The Permanente Medical Group (TPMG) recently completed a pilot of an ambient AI system that doctors would use in real patient visits. The goal was to reduce additional admin time that doctors face because of having to write notes and do more charting after a visit. The AI would listen to the visit and generate a transcript and visit summary for the physician to use when typing their note. They published the results which showed that about 34% of the physicians in the pilot used the tool, in a cumulative 303,266 encounters. There was even one physician that used it on their own in 1,210 encounters - talk about a super user.

The study showed a decrease in “pajama time” which is defined as “time in clerical activities outside working hours” and time outside of 7 AM to 7 PM which sounds like it could also include pajama time but was unclear. Patients also appreciated the AI with 81% reporting that their physician spent less time looking at the computer screen than previous visits.

The adoption numbers aren’t incredible, but overall this shows that there is some promise in having an ambient AI in the patient room with a physician. There are a lot of questions that come up with a study like this, so I want to take some time going through the pros and cons of ambient AI, physician engagement with the platform, and the patient experience.

The Good and the Bad

The ambient AI tool at TPMG sounds futuristic. It’s a mobile device that the physician can turn on (after getting consent from the patient) that just sits in the room and listens to the conversation. It then produces a visit transcript and a visit summary in an app outside of the EMR for the physician to reference when writing their note after the visit.

The biggest pro is the time saved, which is what the study aimed to show. By having the AI summary and transcript, you can write your note faster and have access to the exact conversation for later reference. They didn’t show that appointments happened any faster (i.e. typing things on a computer during the visit doesn’t appear to slow doctors down). This is good if you ask me - doctors are spending more time speaking directly to patients without worrying about real-time documentation. There aren’t many tools that are created just for physician wellness and to prevent burnout, so I’m for this - though I’m sure there are also many other ways that these transcripts can be used to improve billing and things like risk adjustment down the line.

With any new tool, there are going to be hurdles. The biggest one is the 34% adoption rate in the pilot. The biggest reason for this, as the article documents, is that new technologies are hard to use and take additional time to get set up with. The other downside of this tool is that it lives outside of the EMR. Context switching as a physician is burdensome. If you’ve ever used an EMR like Epic you know that it isn’t the most fluid UI. Switching between tabs in the EMR itself takes a second or two and there are countless popups to click through (clinical workflows someone at some point made but didn’t think through the fact that adding workflow after workflow has diminishing returns). This is difficult enough, but having a tool outside of the EMR entirely is a big ask. There are only a handful of tools that have provided enough value (UpToDate or MD Calc) to get doctors to actually leave the EMR. This is a problem the AI team is aware of and addresses. The eventual goal is an EMR integration. Hopefully once that happens, adoption will go up.

Physician Engagement

Even with EMR integration, we need to address behavior change. Doctors are creatures of habit and many would rather brute force their way through tech instead of learning efficient workflows like keyboard shortcuts or smart phrases. It’s not that they don’t care about efficiency, it’s just that learning something new takes time. Time that they don’t have. The way they are going about their work may not be the most efficient but it is efficient enough for them to get the job done. Adding a new piece of technology rocks this boat. It may be interesting to an engineer working at a tech company. But a doctor that has been in practice for many years with a system that is a little annoying but works for them isn’t just going to add a fancy new tool to test out the waters.

In the case of the AI tool at TPMG, using the summary tool is only part of the problem. Even before a physician gets to an AI-generated summary, they have to carry a second device, get patient consent, turn it on, and use it during the visit. TPMG is using “physician champions,” a title I wouldn’t wish upon my worst enemy, to get doctors onboard. The premise is simple - doctors listen to other doctors. If you’re recommended a tool by your peer, you’re more likely to take it seriously. It’s the right way to go about it, but it still has a long way to go.

The Patient Experience

On the other side of this whole experience is the patient. There is a big focus in health tech, especially with AI solutions, on how patients will feel about having more and more AI integrated into their care. This makes sense when thinking about tools like chatbots that may be “auto responding” or at least creating messages for physicians to send. There is a feeling of cringe to know that you’re getting an “empathetic” response from an AI instead of an actual human being. In the case of ambient AI in the doctor’s office, this is actually reduced greatly. If anything, the AI should make the doctor more present in the moment with the patient, focusing less on typing and more on talking and listening to what the patient is saying.

We may also think that patients will feel strange to be consented for AI use. I don’t think this is going to be the case. There may be an initial adjustment period, but there are so many things that happen in the doctor’s office that are out of the ordinary for people (e.g. lab tests, retinal scans, foot exams, (God forbid) DREs), most people are open to what the physician wants to do as long as it is well-explained. There is of course always opportunity to answer questions and there is even the opportunity to decline the AI. This is totally fine and I think the stakes are relatively low here.

The one piece that I’m interested in is how much of the AI output is communicated back to the patient. The 21st Century Cures Act already allows notes and lab results go back to the patient, something that has caused some amount of division on the provider side. With AI generating transcripts and summaries, what will the patient be able to see? Having a fully generated summary keeps the doctor accountable and serves as proof of what goes on in an encounter. This could impact malpractice cases but at the very least, if patients have access to this, they can better understand what happened during a doctor’s appointment. Even better, family members who couldn’t make it or parents who have young children and are trying to juggle multiple things can have the documentation they need to look back on everything that took place during that 20 minute appointment. These are more data points for patients and if communicated properly could elevate the patient experience.

Some Parting Thoughts

It’s still early for AI scribes, but this pilot is telling of what is to come. There’s work to be done around EMR integration, physician adoption and engagement, and the patient experience. It’s not perfect and there is a whole other discussion to have of how adding AI into the mix may impact outcomes but it’s promising. Bringing about changes in healthcare is hard, but starting at the physician level and focusing on workflow is a great way to promote change that has the potential to be adopted. If you can’t get the providers on board, it’s hard to do anything in healthcare and that’s why I’m hopeful for this technology to keep getting better.