Healthcare’s long road to digitization gets an AI boost

The University of Pennsylvania Health System had an enormous amount of anonymized patient data in its Penn Medicine BioBank, and SVP and CIO Michael Restuccia’s team saw an opportunity to use it to benefit the research hospital’s patients.

“We had a conversation about how to take some of the innovation occurring in research around AI and deploy it in the clinics,” he says. That led to the creation of an AI-based diagnostics tool that automates the assessment of the more than 500 images that a typical abdominal CT scan generates to help provide early diagnosis of hepatic steatosis, a condition more commonly known as fatty liver disease. Such systems can also help to prioritize which scans a radiologist should review first.

Charles Kahn, physician, professor, and vice chair of radiology at the University of Pennsylvania Perelman School of Medicine adds that being able to take information about a population and see how an individual differs from the rest of the group makes it possible to intervene by catching conditions early.  “That’s precision medicine,” he says. “And this is just a prototype for the things we can do.”

Penn is just one in a class of innovative CIO100 award winning healthcare providers that are pushing boundaries in the digitization of healthcare. Stanford Medicine Children’s Health, the University of Miami Health System, and Atlantic Health have all moved forward with projects in the areas of precision medicine, machine learning, ambient documentation, and more. “From a clinical perspective, we’re seeing advances in radiology, diagnostic services, and pathology,” says Bill Fera, MD and principal who leads the AI practice at Deloitte Consulting.

An AI first at Penn

The AI-based CT scan analysis system is one of the first to be deployed into a clinical practice, in part because research-driven academic medical practices can build and run their own tools without going through the rigorous process that healthcare product manufacturers face to get approval from the FDA. “We can’t give it away or sell it, but we can use it in our practice,” Kahn says.

The system didn’t come together overnight, though. “It was probably two years before the algorithm was truly ready to go into production,” says Donovan Reid, associate director of information services applications at Penn Medicine, and four years went by before the system finally was ready for production last year.

Because the algorithm requires considerable processing resources, the team decided to host it in the cloud. Data is encrypted and sent to the cloud for processing, and the results flow back into the radiology report. To coordinate this, the IT team developed an AI orchestrator, which it plans to offer to other healthcare providers as open source. “Making it available will be impactful for community service hospitals,” says Penn research professor Walter Witschey.

The team faced a few challenges before the system was up and running as well. IT was concerned about the impact of imaging data flows on infrastructure, and compute resources had to be matched to demand for imaging studies at any given moment. “We received a much larger volume than we anticipated,” says applications manager and project chair Ameena Elahi. And the system had to produce results quickly. “Doctors want the interpretation now, not at four in the morning,” she adds.

Surprisingly, the direct cost — outside of labor — has been only about $700 per month. “We got creative and used tools and apps we already had,” Elahi says. So far the system has processed over 6,000 scans, and now the team is planning to expand to include more of the 1.5 million imaging scans the hospital system performs annually. “We’re scaling up the orchestrator to handle many studies of different varieties,” she says.

Ambient documentation at Atlantic Health

When Atlantic Health System examined ways to reduce physician burnout, the IT team focused on optimizing documentation efficiency during patient visits, and reducing what EVP and chief information and digital transformation officer Sunil Dadlani calls “in basket burden.”

Technology was actually contributing to clinician overload. “Electronic health record (EHR) systems, for example, were one of the most common reasons for stress,” says Dadlani. “Physicians spent more time on the EHR than with the patient.” And physicians often take some of that documentation work home in what Dadlani and other healthcare CIOs refer to as “pajama time.” What’s more, as consumers have become more digitally connected, the boundaries between home and work have eroded. “That was one of the key causes of burnout,” he says.

The project started out with small groups with representatives from all stakeholders. After defining the problem with clinicians, the collaboration team laid out a strategy, with physicians involved all the way through. “They even engaged in the technology selection,” Dadlani says. The IT team then deployed an ambient documentation system to 4,800 clinicians.

With ambient documentation, also called ambient listening technology, an AI algorithm listens to a conversation between the clinician and patient, and creates notes in real-time that the clinician can review and save to the patient’s record. “In many cases there’s no editing required,” says Fera at Deloitte.

With the new system, says Dadlani, physicians have more face time with the patient without having to worry about creating the documentation. Between 80 and 90% of physicians now use it — a number, he says, isn’t easy to attain.

Part of the project planning involved making sure the cutting-edge technology would integrate with Atlantic Health’s existing technology footprint, which is four to five years old. “You also must understand where in the workflow you’re going to use these technologies without creating upstream or downstream issues,” Dadlani says. And IT’s responsibility doesn’t end with deployment. “These technologies have a performance drift, so you have to audit and monitor them throughout their lifecycle,” he adds.

To reduce in-basket workload, the team analyzed inbound content and created categories such as prescription refill requests, appointment scheduling, and things that the care team needed to address. It then created a system that assigned the messages to the appropriate people, and applied gen AI to create responses that can be reviewed, edited, and sent by the physician. No response goes out without human review. “That’s extremely important,” Dadlani says. “Today, we’re in the top quadrant in the nation in terms of provider satisfaction and efficiency.”

University of Miami Health System optimizes scheduling

University of Miami SVP and CIDO David Reis faced a different type of time management problem: vacation and other out-of-office time that surgeons booked in the HR system weren’t connected with the scheduling system for the academic medical center’s more than 25 operating rooms.

“There was nothing telling the OR scheduling people that a physician was off,” he says, leaving precious OR time blocks unused. So, based on a suggestion from a physician leader, Reis put together a team to create a proactive machine learning-based system that provides real-time notifications of conflicts between the time blocks assigned in the OR schedules, and the hundreds of surgeons booked into those slots.

The algorithm needed to integrate two separate platforms: Workday on the HR end and the Epic electronic medical records system. “We wanted a solution that was infinitely scalable and highly reliable,” says Ravi Akkiraju, University of Miami Health System’s chief enterprise architect, but he didn’t want to reinvent the wheel. So the team first approached vendors. “Epic didn’t have an API for block schedule time and neither did Workday,” he says. So, working with Workday, it created an API that interacts with Epic’s back-end database and its in-basket functionality.

Getting the new system up and running was time consuming and required quite a bit of manual work. “There are tens of thousands of data tables and hundreds of thousands of data fields in the Epic system,” he says. “We had to determine which of those applied to the schedule we were talking about because it was a novel integration. There was no data dictionary that said, ‘Go here.’”

With the system up and running, the OR was able to fill 265 blocks of time that would otherwise have been idle. “Two-hundred-and-sixty-five patients were able to get surgery sooner,” Reis says, and over 200 hours of additional surgery were able to be completed in the first six months.

Stanford Medicine Children’s Health builds new roadmaps

Stanford’s hospital for children has a large population of cancer patients, but until last year, the protocols, or roadmaps for treatment customized for each patient, were mostly still on paper.

“Our EMR vendor, Epic, didn’t have anything in this space, and we couldn’t find anything that met our needs,” says Stanford’s CIDO Tanya Townsend, so the team built its own integration that links the electronic medical record (EMR) to a Microsoft SharePoint system that hosts patient roadmaps.

The project took eight months to build. But now, she says, a clinician can open up a patient record, and if there’s a roadmap, a banner appears in the patient chart that takes the clinician over to the SharePoint document. Over 500 staff now use the system, which contains over 1,000 patient roadmaps.

Fera expects we’ll see more open EMR system APIs accessible to provider organizations, which will save time and money for these types of projects. “Regulations are pointing us in that direction,” he says. “Cracking open the EMR, that’s where innovation starts. Open APIs will allow innovation to happen outside the EMR, where we have more opportunity for acceleration.”

What’s next for healthcare digitization

Stanford’s Townsend says more innovations are coming. “We’re looking at using AI for clinician efficiencies and revenue cycle automation to reduce denials and for prior authorizations,” she says. “And we’re automating functions using robotic process automation, which we’re finding saves tons of time.”

The team also has an ambient documentation pilot under way, as well as an in-basket pilot project that lets patients communicate with providers, and uses AI to auto-generate possible responses. “In theory the response time and turnaround time should improve,” she says. “With so many opportunities, it’s important for CIOs to make sure they’re working on the right things for the right reasons, understand outcomes at the outset, be aware that not everything will meet expectations, and fail fast.”

Restuccia says Penn is also rolling out more clinical wellness tools, including ambient listening, and is moving forward with an AI-based chart summarization tool. “By the time we see patients, they often come with a variety of different complications from different health systems,” he says. “The chart summarization tool will reduce the burden on the physician.” In terms of patient engagement, he expects more in the areas of patient self-scheduling, evaluation tools, as well as access to clinical trials information, online payments, and better ways to interact with the care teams. He also expects more use of analytics against data gathered from millions of ambulatory visits and thousands of admissions per year.

The use of AI is accelerating at the University of Miami, too, says Reis. “We have over 100 AI projects underway, with about 20 completed.” And 80% of those are for coding and billing — the low-hanging fruit. Plus, gen AI is poised to improve back-office productivity by automating authorizations, adjudications, and the settlement of appeals, Fera adds.

But today, most healthcare innovations still remain in the back office. “The items are more about reducing administration and documentation burdens,” Restuccia says. “But we’re still a little hesitant as to the ability of AI in direct patient care.”

With ambient dictation, says Reis, “we’re inventing better ways for the EMR system to capture the conversation. What we’re not inventing, yet, is a way for the provider to avoid clicking around the health record.” Providers, he says, still have to “click dozens of times to document a patient encounter.”

Medical records systems are one area that still needs to digitize, adds Fera, noting that much of that data is still more electronic than digital. “With digitization the interaction between patient and clinician will change radically,” he says, pointing out that many important advances, such as ambient listening, are happening outside the medical record systems. “To the extent that we can make the electronic medical record a system of record versus a system of work, we can innovate on top of the EMR to make workflows more tailored and intuitive for clinicians,” he says.

The push now is to eliminate work instead of users using technology to do work. “We haven’t gotten to where technology is eliminating work, but we’re on the verge of that with generative AI,” Reis says.