Before the days of ChatGPT and Gemini, dentist Brian Kirkwood became interested in AI technology. After running into some engineers at a conference, he realized he could work with them to build AI tools that would help clinicians expand their skill sets. 

Kirkwood, now working on a doctoral project at the University of Texas San Antonio analyzing how clinical judgment impacts the development and performance of AI models, sees many models already being used to streamline the treatment process. 

“There are companies that are using AI to highlight where pathology is on a radiograph, bringing clinicians’ eyes to that area,” Kirkwood said. “Other systems are out there looking at ways to optimize a clinical practice. So now you have chat bots or AI agents, where a patient could reach out and get some advice relatively quickly whether they need to go in or not.”

Many hospitals and medical research institutions are focusing their AI developments on diagnostic processes and early detection. A 2019 study by the Mayo Clinic found that applying AI to the results of an electrocardiogram (EKG) is an effective and accurate way to diagnose left ventricular dysfunction, a precursor to heart failure. 

While AI is a tool that can aid clinicians, Kirkwood stresses the importance of not solely relying on these technologies.

“The downside is we could get complacent,” Kirkwood said. “We could get used to using these systems and ignore our clinical judgment and that would be my biggest concern. That's why my interest in research is, how does clinical judgment impact the development assessment and performance, because we all know that not every clinician is the same.”

This complacency can be seen in racial biases that affect the treatment process. When an AI model is trained to interpret test results or medical records based on past patient records, it only learns how to work with populations that historically had regular access to healthcare.

“Those biases in the historical data can be kind of inherent in the models that are developed,” said Cara Joyce, the director of biostatistics at Loyola University Chicago’s Stritch School of Medicine. “You want to make sure if you're applying those models, that they don't unfavorably affect certain subgroups, especially those that are disadvantaged.”

This lack of representation for minority patients in past healthcare records can be seen through disparities in medical insurance rates. In 2022, the Kaiser Family Foundation (KFF) reported that Hispanic Americans had an uninsured rate of 18% and black Americans had a rate of 10%, while white Americans were only uninsured at a rate of 6.6%. Due to their higher rates of uninsurance, minority patients have less affordable access to healthcare and therefore fewer medical records available to train AI algorithms.

According to a U.S. Department of Health and Human Services Office of Minority Health (OMH) study from 2023, the algorithm previously used to determine whether a patient could have a Vaginal Birth after Cesarean Delivery (VBAC) only worked effectively for certain demographics. It predicted that black and Hispanic/Latino women were less likely to have a successful VBAC than white women. The algorithm could only predict VBAC success rates correctly for white women, causing non-white women to have more cesarean procedures than necessary.

That algorithm has since been changed to work for patients of all demographics, but the OMH says the revisions took years of work. Researchers like Joyce are calling for AI developers to implement practices that ensure that algorithms work correctly for all patients from the start.

“I think it's really important to have multidisciplinary teams, so I'm a big believer in team science,” Joyce said. “You want to have representation on your team of those who are going to be affected by the decisions that your models make.”

When AI algorithms that aid in medical practices are developed properly and free of bias, they have the potential to simplify healthcare system for patients. Asha Behrman, a research assistant who studies sleep health at Loyola’s Activity Matters lab, believes AI can increase accessibility to medical knowledge.

“I think it would be really interesting to see how it could increase accessibility for sleep health, for people that can't afford to go to a psychologist or even a doctor that specializes in sleep,” Behrman said. “I think it could offer a lot of helpful advice, especially for youth who are constantly on the internet to learn about the impacts of some of the things that they're doing in their life that might be negatively impacting their sleep.”

Kirkwood sees similar benefits, comparing the development of AI to the invention of the Automated External Defibrillator, a medical device now standard in public spaces benefitingboth doctors and regular people.

“What we're doing here is taking physiological inputs into a system that a novice user could utilize, then that device will assess the patient's condition and determine whether it could administer shock that could help potentially save someone's life,” Kirkwood said. “Using that idea, could we create technologies that would enable healthcare providers to either train the next generation or do our job more efficiently?”