Artificial intelligence has demonstrated promise in healthcare by being able to predict patients' health trajectory based on clinical history, recommend treatments, and increase efficiencies by summarizing physician notes and automating laborious tasks.

AI tools for healthcare vary wildly in their stages of maturity and adoption, ranging from emerging to widespread, according to the US Government Accountability Office.

On the emerging end of the scale, generative AI and other machine learning algorithms hold tremendous promise to improve healthcare and administration of electronic medical records (EMRs). Using them also poses serious risks, including the inadvertent release of sensitive information and erroneous outputs.

Dr. Andrew Albano is the chief medical officer for Prisma Health, a South Carolina-based system of hospitals and physician practices that provide care for more than 1.5 million patients annually.

Albano's job is mostly strategic as he oversees Prisma's clinically integrated network connecting about 5,400 clinicians spread across two-thirds of the state. He must ensure the network remains currently viable for use and must address any potential threats coming down the road. He's also focused on the pathways through which the healthcare system delivers patient care, as well as dealing with insurance companies and technology vendor contracts.

Albano is currently evaluating AI tools that could have a drastic impact on the quality of patient care, as well as massively reduce the amount of time physicians and other clinicians spend looking at monitors instead of their patients. 

Dr. Andrew Albano 

Currently, Albano is evaluating a just-launched precision care platform from RhythmX AI, a subsidiary of SAIGroup. The AI-based platform produces patient-specific prescriptive actions and recommendations for clinicians who can further drill down using a genAI-enabled natural language interface and chatbot tools.

Albano spoke with Computerworld about the current challenges with the use of electronic healthcare systems and records and the potential AI has to solve those issues.

What are some of the issues you face from a medical records perspective? "One of the biggest problems we see with medical records is the "note float" - so what's happening is the clinical documentation, which traditionally was being used to track progress over time and is still part of the purpose of the notes, has now become the instrument for billing and coding.

"So there are other things that need to be included in the notes with each patient encounter. That creates a larger volume of cumbersome notes to navigate for a patient. So what we see from a management standpoint is time is a very limited resource for our clinicians and our clinical teams, and so to try to wade through all the information in a given patient's medical record can be really time consuming and not efficient.

"That's something we're working on - distilling the notes to be more informative and be concise. We want to carry forward all the information that's necessary and remove the information that are distracting components, while still meeting all the requirements from the billing and coding side.

"On the EMR side, the portability is not perfect. Location is seen as an outside entity, and the outside entity doesn't have the same EMR as we do, the internal entity. That poses a significant challenge for us. We have to request records, which often has a delay involved in getting those records. Obviously, there are geographic differences. So, If I'm requesting medical records a couple time zones away, I may have to wait for that facility to open for those records to get sent over, and vice versa. I think that's a huge challenge we're facing from the EMR side.

"And then just time spent in the EMR from a clinician burnout rate perspective. Unfortunately, we see that folks are having to access the EMR before and after work hours, and often times that can really erode into our clinical team's and healthcare professional's wellness."

How is AI helping with all of this? How is your healthcare network using AI?"I don't think we've really implemented AI to the level we need to. I think that AI can help with automation of certain tasks. It can maybe help with note generation. It has language models that can help produce notes with fairly good accuracy. Speaking globally for the system, I think there are tremendous opportunities for AI to do things that we're not currently doing. Is there a way to create better predictive models?

"So putting my CMO hat on and looking at it from risk arrangements from certain insurance products, it's hard for me to estimate how healthy or how sick the population we're serving is. Are we going to be able to move the needle in a favorable direction to reduce healthcare expenditure while still getting the results we all want? I think that's where AI can really be leveraged: to determine with high accuracy that this is the risk score for this patient or this population, and that would really help with contract negotiations or other arrangements.

"The other piece from AI would be cutting down on duplication of resources. With the EMR, that's outside AI's purview, but it would be great if we had interconnected EMRs. Then AI could help complement that. Instead of wading through a patient's medical record the first time we interact with them, or if they're in an acute scenario like an emergency room, AI could really parse out the information that's most applicable and give a quick summary of the patient. I think that is something we really haven't gotten a great command of yet.

"AI's ability to limit some of the volume of work [is another promising area]. I think about our radiology colleagues if they're being asked to turn out static images like from the ER and other venues. They're measured on turnaround times, but if the volume of images requested balloons at certain times of the day or days or the week, and they have to navigate through benign studies like chest x-rays or normal CTs and it's slowing them down from getting to that stat CT of the head or another study that needs their expertise, AI could be helpful in that it could hopefully review the image and then say with fairly high accuracy that the image is likely benign versus needing a radiologist overread. I see that opportunity coming up with chest x-rays, where AI can help filter out the 'normal' and have our clinicians focus on the abnormal.

"We're seeing some of that technology with some of our vendors who sell retinal imaging technology that we use for diabetic patients. That can be really helpful for the accuracy and timeliness for retinopathy or other macular degeneration."

Who are the vendors offering AI-enhanced radiological image reading? "We have a partnership with Iris we've had for a couple years for imaging. Obviously, I'm not married to a single solution. We're always looking for opportunities. We cover our employee practices, so they would obviously partner with a vendor that would be in agreement with Prisma, but there may be other individuals and practices that are independent enough that they're not working with Prisma.

"We're looking for a solution that's not only going to be more effective but also cost efficient, timely to implement, able to be scaled to different sizes. Aside from SAI, I haven't really been meeting with too many vendors at the forefront of where AI is going in healthcare. Imaging technology, so far, seems to be the most advanced for clinical use cases.

"There are those other opportunities I mentioned, for clinical documentation and EMR efficiency, but a lot of that is for language processing and not so much the totality of care. Coordinating care, filtering out unnecessary information, giving timely recommendations in terms of next steps, screening measures, diagnosis codes or hierarchal condition category codes, and then helping with coordination of [that]."

I think there's tremendous opportunities. Opportunities abound for us to better implement AI. I feel like we're in the very early stages of any implantation.

How are you using SAIGroup's RhythmX AI software? "Completely in the exploratory phase. The discussions have been progressing favorably. We're not actively using them at this point in time. From what I've gathered from meeting with their team, it can certainly help both with note generation and aggregating EMR data and making it useful.

"I think there are elements of predictive analytics that could be really helpful, in terms of risk stratification and knowing who to send to which venue or which direction in terms of care delivery. I was speaking with their CEO to ensure we're supporting physicians to do the best job they can at the time they're asked to do it. Basically, we don't want every patient with a heart to see cardiology.

"The problem is we don't have enough cardiologists, and it creates a backlog for the patients that really need it, akin to my earlier example of imaging. You really want the most appropriate study or patient going to the right clinician or surgeon and the things that are non-acuity or non-emergent to be tabled or filtered out to be interpreted by AI or a secondary clinician - a nurse practitioner or physician assistant."

"That's where opportunity exists with SAI's platform. I think the predictive analytics is helpful in that it's much easier and less expensive to deal with early-stage disease. If you get past the primary prevention point, it would be really helpful leveraging AI to aggregate information from internal and external EMRs and craft a pathway that will hopefully keep the patient from decompensating quickly over time; we see that a lot with folks who have renal disease. Often times it doesn't get detected early, because there's no universal recommendation for screening of asymptomatic renal kidney disease - so that can be an opportunity for AI to do a look back of all the patient's kidney tests for every clinical encounter they've had for years on end and then say, 'The trajectory is unfavorable. It's predicted this patient will need dialysis if intervention isn't implemented.' That's the way I envision it working.

"I think it's just a matter of parsing out noise. For a lot of healthcare professionals, they're being inundated by information, and a solution like SAI's would be helpful in cleaning it up. What are the things we need to focus on at the time we're with the patient, so we're not distracted by all this ancillary information that's perhaps not applicable.

"I'd love for AI to take away a lot of the administrative and clinical tasks for folks.When we talk about the billing and coding aspect, being able to not only give recommendations on treatments for the diagnosis - so, 'here's the most evidence-based medication or therapy or imaging' - but also, based on all the work done, 'here's the most appropriate E/M [evaluation and management] code and the associated CPT [Current Procedural Terminology].'

"So that way it does satisfy the billing and coding pieces without tasking the clinician with trying to figure all that out in real time with the patient in front of them. It frees the clinical team member to interact with patients. I don't want to ask AI to supplant or replace our clinical care team. It's really the human connection that makes the difference in healthcare."

How do you see AI eventually connecting to back-end records systems? Do you see some sort of native interface or set of APIs? "That's a great question. I think the challenge we currently face is the patient owns the medical record, and ownership means they have to give approval for others to utilize and share it. I can hold onto that patient's information, but unless they give approval, I cannot use it.

"Maybe not intentionally, it could serve as a leveraging tool for patient retention: 'It's better if you stay within [our] healthcare system, because we have all the information on you and it's easier for our doctors to provide this service.' But it can also be hurtful. If I stay within this one geographical region, it probably makes the most sense for me to stay within one healthcare system. But if I travel or split time - live both in the north and south - then it becomes a restricting model.

"So, I don't know in terms of an interface on the back end, but I feel we have to do a better job linking up any clinical encounter together with the patient being the steward of that information. So I would tie it to a social security number or some other alphanumerical code - something like that. Something the patient can readily access regardless of the EMR that they're encountering wherever the health system is. The AI could pull up that information from a cloud rather than an individual server.

Do you see any innate problems with AI? What's still lacking? "Right now we know AI can look at patterns and develop predictions and do things basic algorithms haven't taught it. [One of] the limitations I see is healthcare is evolving very quickly, so how do we keep AI at the forefront of healthcare and even leading healthcare? I think that's the challenge.

"We want it to be a useful resource, but I don't want it to become a competing entity. I don't want it to be: do you go see your doctor or nurse practitioner or physician assistant, or you do you log into your AI and input your symptoms and it tells you to go do something? That's what I have tremendous trepidation about. I don't think it will be safe for many conditions to have AI serve as the physician. What I hope is we do a good job governing AI and saying that the job of whatever service we have is to optimize a shrinking workforce and stratify the urgency of an evaluation, and then being able to take pieces of information we don't traditionally look at well.

"So, for example, health insurance, social vulnerability and social determinant, and personal preferences. And then be able to help tailor treatment plans around it. So, if I'm your treating physician and I tell you to please go do A, B, C, and D, but to do B you need reliable transportation, and for D you need capital [money] to afford the intervention. So, what I'd imagine AI being able to do is to say in order to for the patient to be able to do all those things, we'd have to factor in the following. So then we could coordinate transportation through Uber Health or something like that. For the capital restriction on D, instead of getting your prescription from a nationwide brand, maybe there's a local pharmacy that has it for$10 cheaper or something like that. And because you don't have reliable transportation, we'll have it delivered to your house.

AI could free up tadiological technicians to focus on the more urgent patient cases.

"That's where I see AI being complimentary and supportive, and not supplanting physicians. I come up with a plan as a physician, and in order for the patient to execute the plan, I have AI figure out the different variables that need to be considered. If we don't do that, I don't think it's going to be as useful as a tool. I don't want AI to be a glorified dictation software or an encyclopedia of knowledge for healthcare and it just pops in recommendations, but it doesn't do much else. Those things are fine, but in and of itself not very helpful."

What concerns do you have with AI in terms of patient record privacy and security? "In the wrong hands, any tool can be harmful. AI is a tool, so as long as the inputs and the information being used by AI are correct and evidence-based, then I think it can be a useful tool. I also say from the security side, if we're de-identifying data to make predictive models, I think that's really helpful. My concern is could it be used to sway insurance products. For example, if we use AI as a predictive model for this population and it says the propensity for this disease in this population is X, therefore we should have a wide percentage increase in premiums, that could be detrimental to patients or populations in a geographic region.

"The other piece with security goes back to portability of the EMR and how do we keep the information confidential and accessible only by authorized parties, but not keep it in siloed 'black boxes' where we cannot access it in the time we need it? Those are my biggest concerns - the inputs that are driving the outputs from AI and the portability of the medical record and accessibility in its real-time use and its use for modeling treatment plans and insurance."