kirill_makarov -

Growth of AI in clinical decision support still slow

The COVID-19 pandemic has given some healthcare technologies a boost. AI in clinical decision support isn't one of them, according to experts.

While the COVID-19 pandemic has triggered growth for some technologies, such as telehealth, it's had less of an effect on others. One such area is AI in clinical settings.

When NewYork-Presbyterian Hospital in New York City first began investing in AI in 2017, it began with easier-to-quantify, lower-risk areas like revenue cycle management and other non-clinical applications, according to Vishal Sheth, the healthcare system's director of transformation.

"We took a deliberate approach of going to that non-clinical area with the expectation that as you start tackling those low-hanging fruits, you can start developing particular portfolios around that with the ultimate goal of going to clinical care as well," he said. "It's been two and a half years and we have a very healthy portfolio of non-clinical projects and we're starting a similar journey now into clinical care, things like predicting patients that are at risk of heart failure."

Even with some of the relaxed healthcare regulations due to COVID-19, NewYork-Presbyterian continues to only slowly introduce AI in clinical decision support, Sheth explained at the recent Ai4 2020 conference. He and other healthcare leaders offered their perspectives on the current state of AI in healthcare and why the pandemic hasn't spurred faster growth of AI in clinical decision support.

AI's current state and limitations

Currently, NewYork-Presbyterian's most significant use case for AI is in the back office within revenue cycle management and the clinical appeals process, which occurs when a patient is discharged and an insurance claim is submitted for reimbursement, Sheth said. He added that, at NewYork-Presbyterian, groups of nurses are responsible for writing the appeal of a denied claim.

Sheth said applying AI specifically to the clinical appeals process provided a clear ROI. To reverse the trend of growing denied claims, he and his team built an AI model that paired claims with nurses in that specialty who would then work with the filed case.

"Our goal was trying to figure out is there a way for us to improve the overall success rate by assigning the right case to the right person," he said. "What was happening traditionally was just first come, first serve and, on average, nurses were doing two to two-and-a-half cases per day. So, if we could have a better optimization of the assignments, perhaps there was a way we could improve the overall success rate."

The AI model had an effect, improving the health system's claims approval success rate by roughly 6%, resulting in millions of reimbursement dollars, he said.

There are just a lot of differences between AI in healthcare versus other industries. Part of that is just the reality that you're dealing with people's lives.
Ylan KaziVice president of data science and machine learning, UnitedHealth Group

While Sheth can quantify the value of AI on the non-clinical side, he said it's not something he can do on the clinical side yet. Sheth said one of the challenges with integrating AI in clinical decision support is jumping through regulatory hoops. But there's also the difficulty of finding a way to explain how the AI algorithms arrive at a recommendation and then remove the ambiguity of those recommendations.

Indeed, Ylan Kazi, vice president of data science and machine learning at UnitedHealth Group, said using AI and machine learning models for clinical decision support is a challenge that requires building trust and maintaining the clinician's role as the final decision-maker when it comes to a patient diagnosis.

"There are just a lot of differences between AI in healthcare versus other industries," he said. "Part of that is just the reality that you're dealing with people's lives. We do have to figure out how do you continue to balance innovation with all of the regulatory pressures that exist."

AI during COVID-19

COVID-19 may not have fast-tracked AI in the clinical setting, but the technology has played an impactful role during the pandemic in areas where it has already proven its value.

Mark Saroufim, a machine learning engineer at Graphcore, which makes AI processors, said the company partnered with Microsoft to build a machine learning model that could analyze chest X-rays and determine whether a patient has COVID-19 or not.

"When you look at these large models that work on very high-density images, they can be slow to train and slow to actually get a prediction," he said. "On this project with Microsoft, we were able to increase our training time by 8x. Whereas a model would take a week to train … now it takes a couple of hours."

But beyond specific use cases such as chest X-rays, AI has struggled to play a more significant role in clinical decision support during the pandemic. Challenges have included changing testing protocols, which results in inconsistent data collection across states, UnitedHealth Group's Kazi said. That creates a nightmare for machine learning, he said.

Anthony Chang, M.D., chief intelligence and innovation officer at Sharon Disney Lund Medical Intelligence and Innovation Institute at the Children's Hospital of Orange County, said the issue with AI in clinical decision support during COVID-19 is the lack of a good public health strategy for data collection, which would require consistent testing and data collection across states. Chang said while the larger public health strategy serves as the armor in the fight against COVID-19, AI acts as the weapon against the virus, and both are needed to successfully manage the pandemic.

"The other problem with COVID-19 as a challenge for AI is that pandemics, just like a lot of biomedical issues, are very complex," he said. "They're highly unpredictable and it's hard to pin it down with data science alone."

Dig Deeper on Healthcare IT systems and applications

Cloud Computing
Mobile Computing