Is the U.S. healthcare industry—both on the provider and payer sides—on the verge of a breakthrough in terms of the leveraging of artificial intelligence (AI), machine learning, and other advanced forms of analytics, to transform care delivery and care management? What seems clear, an expert panel’s members agreed on Tuesday, Aug. 10 at HIMSS21, being held at the Sands Convention Center in Las Vegas, is that analytics work is advancing rapidly, and that the most pioneering healthcare organizations are moving ahead by leaps and bounds.
In a discussion session held on Tuesday afternoon, under the session title “Making Sense of Health Data to Accelerate the Shift From a Reactive to Proactive Healthcare System,” Taha Kass-Hout, M.D., chief medical officer and director of health AI at Amazon Web Services, moderated the panel. His fellow panelists were Bala Hota, M.D., M.P.H., vice president and chief analytics officer at Rush University Medical Center and Rush Health in Chicago; Ashok Chennuru, chief data and analytics officer at the Indianapolis-based Anthem; and Tanuj Gupta, M.D., vice president and physician executive at the Kansas City-based Cerner.
Each of the panelists shared with Kass-Hout and the audience their perspectives on their current work, beginning with Hota. “It’s an exciting time,” Hota said. “My system is Rush in Chicago. Our main anchor hospital, our 600-bed hospital, is Rush University Medical Center. We had typical academic medical center goals prior to COVID. And health equity has always been a main pillar of our strategic plan. And the way I look at digital transformation, it’s in terms of pre- and post-COVID. Pre-COVID, everything had to be driven by the bottom line. There were lots of pilots going on. But a few elements that had traction pre-COVID, included the development of our data warehouse; our use of telemedicine; and our patient portal. But what we learned is, some of these things took huge amounts of time and effort, and not a lot happened. Then COVID happened,” he said.
And, Hota said, “Patients loved telehealth after the pandemic emerged. The whole organization went through an agility transformation. Things that had taken forever to decide, happened very quickly. And much of the pressure fell to the IT staff. And COVID really provoked a data-driven response. Our dashboard around COVID became distributed across the whole system. And the liquidity of data we had wanted became the number-one thing. Now, we’re at the phase where, how much change are our providers able to accept? There’s fatigue with the pace of change.” But that change is happening nevertheless, he said.
“I represent Cerner in our AI and ML solutions,” Gupta said. “I consider us part of the digital world already, but there’s a transformation happening with health IT. Typically for an EHR [electronic health record] company, we’re trying to help providers capture data, summarize it quickly; coordinate care; and secure reimbursement.” With regard to the tools involved, he said, “For capturing data, it’s been dictation, templates, and macros; for summarizing information, it’s been tabs and sections, tables, in the UI [user interface]; for care coordination, it’s largely been workflows, order sets; for securing reimbursement, it’s been revenue cycle management tools; for capturing data, it’s analytics. And in capturing data, we’re seeing the rise of ambient note taking. And voice is becoming a big play” in clinical documentation. Meanwhile, he added, “We’re seeing robotic process automation both with care coordination and with revenue cycle management. And in critical decision-making, the same thing. So the jobs are the same, the tools are evolving. So you’re going to see a lot of the HIT companies evolve to bring you those new tools.”
“I work for Anthem,” Chennuru said. “We serve over 115 million consumers, including 24 million in our health plan. So we have a lot of data, structured and unstructured. And we’ve embraced a digital-first strategy, and are developing a data platform that leverages the power of AI, and being able to translate that data into proactive and predictive insights. We accelerated the digital adoption during COVID,” he noted. “Over 15 million of our members are on a digital platform. In fact, 66 percent of our interactions with consumers are on a digital platform,” with virtual care encouraged where appropriate.
“And the power of the data we have comes from a combination of from claims, eligibility, provider data, and marketing, and getting the medical records from the providers we partner with, and combining that with emerging data sets, whether it’s longitudinal data or other types,” Chennuru. “So how do we leverage the data, and share that with our care team working with our members and with our providers in value-based care? Providing them with timely gaps in care. A PCP [primary care physician] might not know about a hospitalization, but we’ll share that with them, so they can focus on the post-discharge care. We’re also leveraging AI to support members who need care management. Our focus is on how to translate data into meaningful encounters.”
Looking at the process issues that complicate the use of analytics
After those initial descriptions were shared by the panelists, Kass-Hout led his fellow panelists in a nuanced discussion of the complexities involved in developing and leveraging advanced analytics, including AI and machine learning. What does “learning by doing” mean, in this context, Kass-Hout asked the panel.
Rush Health’s Hota testified that “There’s a lot of skepticism at the health system level about innovation that goes a little bit too far beyond where we are currently. And for health systems, if something’s not ready to deploy, it consumes time from people who really just want to see patients. So getting useful tools to people is the key,” he said. “And people are frustrated with clunky user interfaces. People really want usable interfaces. Lessons we’ve learned: where you can, look for things that have a good, solid mature back end or system in place, so we can work on the workflows. And what’s so satisfying is that tools that used to require retooling or customization—there’s a lot of tooling that we don’t have to do. So, making sure it fits into the workflow, is important.”
Meanwhile, said Cerner’s Gupta, “It’s not as simple as developing the model and letting it go. There’s a life cycle to the models. You have to touch the models and calibrate them a lot more frequently. And for eight years, we had developed a library of 50 clinical models; and we deployed three or four; the deployment process took a long time. In the last year, we’ve started shifting our interest to data ops and dev ops people. We worked on two models, 35 times; that was a 10X difference for us. And we’re on track to develop 10 models 50 times. So we’re on the verge of solving that development problem. And it’s not as simple as developing 10 different models to a solution.” Indeed, he said, quoting a recent industry analysis, it takes about $100,000 to build a data analytics model for the first time, while maintaining an analytics model costs about $15,000 per client per year on average. “So you can imagine the cost,” he said. “That was the barrier and learning for us, that distribution, deployment and monitoring, is a much more challenging problem for us than it is to create the next solution. The monitoring part is where a lot of innovation is going on. It’s not just the model drift; we’re getting questions about gender and racial bias; there are data breaches. So the monitoring of these models becomes a big deal.”
“We have over 900 data scientists and AI engineers,” Anthem’s Chennuru noted. “So pushing data to the cloud, to develop actionable insights, is key. But how do you create all these models and integrate them into the workflow? We operationalize a model and put it into the consumer workflow—and we’ve learned that it’s not really the outcomes it’s generating. But we’ll quickly pivot from that learning and deploy. Also, when we are leveraging data to automate prior authorization. How do you look for unconscious bias in race and ethnicity data, for instance? Being able to explain the models to regulators, to auditors. So as we get more and more data and mine it, we’re able to capture real information; and you want to explain it, and it becomes challenging. So we talk about trust AI, explainable AI. The other aspect of putting data into the cloud is for security. We have data for over 200 million lives. We’re talking about large data—we’re talking about one of the largest data sets in the industry. How do you perform analytics without issues? So we’re partnering with you. And how do we layer in SDOH [social determinants of health] data, genomic data? The complexity keeps getting multiplied. But the key in all this is focusing on things that matter, and being able to get to improved outcomes. And outcomes are really valuable when they come through the workflow.”
At Rush Health, Hota said, “Coming to COVID, we had gone through some attempts to bring AI and ML into our clinical workflows. They were sort of helpful. But we found, four years ago—the hunger was really for descriptive analytics and dashboards just to help the business. And again, COVID changed the whole expectations and structure. Simple predictive models to help us anticipate bed capacity, were a massive hit. They helped us anticipate when we had to open up beds. And the forecast models were built into our Epic HER and our command center. So now we’re trying to build forecast models into our revenue cycle management processes. We’ve always had targets, but what’s it like to be able to forecast revenue? It’s almost like a weather report. And we’re anticipating rolling that out to value-based care, and how are we doing in terms of PMPM costs? I’m excited by the project we’re doing with Amazon.”
What’s more, Hota said, “One of the things we have at Rush in our ED is detailed screening for health equity needs, SDOH: legal issues, food insecurity, housing insecurity, and other issues. Plus, we have a closed-loop ordering system to order referrals to community resources. But we can’t really scale this to everybody. How do you identify the need across your system without having to go and ask questions of the triage people? We can get to about 85 percent through structured data, but when you add in notes, you can get to 90 percent. What excites me is that some things that used to need to be custom-build are now a commodity. And where I think this is going is risk adjustment. The federal government does risk-based models, insurers do. But being able to bring in additional information, like SDOH, into risk models, really helps a lot.”
Amazon’s Kass-Hout noted that “We’re developing machine learning tools,” involving “purpose-built services, trained around a wide variety of examples. A couple of weeks ago, we launched Amazon Health Link, to help you analyze, structure, and aggregate a patient’s entire history, to extract indications like diagnoses, tests, procedures; and using FHIR standard to aggregate the data through APIs. You can integrate interrogated data, to develop predictive models for population health, predictive analytics. In the middle of the stack is Amazon Sage Maker, to help you deploy models at scale in the cloud.”
In terms of people and process issues, Rush Health’s Hota reported that “We have a data science team. We’re trying to figure out how we can reduce time to market for these tools. By the time you implement your model has changed, or it may not be good enough for a clinical setting, or there’s an explainability issue. So reducing the time to deployment has been a very good thing for us.”
Looking forward into the future, Anthem’s Chennuru said that “It starts with the longitudinal view of the member: through claims data, EHR data, etc. And being able to generate precision insights, based on rich data, and being able to hand that to case managers. With such a large database, we can do analytics. It’s all about the data-driven insights. The other thing involves providing the right care: how do you match the member to the right provider? Rather than making it just code-based, we’ve factored in, if you have comorbid issues, who is the right provider? And the more data we have, the more we learn. Also, per administrative burden,” he reported that he and his colleagues have been doing work around improving prior authorization processes, to benefit both providers and consumers. “When we’ve been able to get the right data, we can provide an instant response on authorization,” he said. “That helps the provider as well. So, improving the consumer experience, cost, and administrative burden issue—all are improved.”
