In the world of clinical AI, there is imaging and there is everything else.
Researchers in Future Healthcare Journal say that while there are claims from organizations that they’ve developed an AI solution that can “diagnose and treat a disease with equal or greater accuracy than human clinicians,” most of them are based on radiological image analysis. A survey of health care leaders found that while 58 percent are using AI in clinical practice, the most common applications are all related to imaging. In fact, imaging apps approved by the FDA were primarily for the field of radiology. In the Department of Defense, which has invested significantly in health care AI-related applications, the agency has created a massive imaging portfolio.
In other words, the rest of clinical medicine has a way to go before catching up to imaging-based AI applications. “You don’t see that level of accuracy in some of the other applications of AI yet and so I think it sort of inflates expectations when you see it get knocked out of the park in a couple of specific areas,” says Michael Matheny, MD, Co-Director Center for Improving the Public’s Health through Informatics, and Associate Professor in the Departments of Biomedical Informatics, Medicine, and Biostatistics at Vanderbilt University Medical Center.
“The problem that we see right now in the health care AI space is that there’s a ton of remarkable work going on in the basic science space, where models are becoming increasingly sophisticated, increasingly impressive and powerful, but when you take a look at what’s happening on the front lines with patients, providers and health systems, very little of that is actually being used at the bedside for the benefit of patients, providers, and society,” says Steven Lin, MD, Founder and Executive Director of the Stanford Healthcare AI Applied Research Team (HEA3RT).
But that doesn’t mean health care executives are dismissive of other AI applications in clinical care. In fact, 93 percent of respondents in the survey of health care leaders above say AI is essential to their strategy. Another survey, from Optum, found that while only 20 percent are in late stages of deploying AI, a whopping 98 percent of health care leaders either have an AI strategy in place or are planning to create one. Nearly 40 percent are interested in using AI to accelerate research for new therapeutic or clinical discoveries.
In part one of our two-part series on clinical AI, Health Evolution looked at the barriers preventing wider adoption of AI in clinical settings. In part two, we examine the most promising clinical areas for AI usage that go beyond imaging.
Lin’s HEA3RT team is undergoing a number of projects with some of the top names in AI including Google, DeepScribe, Amazon and others. There are AI projects around improving clinical diagnoses, population health, optimizing inpatient care, increasing education around health policy and equity, and using voice-enabled applications. He is excited, in particular, about a couple of projects that use AI to create risk assessments for primary care patients to prevent hospitalizations.
“We are taking the de-identified version of these records and then building a machine learning algorithm in order to identify patients by risk and see if we can predict their probability of ending up in the emergency room or hospital for the next 12 months,” Lin says. “This is important because the only way we currently know if a patient is at risk for the ED or hospital is if they are in the ED or hospital.”
The idea behind using AI for risk assessments is to build a prospective predictive algorithm without relying on claims data, which will always be retrospective, Lin says. Using the machine learning algorithm can help primary care physicians be more reactive in helping treat their patient base. Risk prediction, population health management, and clinical decision making are three of the ten areas where primary care doctors could benefit from AI usage, Lin wrote in an article for Journal of General Internal Medicine.
John Halamka, MD, President, Mayo Clinic Platform, says that when you see the hundreds of places in medicine where you could apply AI, a few areas rise to the top. Considering his wife is a cancer survivor, Halamka is bullish on oncology as one of these leading spots to use AI.
“There’s clearly a need for being able to predict who will develop cancer to help them avoid cancer. Or once a diagnosis is made, what treatment will be best for their care,” Halamka says. “The increasing use of multi-modal information—phenotype, genotype, exposure information—coming together in AI algorithms, we’re seeing a lot of focus on that. It’s happening both in academia and in the industry.”
He is excited about the use of AI within radiation oncology as well. He says that typical radiology treatments require a lot of work to prevent nerve or artery damage. At Mayo Clinic, Halamka says the organization has developed a series of algorithms that can reduce the time needed to build the three-dimensional model needed to administer the radiation.
John Halamka, MD, Mayo Clinic Platforms
Strokes and neurological care
Greg Albers, MD and Roland Bammer, MD started a company called RapidAI in 2008 to improve care for patients who suffer from a stroke. RapidAI uses AI-based image processing technology to help a neuroradiologist remotely and quickly identify when a stroke patient is suffering from a brain hemorrhage or if there’s a large vessel inclusion. Beyond the current uses, he sees promise for using AI to help clinicians treat complex medical issues, like strokes and neurological diseases.
“Ten years ago, there were no AI products for neurological diseases. The technology has moved along quickly in the last five years, where people have learned that you can train these neural networks relatively easily,” says Albers, who is the director of the Director Stanford Stroke Center. “Just like you can train AI to recognize the differences of cats and dogs, you can train it to recognize a brain hemorrhage from a non-brain hemorrhage.”
Like with primary care risk-based assessments, there may also be opportunities to use AI in preventive care when it comes to strokes and neurological diseases. Researchers from the China National Clinical Research Center for Neurological Disease in Beijing published a study in Stroke that says that AI can be used for early intervention to reduce stroke risk.
Other clinical areas
Halamka says that cardiology-based AI applications have been a focus at Mayo Clinic. The organization spun out a company, Anumana, which uses cardiology algorithms that can measure ejection fraction to diagnose diseases based on data from remote monitoring devices. Anumana uses neural network algorithms based on billions of relevant pieces of heart health data in Mayo’s Clinical Data Analytics Platform, for this early detection capability. It’s something Halamka sees happening more frequently, across different specialties.
“As we look forward to delivering democratized access to specialty care to more people in more places, this model of being able to ingest data from a sensor you wear, [connect it to] an AI algorithm and produce a result, will become more common,” Halamka says. “Once you have an AI algorithm that says, ‘This patient has a weak heart pump’ or ‘This patient is about to develop afib,’ then the human decides what action to take.”
Not surprisingly, in the wake of the COVID-19 pandemic, many are looking at AI as a way to control and reduce the impact of infectious diseases. There has been a lot of research on using algorithms to predict COVID-19 detection, although researchers from the University of Cambridge in the U.K. said they weren’t all that effective due to biased data.
What has been effective is the ability to detect sepsis through AI, according to Suchi Saria, Founder and CEO of Bayesian Health, an AI-based clinical decision support platform, and John C. Malone Endowed Chair and Director of Machine Learning and Healthcare Lab at Johns Hopkins university. Bayesian Health’s platform, in fact, resulted in patients receiving faster life-saving treatment by an average of 1.85 hours.
“In this study, we pointed the platform towards the early detection of patients at risk for sepsis. We applied it in the ED, the ICU, across clinical floors in five different hospitals…and this was at a high-performing system like Johns Hopkins, which has baseline outcomes that were already pretty good. We were able to show high sensitivity, precision, and significant early detection rates. When the tool was applied prospectively, patients received antibiotics significantly earlier with significant associated reductions in length of stay and mortality. Most interesting part of the study was that although it was implemented as a passive flag and the use was not mandatory, we saw provider adoption around ~90 percent sustained over two years. Most CDS tools struggle with provider adoption,” Saria says.
5-10 years down the road
The barriers that are preventing widespread clinical adoption of AI, as reported in part one of this series, have not gone away. There is a reason that only 20 percent of leaders surveyed above are in the latter stage of AI deployment.
However, for a number of health care AI experts, such as Matheny, many clinical AI uses are ready for primetime as it stands today. These are primarily imaging-based diagnostic and cognitive support tools. Five years down the line though, he says there will be more AI usage in clinical care across the board. He breaks it down into three categories.
“You have the AI imaging informatics side, which is mature and will be relatively ubiquitous in five years. You have the autonomous AI that will slowly grow through fits and starts. You may see some high-profile missteps in that area. Then you have the third bucket, the clinical support AI, which I see expanding rapidly as the industry learns how to successfully integrate it into a clinician’s work flow,” Matheny says.
Halamka is bullish on the impact of AI in the next five years as well.
“We are going to see the explosion of organizations offering various kinds of AI services,” he says. “Using a platform approach, you should be able to connect any data source with any provider of algorithms to create an ecosystem where people with a given disease state or concern can get guidance.”