Enhancing Patient Outcomes Through Thoughtful AI Adoption in Family Medicine

In August, the American Academy of Family Physicians (AAFP) and Rock Health co-published a new whitepaper offering a roadmap for responsible AI (artificial intelligence) integration in primary care. “While AI is poised to reshape primary care, the critical question is how it will do so,” the study stated. “If implemented thoughtfully and in close partnership with clinicians, AI could help stabilize primary care and reverse the impact of decades of underinvestment and fragmentation.” 

The report noted that “Primary care touches more patients than any other part of the healthcare system and handles a wide range of issues, often with incomplete data and limited time. That makes it both a clear proving ground for AI tools and a perfect testing ground for their ability to support real-world care.”

The study's research included perceptions from 1,300+ clinicians, as well as insights from the Starfield Summit on AI and Digital Health, and guidance from leading primary care experts.

Healthcare Innovation recently spoke with AAFP’s CMIO, Steven Waldren, M.D., about the report's findings. Waldren is a family physician by training and has been a clinical informaticist for more than 20 years. AAFP, he explained, has been advocating for safeguards to be put in place for adopting AI in the field.

What do you see as the urgency of having AI become part of operations in primary care practices?

One is just the rapid advancement of it. Everybody's trying to think about how they can leverage it to drive value in the marketplace, either to create a new business, drive revenue, or really help primary care or other parts of the healthcare system. I think our colleagues in radiology and nephrology have been leading the way in regards to putting AI in the clinical space.

The other thing in primary care…is the epidemic of administrative burden, and how the current technology in the EMR space really hasn't met the needs of the frontline clinicians. We’ve seen docs talk about these types of AI solutions being game changers. We've done some evaluations in the AI scribe and ambient listening realm on patient record summarization, and highlighting the care gaps that are there, and the quality measure gaps that are there, especially for those docs who are in value-based care.

Then, some newer technology is coming to market around the EHR inbox. The pandemic has led to a large volume of messages compared to pre-pandemic levels. Docs are trying to struggle how to fit that in.

What other developments are you seeing?

One thing is Open Evidence, which uses a large language model (LLM) to allow people to ask questions of the scientific literature. I keep hearing more and more docs talk about leveraging that. The other are just generic LLMs, like ChatGPT and Gemini Claude. Docs are starting to use those to create patient instructions.

On page 12 of the report, it said AI solutions may handle more routine tasks, such as documentation triage and patient education, while humans focus on empathy, complex decision-making, and relationship building. Could you speak to that?

You take the great advantages of AI and the great advantages of human intelligence, and put them together. I think that's where we'll see a lot of use of AI. I think there's this discussion around what is the best collaboration between AI and humans, and what are humans better at? What is AI better at? And how do you put those together in a workflow that makes sense and demonstrates the highest value for patients and the industry. I think that's the work that needs to be needs to be done. I think we have some good examples of that right now in some of the ambient scribing stuff. I think more needs to be done on the clinical side to figure out how to best leverage that. If you can get rid of a lot of these administrative tasks, then you're able to give physicians…breathing room to start thinking about having to be more empathetic.

There was a study with patient notes or patient message responses in Epic, and patients felt it was more empathetic in the study. It's not because the docs are less empathetic. It's just that they don't have the time to sit down and write a lengthy note. What they found in that study was that it didn't save the docs' time, but they felt like their messages to patients were better. Adding more to that capability, I think you can actually reduce the burden as well as create better messages back to patients in which patients feel like they're being heard, make them more engaged and more likely to be adherent with their meds, and therefore have greater outcomes.

On page 14, it states that AI can help surface meaningful, human-centered insights related to social barriers, behavioral health needs, learning preferences, and communication styles. Could you explain a bit more about that?

That sentence talks about kinds of inputs and outputs. On the input side, being able to pull in data. Some people call this data exhaust. We go about our days right now with our smartphones; for example, you have geo data, and you may have reviews. You may have a diabetic who seems like they're always going to KFC. There's now an opportunity to say, well, what are some alternatives that you could do either at KFC or someplace that's near or on the route? Being able to do that work would be a fair amount of work. If AI can start to bring some of that data together and create some insights on top of large volumes of data, that could be a great opportunity.

On the output side, being able to meet the cultural style of the patient. You have that ability, potentially, to be able to send them information in their primary language without a significant amount of cost in transcription or translation services. You also have the ability to say: "This is a person who has an eighth-grade reading level; take my message and create a version of it that's at an eighth-grade reading level to make sure it’s easier for them to comprehend.” Or if you have somebody that's super anxious, as an example, can you do that with a more reassuring tone? They're (patients) getting the trust of their providers, and therefore are more likely to fulfill the recommendations that are coming from the docs.

What does shared decision-making look like with AI?

I have a vision where the doctor has their AI, and the patient has their AI. The patient may say: I'm going to see my doctor and follow up…based on my medical history…what are the top five questions that I should ask my physician? They come prepared. Then on the physician side….What are the top three, five things that are going to be the most impactful for this patient's health and wellness?

I think there's also some AI there that could actually then do a follow-up afterward. You have the ability for both to be informed. When they come together and make a recommendation, the patient has trust that that's the right one. That's where the traditional kind of…shared decision-making came from. I think now you layer on AI on both sides to make that more effective and more efficient.

On page 16, there was a bullet point stating that AI-powered panel stratification sorts patients by risk, complexity, or social needs and guides the team on what actions to take and when. Could you walk me through this?

I think what this is talking about is what some people call panel management. Some people call it population management. If you get into the public health sphere, they think population management is a much larger concern. We agree with that. What this was really focusing on is the notion of panel management.

As a family physician, you normally have somewhere between 2000 and 3000 patients on your panel. Not all of those folks come in to see you, even though they should. You don't have enough time to do all the primary care for 2500 patients. What this technology has allowed you to do is triage, so that you say, the patients that need to come in aren't coming in. How can we send a message to encourage them to come see us so that we can help them manage some of their chronic disease? Or, they haven't seen us for a while to do prevention-related stuff, and then you have the opportunity, maybe even with AI, to say, here are patients that meet all criteria to order a mammogram. The AI goes out and helps them, encourages them on why it's important.

Those are some of the things that I think AI has the capacity to be able to do. I don't know if that's in the market, but that's within the current state of the art.

How do prompts and clinical decision support tools during visits identify care gaps?

We did an evaluation of a product which…summarizes the chart. It creates a problem-oriented summary. It pulls the data all together in one place. If it's connected to the EMR…it looks through all that data to see if there are care gaps that this patient should have based on their age and diseases. It puts it out there, saying that this patient has not had certain preventive services in the last 10 years, and based on guidelines, they should have them. It highlights all those things.

Could you speak to some of the barriers when it comes to AI adoption?

One of the biggest challenges is that the current EMR data is not being captured to really drive good training data for good models. It's all focused on billing and documentation for building purposes, and to give a record for the docs that are seeing a particular patient. I think that's a challenge.

I've been working on interoperability, and while we've made a lot of great progress, we're nowhere close to where we need to be to make sure that the data can flow like it needs to, so that these AI models can deliver that either for training or for the actual prediction that they need to do.

The interfaces on the APIs for the EMRs still are not super robust. There are requirements for that in the certification process, but there's not a requirement to build the other infrastructure that's around that.

We don't have a good model for determining risks for these systems. Where are the lines to say that this is high risk, medium list, low risk? We still need to do that as an industry.

Who do you feel needs to sit at the table when decisions are made regarding AI implementation?

I think you have to have a multidisciplinary team. In our view, the physician clinician needs to be the leader to make sure that you get the right problem and that the solutions that you're looking at will actually solve the problem. Healthcare is a complex adaptive system. You need to have your IT folks in there as well. You have to have your security folks say: What new types of threat factors does this open up for us, and how do we mitigate those? Your operations and other support staff are going to be interacting with the system as well.