As artificial intelligence (AI) use continues to grow in nearly every industry, it is important to establish guardrails to make sure the technology is used ethically and responsibly. This is especially true in the field of medicine, where errors can be a matter of life and death and patient information must be protected.
A group of stroke physicians, researchers and industry representatives discussed the current use and future of AI in stroke clinical trial design at the Stroke Treatment Academic Industry Roundtable meeting March 28. Led by the University of Cincinnati's Joseph Broderick, MD, the researchers published an article in the journal Stroke Sept. 30 summarizing the group's discussion.
Stroke physicians already use AI to aid clinical decision-making, particularly when analyzing brain and vessel imaging. It also alerts physicians about potential participants for clinical trials.
But with these and other expanded uses of AI, Broderick and his colleagues emphasized the importance of designing "human in the loop" systems that require human input and expertise in the training and use of AI models.
"Think about AI like a toddler learning to ride a bike," said Broderick, professor in UC's College of Medicine, senior adviser at the UC Gardner Neuroscience Institute and director of the NIH StrokeNet National Coordinating Center. "It is an amazing feat to ride a bike, but there are a lot of falls (mistakes) in the learning. Having an expert, and even training wheels, to help support the bike while the child is learning is helpful. Eventually children do learn to ride the bike very well."
Broderick and his colleagues compared the usage of machine learning (ML) with generative AI in stroke applications.
Machine learning trains AI models on a structured and human-curated dataset to classify or predict outcomes known as the "ground truth." While it takes more human effort to train these models with large data sets, most machine learning can be done efficiently with standard computing power.
"A major advantage of these ML models is that their methods are generally more interpretable and their decision-making process more transparent, so they can be understood and traced, which is critical for medical validation and biological plausibility," the coauthors wrote.
Generative AI is trained on massive, unlabeled bodies of text from the internet, books and periodicals before being fine-tuned on more specialized sets of data. This generally means less human intervention in training the model, but it requires massive computing power and electricity.
"The (generative AI) models themselves have billions or trillions of parameters, but they operate as a 'black box,' making it difficult to fully understand how or why a specific output was generated," the coauthors said. "Explainability of large language models is an active area of research."
Whether using machine learning or generative models, stroke researchers will need to be proactive in making sure data sets are robust and account for data from different scanner manufacturers, institutions and patients to improve generalizability.
"If we use bad or limited data and human experts don't correct the bad data or classifications, AI can produce inaccurate and wrong recommendations," Broderick said. "My biggest concern is when AI is trained on bad data and gives answers that can harm."
Researchers will also need to develop strict protocols and safeguards to keep patient information used to train the models private and HIPAA compliant. This could look like independent third parties such as the American Heart Association centrally collecting anonymized patient data before it is fed to AI models, or training models with data only from each individual institution before sharing the learned parameters more broadly.
"Protection of patient privacy represents a major challenge to the use of clinical data for training AI in healthcare, and sharing of even de-identified data between countries is made more challenging by different laws regarding data sharing in various countries," the coauthors wrote. "New methods of model development hold promise to address some of these privacy concerns."
After robust stroke AI models are developed and validated by humans, Broderick said potential applications include better identification of potential trial participants, communicating trial designs to patients in lay language, translating trial information into different languages for non-English speaking patients and helping identify the best treatment for each individual patient.
"We have been talking about precision medicine for some time, but AI is a major step forward to accomplish this," he said.
In addition to AI, the authors discussed new clinical trial designs, such as platform trials, which can more efficiently test several research questions at once and add new questions as older questions are answered. Another major focus going forward is pragmatic trials, which aim to assess the effectiveness of treatments when they are implemented into routine clinical care rather than under idealized conditions.
By comparing existing treatments, embedding trial procedures into normal clinical workflows and using data from the electronic health record, researchers and organizations can lower the costs associated with these types of pragmatic trials and simplify their infrastructure. Pragmatic designs hopefully increase the chances that a trial is accomplished successfully, timely and inexpensively.
Finally, the stroke research community needs more community and patient engagement. This should include input from the boots-on-the-ground medical personnel (EMTs, physicians at transferring and receiving facilities, and study coordinators) who enroll and treat stroke patients in clinical trials.
Common goals for a trial should be established to minimize patient and investigator burden in trial participation, extend trial participation to community-based settings whenever possible, and quickly disseminate trial results to patients, clinicians and the public.
"The future is bright, and we will make great progress in research with these new tools," Broderick said. "At the same time, the real test of our current age with the rapid expansion of AI into our daily lives is recognizing accurate data and truth amid a sea of words, images and videos that can be wrong, harmful or inaccurate."
"Fire can burn down a house as easily as it warms the body or cooks a meal," he continued. "AI is a fire that is rapidly spreading, but we are just beginning to learn how best to use it safely and wisely."
Other article coauthors include UC's Eva Mistry and Paul Wechsler, Mitchell S. V. Elkind, David S. Liebeskind, George Harston, Jake Wolenberg, Jennifer A. Frontera, W. Taylor Kimberly, Christopher G. Favilla, Johannes Boltze, Johanna Ospel, Edgar A Samaniego, Opeolu Adeoye, Scott E. Kasner, Lee H. Schwamm and Gregory W. Albers.
