Despite Slow Adoption, Artificial Intelligence Pilot Programs Yield Practical Results

Artificial intelligence (AI) has made inroads in many industries—banking, finance, security—but its adoption in healthcare has been lagging and real-world clinical implementation has yet to become a reality. Nonetheless, proponents say it is only a matter of time and pilot programs are starting to yield some practical results.

Two artificial intelligence pilot programs launched by Northwest Medical Specialties in Washington and The Center for Cancer and Blood Disorders in Texas were designed to track high-risk factors for patients. These factors included mortality, pain management, depression, deterioration, avoidable admissions, emergency department visits, and readmissions. The pilot program was launched in July 2017 and was completed in September 2018 for both sites.

Findings “identified patients who could experience up to a 30% reduction in loss of function and activities of daily living (deterioration), increased the diagnosis of depression by 22%, and reduced moderate and severe pain by 33%,” said Amy Ellis, director of quality and value-based care at Northwest Medical Specialties, during a panel discussion at the Community Oncology Alliance’s 2019 Community Oncology Conference.

The practice also reported a 225% increase in hospice referrals per 1000 patients/month and a 35.3% increase in palliative care referrals and supportive care consults per 1000 patients/month.

At The Center for Cancer and Blood Disorders, similar impacts to patient care were also reported.

“We saw a 17% reduction in loss of function, a 33% increase in depression diagnoses, and a 28% reduction in moderate and severe pain,” said Ray Page, DO, PhD, the president and director of research. Hospice referrals increased 113.3% per 1000 patients/month and palliative care referrals were up 218.8% per 1000 patients/month.

Both practices also reported decreases in depression among patients.

In general, the machine would integrate data about the patient that were collected from the oncology practice and findings from the clinical workflow and examination. Those data were processed and integrated into the patient’s electronic medical record, which the oncologist would review and refer to during the patient’s visit. Findings from the visit would then be input into the machine, synthesized, and conclusions drawn about the findings were shared with the oncologist after the visit. If any high-risk factors were identified, an alert was sent to the patient’s care team for follow up.

AI applies to computers that perform tasks that are assumed to require human intelligence. It represents a collection of technologies that enable machines to comprehend, act, and learn so they can perform administrative and clinical healthcare functions.

Legacy technologies only use algorithms and tools to complement human understanding, but AI in healthcare can augment human activity said John Frownfelter, MD, chief medical information officer at Jvion, Inc, a technology company based in Johns Creek, GA.

“First of all, AI is a concept of learning. If we’re using AI, in some way or in some form, it is learning from the data that are feeding into the ‘machine,’” said Frownfelter. “The second concept about AI is that it is using large amounts of data. It will take large amounts of data and learn and draw associations from pieces of information in the data that we [as humans] may not even take into account.”

Frownfelter also addressed the reason the adoption of AI has been lagging in healthcare. In medicine, “we have been trained to try to understand fully, to be able to touch and feel the results. We need to dissect and understand where the flaws in logic might be.” This is an approach that helps protect the patient. However, AI does not explain how it reaches its conclusions, about how it got the answer.

He also pointed out 2 examples of what is not AI. Merely following guidelines or noting when deviation in guidelines occurs is not AI, he said. “Physicians who come to a point in care and can choose option A or option B and who then refer to guidelines are not using AI, they are following an algorithm,” Frownfelter said.

AI can provide more than just predictions, noted Frownfelter. AI requires “deep-machine learning and applying [the learnings] to the next patient in a way to anticipate not only what will happen but also to give insights. AI helps to answer ‘What are the interventions that are most likely to be impactful to the patient?’”

An advantage that community oncology practices have over academic centers when implementing AI pilot programs is that these practices are more agile, said Aaron Lyss, MBA, director of strategy & business development, Tennessee Oncology. This agility and adaptability allow oncology practices to determine where the best return on investment (ROI) lies.

“To the extent that we can find clinical and operational workflow efficiency, that’s where we can see ROI in the near term,” Lyss said. “One example is trying to automate prior authorization in a way that doesn’t require manual human intervention.”