Session participants to examine potential of artificial intelligence in stroke care
The technology shows promise, but many hurdles remain.
When it comes to artificial intelligence managing practices and patient care, the stroke sector isn’t much different than any other medical sector: The potential is there, but so, too, are roadblocks to maximizing the tool.
For stroke care professionals, the biggest impact artificial intelligence has had so far is in the area of imaging. David Liebeskind, MD, FAHA, professor of neurology in the UCLA Department of Neurology and director of the UCLA Stroke Center, said the stroke field has been using automated imaging for close to 20 years to create instantaneous and real-time imaging reports. Those results, however, were not based on artificial intelligence, but rather used as a fast-processing system without a machine learning algorithm.
“But that has changed as certain groups began to use machine learning for select components of the imaging,” he said. “As an example, on perfusion imaging, we’ve been able to do the selection of the arterial inflow function using machine learning, so you don’t have to manually look at where the blood flow is coming into the brain, which is a requirement to process the perfusion images.”
Dr. Liebeskind, along with Oana Dumitrascu, MD, MSc, associate professor of neurology and co-director of the Neurology Artificial Intelligence Program at Mayo Clinic College of Medicine in Scottsdale, Arizona, will lead a discussion on the current uses of AI and its potential in Wednesday’s session, Artificial Intelligence and Big Data in Stroke Care: From Hype to Reality.
Dr. Dumitrascu said artificial intelligence in stroke care currently serves more of a support function.
“We use AI in stroke as a decision support tool to predict outcome based on imaging and clinical features,” she said. “The way we work as humans, we use imaging interpretation and clinical assessment in order to guide our diagnosis and treatment decisions. AI, though, uses automated imaging and clinical feature identifications.”
But there is tremendous potential beyond that. Dr. Dumitrascu said she envisions a future in which AI can help underserved areas that do not have access to advanced imaging or stroke specialists.
“As an example, in order to determine patients who are candidates for mechanical thrombectomy in late windows, we are using computed tomography (CT) perfusion,” she said. “There are developing AI tools that automatically detect the core and the penumbra and are not using CT perfusion. They are only based on the point of care head CT that is available in most hospital settings.
“CT perfusion is not available in many rural hospitals. It is more expensive, it requires more intravenous contrast and more radiation. If we can alleviate all of these downsides by using an accurate AI software that can predict the core and the penumbra from the plain head CT, then we can apply it both in pre-hospital settings as well as in under-resourced rural hospitals that don’t have advanced imaging technology. This way, it can increase the utilization of AI across geographical disparities; it’s going to improve access to treatment and improve health equity.”
Dr. Liebeskind said he sees AI and machine learning potential in a couple of other areas more closely related to the day-to-day clinical practice of the stroke specialist.
“One is using it in medical records in terms of large language models,” he said. “It will undoubtedly be useful there in the future, but how it’s integrated into our electronic health records is still an open question at this point. It’s not one clear data set that the algorithm has access to. How it’s organized and how it’s deployed is a question. The other iteration is the communication aspects of lab results and other features to help accelerate what we normally do.”
Dr. Dumitrascu said speech recognition software is already being developed that would automatically create a note while the patient is being seen in telestroke services.
“That will save us a lot of time, and we can move forward to the next acute care patient,” she said. “AI will enable our documentation and will enhance our workflow by providing quality data collection and feedback.”
Both doctors said outpatient monitoring is another potential use for AI.
“This would involve using data that’s available to the outpatient in the form of wearables and personal health-related devices, even from your mobile phone,” Dr. Liebeskind said.
“There is a lot of work in progress (for remote care using AI), but nothing is available yet for us to use,” Dr. Dumitrascu said. “One of the promising tools is a natural language processing tool — AI can read through the entire electronic medical record and can identify what clinicians need to make a decision: vital signs, past history, details that would impact management. It can automatically calculate scores for us using certain parameters.”
However, AI needs to overcome many hurdles before it can get to the point of everyday use, not the least of which are the legal implications. Although the FDA approved some automated imaging and computer-assisted diagnosis techniques, the agency issued a warning letter in April 2022 that none of these methods should be used to replace human interpretation of the scans.
“Subsequently, the use of machine learning has evolved to other aspects of stroke imaging,” Dr. Liebeskind said. “We’ve developed our own platform to do true machine learning for every component of stroke imaging.”
Then there is the question of liability — who is medically and legally liable if the AI tool is wrong about a diagnosis or treatment? The answer as of now is the stroke neurologist or the radiologist.
Dr. Dumitrascu said there is still a long way to go and much work to be done before AI can be a useful part of a physician’s daily practice across various practice settings.
“All of these AI interventions that we’re all hoping to have one day in our routine practice need to undergo rigorous, unbiased and prospective evaluations to demonstrate their true impact on stroke outcomes,” she said. “Are they really helping us, or are they creating more false positives or false negative scenarios in our services? I don’t think we have that answer yet.”
Dr. Liebeskind agreed, adding there are still many questions to answer.
“How is machine learning improving our lives and our daily workflow in the care of the stroke patient in front of us?” he said. “It varies — in some situations, it’s extremely helpful because it’s faster and computer-generated, but we have to be careful. It’s enticing, it’s intriguing, but it all depends on the details in terms of how it’s actually deployed and used on a daily basis.”