Redefining the Doctor Visit: Evaluating AI’s Role in Clinical Decision-Making

Artificial intelligence (AI) is playing a growing role in healthcare, expanding beyond administrative automation to shaping clinical decisions. Today’s AI tools can process large volumes of data to detect disease trends, support medical imaging analysis, and recommend treatment plans grounded in clinical evidence. These advances offer opportunities to boost efficiency, improve diagnostic accuracy, and elevate patient care. At the same time, they bring challenges, including algorithmic bias, data privacy risks, and the essential need for physician oversight. For AI to truly benefit healthcare, its integration must be thoughtfully guided to support but not replace clinical judgment and ensure better, more accessible care.

A study evaluating a clinical application of an AI model was presented by Zehavi Horowitz-Kugler, MD, a primary care physician at a private practice in Israel and vice president of Medical Sciences at K Health at the American College of Physicians (ACP) Internal Medicine (IM) Meeting 2025, held April 3-5, in New Orleans, Louisiana.

HCPLive spoke with Kugler and Ran Shaul, cofounder and chief product officer of K Health, while at the meeting, to learn more about the model they developed, the study that evaluated it in a virtual urgent care setting, and the potential clinical impact of the model for patients around the world.

HCPLive: How have you seen AI been moving into the medical field a little bit these last couple years?

Zehavi Horowitz-Kugler, MD: There is a growing interest in AI, and we feel it with the hospitals that we've been partnering with. I think physicians are, over time, more open to it and using it. And of course, studies like the one we've just published will help gain trust. I think this study is unique because we are talking about AI that can really be part of the clinical workflow. We're not talking about all the more technical issues in the visit, like scribe tools and things that help physicians be more efficient. We are talking about clinical quality here and actually being part of the clinical workflow with the patient.

Can you discuss the study itself and what you did?

Kugler: The study compared AI recommendations to physicians' decisions in real visits on our virtual clinic. It was a joint study between K Health and Cedars-Sinai. What we did is we took a sample of cases that focused on the most common acute complaints—like respiratory issues, sore throat, cough, etc.—genitourinary issues, and eye and dental complaints.

For this sample of cases, we had expert adjudicators review all of the cases where there was any discrepancy between the recommendations the AI gave and the actual physician decision, and also a subset of the cases where they fully agreed with each other. The adjudicators rated both of them on a 4-point scale from optimal to reasonable, inadequate, and then potentially harmful.

The results of this comparison were that the AI was rated as optimal in 77% of the cases, compared to physicians’ decisions, which were rated as optimal in 67% of the cases. The AI was also less frequently rated as potentially harmful compared to physicians.

Notably, the AI and physicians were rated equally about two-thirds of the time. But if you look at when each of them was rated better, in 21% of the cases, the AI was rated better than physicians. You can see in the study that this is mainly because of the AI's ability to better use EMR data as well as the intake data to better capture red flags and alarming symptoms, and it is more guideline-adherent than physicians. Physicians were rated better than the AI 11% of the time, and that was mainly in cases where physical findings were very important to get to the diagnosis and management.

The next step would be going to a randomized controlled trial, a prospective study in which we will compare 2 arms: 1 where the physicians are going solo, and the other where they use the AI. In this study, we're planning to look at outcomes—not just comparing the decisions of the AI and the physicians but actually seeing what happened afterward with the patient. So that's coming soon.

How was the AI model integrated into care in clinical centers in the study?

Ran Shaul: The thing I'm most proud of is not about AI or a celebration of a model or algorithm. What was described already works, and many, many patients are actually using it. We work with leading health systems like Cedars-Sinai, Hartford Health in Connecticut, Hackensack in New Jersey. Those systems are basically working with us in a partnership to put a product in front of their patients that pretty much reengineered the idea of going to a doctor.

Imagine going to a doctor today, calling someone and making an appointment, and that's pretty much it. Maybe somebody asks you a question like, “Why do you want to be seen?” and you say, “I'm not feeling good,” and that's the experience. Maybe you go to a brick-and-mortar clinic and then you fill out a form. We completely revamped that.

The idea of going to an appointment in our system with technology in those health systems has changed. You're going into a very intelligent conversation with a machine that is as smart as a doctor we just described. You can express all your symptoms and go through a process that is intelligent and works in an engaging way. Most people are highly engaged in the process. You give all your complaints, all the reasons, when it started, what else is painful.

Then, when you see your provider, everything is already served for the patient and the physician. So the consultation is actually quality time. It's not like, “Wait, wait, what did you say? Let me type this. Oh no, I forgot.” The time is really focused on giving you and the physician the opportunity to build on everything that has been collected and make better health decisions. It leads to better, high-quality care. Both the patient feels empowered, and the provider feels empowered. We think we're really changing the way we consider what a doctor consultation of the future looks like.

How has the clinician response been to the AI model?

Kugler: We've been working with medical groups in our partnerships, so you see all different kinds of responses to the AI. I think you have physicians that are more experienced or more tech-savvy that embrace it and use it very often—even just to reassure themselves before making a decision. And other physicians that are more attached to the way they're practicing might even tell themselves that this is trivial, that treating a UTI or an upper respiratory infection is not something that you need help with. I think the study demonstrated that that's wrong and that it could be valuable even for the more simple, bread-and-butter cases.

How could the AI model help improve access to care?

Shaul: We live in a country where you can still see a physician. But there are many countries out there where, no matter what you do, you're not going to be able to see one. So, wouldn’t it be helpful to have a high-level, accurate diagnosis and treatment option for the masses? I think there's an opportunity here, because we know that healthcare is in crisis globally, and people are just trying to get better.

This content has been edited for clarity.