A tailored approach: Spears professor creates AI models to analyze stroke risk

A tailored approach: Spears professor creates personalized AI models to analyze stroke risk

Friday, November 1, 2024

Media Contact: Terry Tush | Director of Marketing & Communications | 405-744-2703 | [email protected]

As a Spears School of Business management science and information systems professor, Dr. Xiao Luo has a keen awareness of our increasing reliance on artificial intelligence to improve our lives.

From GPS navigation systems to music streaming services, AI technologies offer convenience, entertainment and connection.

Luo, who holds a Ph.D. in computer science from Canada's Dalhousie University and a bachelor's degree in the same major from China's Huazhong University of Science and Technology, is using her expertise to think beyond those everyday uses.

How can AI transform health care to keep us alive?

Luo started focusing on this big question when she worked in the Canadian health care system, noticing troves of patient data that often went underutilized. Instead of stowing away electronic health records like boxes of valuables in a dusty attic, Luo wanted to unlock their tremendous potential to help patients.

"I was like, 'OK, maybe I should come to academia to do some research, to work with physicians on how we can improve the health care system in general by utilizing those data," Luo said. "That's the first motivation. That's why I work with physicians in the health care domain."

The Oklahoma State University faculty member collaborated with Dr. Alan Sawchuk, an Indiana University professor of surgery, for her ongoing study, "Developing Large Language Model (LLM)-powered interpretable AI models to predict stroke risks in patients with asymptomatic carotid stenosis (ACS)." This summer, Luo's proposal was one of seven to receive a $10,000 seed grant as part of a new Spears Business internal program to catalyze research funding.

With her groundbreaking research, Luo strives to make lifesaving health care more accessible. Particularly in rural areas, patients might not have access to advanced biomedical imaging systems to detect stroke risk, but electronic health records are readily available tools that do not require excessive costs.

After arriving at a doctor's office, a patient typically fills out paperwork in the waiting room and then answers a nurse's questions about health history and habits. Those seemingly ordinary steps produce the valuable data points that could - thanks to Luo's models - prevent stroke.

The key lies in AI's ability to create personalized risk-assessing health profiles.

Many AI technologies are popular because they make customized predictions for individuals. Think of the earlier examples: GPS systems can guess your destinations after learning the routes you regularly travel, and music streaming services create "For You" playlists based on your listening patterns.

In this fundamental way, Luo and Sawchuk's technology is no different. AI can synthesize an individual patient's medical data, in this case creating unique predictive profiles for those who have ACS, a diagnosis characterized by artery blockage or narrowing. These personalized data sets are intended to allow health care providers to assess patients' stroke risks, weighing numerous factors to determine if a preventative operation is necessary. In clinical settings, patients would sign waivers agreeing to input their data into the system as they already do with electronic health records.

"We can analyze large amounts of data, look at the changes in the data, look at the trends in the data and do the personalized analysis," Luo said.

This decision-making tool could preserve a patient's quality of life and drastically reduce medical expenses. According to the Centers for Disease Control and Prevention, stroke occurs in more than 795,000 people in a year in the United States. Additionally, expenses associated with stroke totaled almost $56.2 billion in the U.S. between 2019-20, per the CDC.

A patient with ACS might never have externally visible health complications until a stroke occurs. Luo said often, these patients undergo operations to prevent stroke, but the procedures are sometimes unnecessary. Occasionally, the surgeries themselves increase health risks, particularly for aging patients. Sometimes, a lifestyle change might be enough to minimize risk. No two humans are the same, so health care providers cannot rely solely on generalized guidelines to evaluate the probability of stroke.

This is why Luo and Sawchuk are assessing individualized data. The AI models synthesize information that includes demographics, comorbidities, vital signs, medication history, health habits, laboratory test results and more. Luo is training the AI models to not only determine a patient's stroke risk, but also communicate with physicians and patients to help them understand the calculation.

"If AI only produced a probability, like a 90% chance this patient is going to develop stroke, it's not convincing because physicians need to know why," Luo said. "AI needs to explain. How did you come to this conclusion? We call it explainable AI. In order to have physicians utilize these in the clinical setting, it has to be explainable."

For AI to communicate like a person, the machine needs human coaching.

To build the model, Luo and Sawchuk reviewed data from 872 ACS patients, whose identities were not disclosed in the study, in the Indiana University School of Medicine system from 2009-22. They partnered with physician residents and students who annotated those past records, able to note if the patients did or did not experience stroke.

Compared with the AI-generated risk assessments, these historical outcomes could verify the accuracy of the AI and guide Luo in fine-tuning it. Seventy percent of the initial patient data was used just to train the AI, while 30% was devoted to testing it.

Jackson Silvey, a business analytics and data science master's student, assisted Luo with this preliminary work, focusing on data processing and engineering algorithms for machine learning. Luo has been creating this AI model to utilize scientific publications and clinical guidelines to gain deep knowledge about ACS, essentially building a medical encyclopedia for disease prediction and analysis.

With this knowledge base, the AI models can offer an in-depth look at health care as opposed to scratching the surface. A busy physician typically has a short window of time to meet with one patient before the next appointment begins, Luo pointed out, so they might not be able to review a patient's comprehensive health history.

"The physician may only look at the recent lab test or compare it to a few previous ones," Luo said. "For us, we can consider long-term and complex medical history by using artificial intelligence."

As Luo strives to include clinical data from electronic health records systems across the United States, she is cognizant of the difficulties. While the AI models start as neutral entities, they can, like the mind of a child, develop bias depending on how they are taught. If the sample data were to include only male patients, for example, then the model would not be adequately equipped to guide health care decisions for women.

Luo has paid careful attention to these factors, making sure to evaluate men and women in the initial patient sample and note differences that could be due to biological sex. She said the research team is also striving to add more Black patients and Hispanic patients to the database, groups that were underrepresented in the existing IU records. After their pilot study using IU's patient files, Luo and Sawchuk intend to present a proposal to the National Institutes of Health to expand their research to OSU Medicine and other medical institutions, reaching new populations.

"AI needs to consume a lot of data in order to make correct decisions," Luo said. "So, that's a challenge of ultimately utilizing AI in the real world. We must have an unbiased algorithm, although data might be biased."

Once the AI model is sharp enough for clinical settings, it does not have to be limited to ACS patients. This predictive technology could apply to numerous areas of medicine, including cancer research, diabetes research and more. Luo, a widely published scholar, thrives at the intersection of AI and medicine, constantly seeking novel ways to increase the quality and accessibility of health care.

As a Spears Business professor since August 2023, she is already exemplifying the difference-making potential of the school's elevated focus on faculty grants.

"I feel very appreciative to receive the seed grant, and I also feel very proud to work with the school that provides faculty the seed grant to advance their research," Luo said. "It's an important research collaboration with different institutions and across different departments at OSU.

"This seed grant not only advances faculty research and encourages more collaborations but can also provide students opportunities to research with me in my lab to understand how we can utilize AI to advance human health, to improve health care outcomes."

Photo by: Devin Flores
Story by: Hallie Hart | Discover@Spears Magazine

Public link

Please use the above public link if you want to share this noodl on another website.