By Christina Bennett, MS

The demand for data analytics is at an all-time high in oncology care, and medicine as a whole. Faced with an ever-increasing pressure to control costs and deliver higher quality care in the era of value-based reimbursement, oncology practices are demanding deeper knowledge and insights into how best to care for the right patients, at the right time. Practices and payers are asking questions that reflect their value-based care reporting requirements, such as how to reduce costs, avoid unnecessary care, meet the needs of a diverse population of patients, and determine whether interventions are actually working. Emerging as a potential answer to some of these questions is artificial intelligence (AI).

AI has shown utility across several industries—banking, finance, security—and has different applications being explored in medicine, one of which is clinical prescriptive analytics. Clinical prescriptive analytics involves helping guide decision-making, such as identifying high-risk patients so that care teams can intervene before an adverse outcome occurs. Furthermore, clinical prescriptive analytics is more than establishing a risk prediction for a patient—it involves applying AI to the next patient in a way that anticipates not only what will happen but also to provide insights into interventions that are most likely to be impactful.

Entities are already recognizing the potential AI has to improve the delivery of care. Specifically, earlier this year in March the Centers for Medicare & Medicaid Services (CMS) in collaboration with the American Academy of Family Physicians and the Laura and John Arnold Foundation announced the Artificial Intelligence Health Outcomes Challenge. The challenge aims to find AI solutions that can improve the quality of care by better predicting health outcomes, such as hospital admissions. In addition, the payment model proposal for Making Accountable Sustainable Oncology Networks (MASON) described in depth an AI solution targeted at benchmarks for clinical quality and cost-related outcomes. It was recently approved by the physician-focused technical advisory panel for CMS.

“This is only the beginning,” said Barry Russo, Chief Executive Officer, The Center for Cancer & Blood Disorders. “AI is something that we hear a lot about in our practice and we’re really excited about the opportunities that AI presents.”

AI Gets a Test Run

The utility of AI in oncology was recently demonstrated in a pilot study between Cardinal Health, Inc. and Jvion, Inc. in which The Center for Cancer & Blood Disorders, Tennessee Oncology, and Northwest Medical Specialties incorporated an AI tool from Jvion into their clinical workflows. The community oncology practices’ experiences with the AI tool were discussed during a panel session at the 2019 Community Oncology Alliance (COA) conference.

During the pilot study, practices identified high-risk patients for seven possible adverse outcomes, including 30-day mortality, 30-day pain management, 6-month depression, 6-month deterioration, 30-day avoidable admission, 30-day emergency department visit, and hospital re-admission.

Amy Ellis, Director, Quality & Value Based Care, Northwest Medical Specialties, PLLC shared the changes the practice saw, which overall were improvements across all adverse outcomes assessed. In particular, stark improvement was seen in the mortality metric for hospice and palliative referrals. The practice had a 225 percent increase in hospice referrals and a 35 percent increase in palliative care referrals. Ms. Ellis noted that Northwest participates in the Oncology Care Model (OCM) and has a “pretty big” emphasis on end of life care for patients.

“The most surprising piece for me was the mortality dashboard,” said Ms. Ellis. She recalled that during the first initial data release, she went through the patient list and found that it was “very” accurate. She said there were patients on the list that had within the last couple of weeks already passed away or were recently referred to or in hospice. “That’s actually when I became a believer and a driver of this tool for value-based care—because I saw that first hand,” she said.

The underlying benefit of piloting AI is that it surfaced up patients that physicians didn’t anticipate being high risk, allowing care team members to intervene and potentially mitigate or reverse the adverse outcomes.

“There may be patients that fall into that risk group that you didn’t predict,” said Ray Page, DO, PhD, President & Director of Research, The Center for Cancer & Blood Disorders. “In that situation you really want to get the caseworker and look at those people a little more closely to make sure that they’re doing okay.”

For example, one of Dr. Page’s patients with metastatic lung cancer was brought to his attention via the AI tool. “We’re predicting that he’s going to die within the next 30 days,” said Dr. Page. The patient shows up on several of the adverse outcomes assessed and has a history of a brain metastasis as well as poor kidney function. “He’s got a lot of issues going on, but when I look at him in the clinic, I don’t think he’s going to die within 30 days—I think he’s going to do well,” said Dr. Page. Usually Dr. Page doesn’t discuss the AI tool with patients because it hasn’t been validated, but for this situation he decided to inform the patient of the prediction. Dr. Page told him, “I want us to work together to make sure that [prediction] doesn’t happen.”

Given the prediction, the patient is being monitored more closely and caseworkers are calling him after each chemotherapy treatment. The patient also now uses a cane because of mobility problems and he’s undergoing rehabilitation to regain strength. “He’s being touched by somebody in the clinic virtually every day of the week,” said Dr. Page.

“All this is still very early, but I think it definitely is a very interesting tool,” said Dr. Page. He said that the tool is not going to replace doctors. “It gives us pause to sit there and think, ‘Well, what’s going on with this patient? What can I do to make a difference?’”

With AI Come Challenges

The pilot study also revealed challenges to incorporating AI into a practice. For instance, Dr. Page highlighted concerns among physicians about breaching patient privacy (given that AI uses patient data) and adding to the already high administrative burden of an electronic health record. To avoid creating more administrative work, care team members were given authority to identify high-risk patients and address their issues, without the doctors having to triage and “make more clicks on the computer.”

Getting buy-in from physicians and care team members who may not understand what AI is can also be a challenge. “In the beginning we were all skeptical a little bit, until I started to manually validate [the data],” Ms. Ellis said. “Now [care team members] are very open to it and they stand by me if I want to make a change.” Even with provider buy-in, however, knowing how to intervene can be still be difficult. Ms. Ellis explained that, for example, AI may indicate who is at risk for a hospital re-admission, but “what you do with that information—that’s a learning curve.”

Beyond that are the pragmatic difficulties of implementing any new initiative into a practice. Aaron Lyss, MBA, Director, Strategy & Business Development, Tennessee Oncology, explained that Tennessee Oncology has implemented a myriad of practice changes during the past several years, and that it was difficult to precisely match outcomes seen during the pilot to which practice changes were implemented.

“We’re never working with sort of a clean laboratory environment in which we’re testing a lot of these new initiatives,” said Mr. Lyss. He stressed that community practices have to pilot in a way that’s “very agile and adaptable” and figure out whether a new initiative, such as AI, ultimately generates a return on investment (ROI). If a new initiative does not meet its goal, the consequences can be a costly; for example, the failed piloting of IBM Watson at MD Anderson Cancer Center cost the institution more than $62 million.

“To the extent that we can find clinical and operational workflow efficiency, that’s where you’re going to see ROI that’s measurable in the near term,” Mr. Lyss said. For example, automation of prior authorization and various payer reporting requirements that don’t require manual human intervention may generate a measurable ROI.

Changing Culture to Build Trust

Compared with other industries, healthcare has been slow to adopt AI and the reason for this is largely the culture of medicine. “The way we have been trained to think in medicine is that we need to be able to understand fully and we need to be able to touch and feel the results,” said John Frownfelter, MD, FACP, Chief Medical Information Officer, Jvion. With AI, physicians cannot see a direct explanation of the result like they would in a journal article. AI takes large amounts of data to go beyond what physicians already know, and learns and draws associations from pieces of information that physicians may not have even considered.

In fact, Robert Pearl, MD, former CEO of The Permanente Medical Group, wrote that the biggest barrier to AI is medicine’s culture. What Dr. Pearl meant by this, Dr. Frownfelter explained, is physicians have a wrestling moment to say, ‘I may not know everything there is about my patient, that I am learning something from the insights that the AI is providing that I didn’t think of, and how do I begin to trust that.”

© Caribou Publishing. All rights reserved. Reproduction in whole or in part is prohibited.