Microsoft Chief Medical Officer David Rhew, MD, gets beyond the hype on the value of generative artificial intelligence (AI) in healthcare today.
Cure: There is real hope for AI to do good in transforming the way we conduct healthcare. In day-to-day terms, how is AI going to be applied to healthcare in the immediate future?
Rhew: One of the biggest challenges in healthcare today is that clinicians are burned out. We have a workforce crisis. Many individuals find that the administrative workloads are crushing their ability to take care of patients and their morale. How can we reduce that? Typically, we’ve thought maybe we needed to have scribes. But AI can do a really good job at this — better in many ways than a human.
As we’re having a conversation with a patient, with AI in the background, we now have the ability to have that information captured and be converted into a clinical progress note. Then, with a few edits made by the clinician, the note is integrated back into the medical record.
This saves each clinician about seven minutes per encounter and reduces the total amount of documentation time by 50 percent. Immediately, there’s value. And the patient enjoys it because now you’re actually looking at them, and really paying attention to what they’re saying.
This is real. It's called ambient clinical intelligence[EH1] , and it's something that we could see changing the way that doctors and other clinicians interact with patients and that would allow us to provide better care.
Cure: In your examples, the physician is still involved. Although AI is processing all of this big data, at the end of the day, a human is QCing it and making decisions. Should people be reassured that AI is not taking charge of our healthcare?
Rhew: That’s right. We like to call it ”the human in the loop,” and at Microsoft we refer to the technology as our copilot.
It’s helping with a lot of administrative tasks, everything from finding information from disparate sources, pulling it together, synthesizing it — in this example, into a clinical progress note—and then presenting it back in the right structured format, such as the SOAP fields, the subjective objective assessment plan.
We’re finding this type of capability — oftentimes done by highly trained individuals, perhaps even physicians or other clinicians — can be done through automation and through AI. That is just an extraordinary opportunity for us to think about how large sequence models can help the clinician in their day-to-day.
Cure: So the data need to be complete, representative and accurate. What else?
Rhew: With regard to the data, it's important that we understand that AI is very dependent on the size and diversity of the dataset. That's true for all AI, traditional AI as well as generative. But what we find with generative AI is that because the datasets are so large, oftentimes the key factor is around how we provide the prompting for the specific information. What we're specifically looking for may take us down different paths with potentially different answers. Because of the variability in how we may ask the prompts, we may end up with different or potentially erroneous responses. So it's really important as we think about this human in the loop to be aware that AI is here to help us, but it's not always perfect, and that's why we have to make sure that this is still being driven by the human.
Cure: How can AI transform drug discovery?
Rhew: When you look at clinical research, a lot of it is about gathering information, synthesizing it, creating a hypothesis — which is a very creative process — and then testing it. Then from there you pause, you look and you pivot. This process of iteration is something that can be greatly accelerated if you have a tool such as AI. I oftentimes find that humans, in their hypothesis generation, may only be aware of certain aspects. But the computer can actually think of things that we may not have ever thought of before. So in that context, we may be able to explore things more comprehensively much faster.
Cure: Do you think that AI can get us to what we ultimately want, which is cures and preventions?
Rhew: Oh, absolutely. And largely because our systems have gotten better, we have greater capabilities, we also have models that have matured in many ways. Then the whole advent of the generative AI has given us a new set of tools that can leverage things. And we've seen this, whereas the iterations have gotten at more parameters better, the outputs are more accurate, more reliable, and we're learning a lot more how to use it. We're really at the precipice of some extraordinary things with the use of AI.