with these tools, saving providers time in their day. That sounds great to me, but also scares me in what providers will be asked to fill that with now that there's available time.

Liability And New Rules On Review

SPEAKER_00 23:15

I want to move on to talk for a few minutes about some potential consequences of all of these uses of AI in healthcare for providers. I'll start with the idea of liability because it's such an interesting theoretical concept. Under the current legal framework, doctors and providers are the one responsible for decisions that they or people under them make. If use of an AI tool contributes to a medical error, the provider is still the one who's legally responsible. Meaning, theoretically, because I couldn't find a single legal case like this yet, a doctor can be found at fault for trusting an AI recommendation too much and not double-checking it. And to take it further, I wonder if a doctor could also be found at fault for not using an available well-validated AI tool that may have caught something. Hospitals and the companies that build these tools might also carry some liability if the system itself was defective or poorly implemented. But at least for now, the physician and their own decision making remains the main issue. Picture a radiologist reading a chest x-ray with an AI tool running alongside them, flagging a small nodule as likely benign. If the radiologist agrees and moves on, and it later turns out to be cancer, the fact that the AI also missed it probably won't shield the radiologist. They're still the one who signed the reading. Now flip it. Imagine a hospital had that same AI tool available, but the radiologist's group decided not to license it to save money, and a nodule gets missed that the tool would have caught. Perhaps that hospital and that group could end up defending a lawsuit on the premise that a reasonably available safety tool should have been in use. This is going to be such a fascinating trend to watch over the coming years as AI is threaded into everyday healthcare life for physicians and other providers. There is some legal movement on this front, though. Texas now requires hospitals to have a real human review AI outputs in the medical record and to actually tell patients when AI assisted in their diagnosis or treatment. Alabama, Indiana, Utah, and Washington have gone after a narrower but very real problem. They now prohibit insurers and providers from relying on AI alone for adverse decisions or denying a prior authorization, for instance, without independent human judgment behind it. And this isn't hypothetical concern from lawmakers with too much time on their hands. In December of last year, a coalition of 42 state attorneys general sent a formal warning letter to AI companies after chatbots were linked to multiple deaths, putting them on notice that they could face liability under existing state law. So the legal system is still figuring out how to handle AI in healthcare and medicine, and lots of changes are likely coming.

Burnout Relief And Training Gaps

SPEAKER_00 26:09

I also want to circle back to an issue that I did a whole episode on a while back, and that is doctor burnout and moral injury. And this is another area where there's some potentially direct consequences from AI. Since the medical world and the US likes to study everything, there are, of course, studies emerging on this very issue, AI affecting doctor quality of life. This is important because in the U.S., physicians experience burnout at a substantially higher rate than for other professions. Currently, it sits at about 40% of U.S. docs experiencing at least one symptom of burnout regularly. This leads them to leaving medicine, like me, or perhaps switching jobs or just being in a bad mood. So can AI help with that? Maybe. A couple of recent studies looked at physicians using ambient scribes and found that just 30 days after starting to use one, burnout in outpatient clinics dropped from about 52% down to a reported 39%. MassGeneral saw something similar system-wide, a drop of reported burnout from about 53% down to 31%. And this isn't just doctors feeling better for their own sake, because 56% of the patients in that same research said the ambient AI tool had a positive impact on the quality of their own visit, which tracks with exactly what I described at the start of this episode. It's nice to visit with a doctor who can actually look at you instead of a keyboard. And in case you're wondering, your provider or your doctor is enjoying it as much as you are. If you've listened to my burnout episode, which is episode seven, by the way, you know how much of that crisis comes down to documentation burden specifically, hours of charting stacked onto an already full day. Though, these tools will only potentially help if providers actually use them. From docs that I talk to, one of the biggest barriers to use is simply not understanding or knowing how to effectively use these new tools and not being given protected time from work to actually learn how to use them. If you never have a moment free, like most doctors, then when or how will you learn to use these tools? In addition, I do still have a concern that as more AI tools are put into use for providers' days, that will free up time, but what will happen with that free time? Will it be given back to the provider and their patient, or will it just be filled with more work to do? Medical centers, hospitals, and practices have to acknowledge these limitations and dangers and set up that protected time for providers.

What Patients Can Ask For Now

SPEAKER_00 28:44

So standing back, here's the thread running through everything in this episode. Ambient scribes, portal drafts, radiology flags, sepsis alerts, literature searches, it's this. AI is doing well whenever it's handling a task that has a clear right answer. And when a human still checks the output before it affects you. And it's doing poorly or at least unevenly, whenever the checking is inconsistent or the incentive to skip it is strong. A scribe drafting your note gets checked by your doctor before it's final. A radiology flag gets checked by a radiologist before anyone acts on it. A sepsis alert gets confirmed by a clinician before the treatment changes. That single difference, a person actually reviewing this before it affects you, is one of the most significant parts of whether a given use of AI in a hospital right now is something to welcome or something to watch closely. The tools working best are the ones handling documentation and administrative drag, freeing up a doctor's attention for you. The tools with real unresolved risk are the ones where the checking is thin and consistent or where the financial incentive quietly favors moving fast over reviewing carefully. I'll close with a few practical things worth knowing as a patient sitting across from a doctor who might be using any of this. If your doctor's office uses an ambient scribe and discloses that to you, I mean, I'm not sure they have to in some states, you can ask them to turn it off for any part of the visit. That might be your right depending on where you are. If your doctor mentions using AI to look something up mid-visit, that's very likely a literature tool like open evidence, helping them double check current guidance. It's not a red flag. Doctors using AI to stay current is one of the more low-risk, high-value uses that I mentioned in this whole episode. If you're ever told a scan was flagged or reprioritized by AI, know that the final read itself still comes from a radiologist. The AI just helped decide what order things got looked at in. None of this technology is going away, and most of it is going to keep improving. The job right now for doctors and providers and for you is knowing which parts of it are ready to lean on and which parts still need a real person checking all of the work.

Book Pick, Subscribe, And Disclaimer

SPEAKER_00 31:05

That's part two of my look at AI and healthcare in a tiny nutshell, meaning I probably could talk for a couple more hours about this topic, but I want to be kind to your time. I strongly recommend the book A Giant Leap by Dr. Robert Walkter, a professor and chair of the Department of Medicine at the University of California, San Francisco. He's written hundreds of articles and a handful of books, but this one on how AI could be useful in medicine is thorough and fascinating. And if you missed part one on how patients are using AI for their health care, go give that one a listen as well. Thanks for listening. Thanks for listening today. To catch up on more episodes and to get new ones delivered directly to you, subscribe wherever you find your podcasts. Apple, Google, Spotify, iHeartRadio, and more. If you'd like to be a guest or have an idea for an episode, let me know at www.drpatientpodcast.com. That's doctorpatientpodcast.com. Here's the disclaimer. Even though I am a doctor, I'm not your doctor. These stories, my comments, and all discussion is purely reflection about what's working in the healthcare system and what isn't. Don't use any medical information that you hear in these episodes to diagnose or treat yourself. If you have a question about your health, get in touch with your doctor or local health clinic.