Researchers at the University of Pittsburgh and Carnegie Mellon University collaborated to create a customized app based on artificial intelligence that may help people with multiple sclerosis track their symptoms — and even predict them.

The app passively pulled health data from the personal digital devices — like smartphones and fitness trackers — of participants with multiple sclerosis to understand their depression and access more precise treatment for them. Study results were published in the Journal of Medical Internet Research in March.

After finding their machine learning model was able to accurately predict a patient's depression burden for two weeks out, the researchers hope in the future that it can be applied to different populations, such as those with breast cancer or suicidality.

Around three million people worldwide have multiple sclerosis, per the Atlas of MS map, including around 1 million in the U.S. The immune systems of people with MS attack the insulation that coats brains' neurons, called myelin. Just as insulation protects wires and helps send electric currents faster, myelin speeds up signals in the brain. When that coating decays, the connection between the brain and body frays too, leading to motor problems, speech delays, vision changes, numbness and memory issues.

MS is commonly correlated with depression, with 30-50% of those with MS reporting the mental health condition. Depression can also worsen fatigue, motor problems and memory issues someone with MS may already be experiencing.

"Depressive symptoms can reduce the quality of life as much as physical disabilities," said Zongqi Xia, associate professor of neurology at the University of Pittsburgh School of Medicine and lead author on the paper. "Early recognition of depression can help patients get effective treatment sooner.

"This approach helps patients with continuous self-monitoring in their own living environment between their routine clinic appointments," he added.

Mayank Goel, senior author on the paper and an assistant professor in CMU's computer science department, specializes in the application of machine learning in health care, and had already created a machine learning model to passively collect phone and fitness tracker data, in part to improve doctor-patient communication.

Xia read one of Goel's published papers and thought the technology could apply well to a cohort of MS patients he treats in the clinic at UPMC. He sees depression and MS co-occur in about 25-30% of patients he treats, in addition to fatigue and poor sleep, and he expects the true number may be even higher.

The project combines the strengths of both Pittsburgh-based institutions: the clinical practice at UPMC and the computer science at CMU.

At its face, it establishes a "warm start" for others to create machine learning models based on their methods, said Goel, with the overall goal of expanding knowledge on patient quality of life.

"A lot of people in the past have looked at the physical disability of multiple sclerosis, but the mental health burden, the behavioral burden, like the sleep quality, depression symptoms, is something that has not been studied that much," said Goel. "So we believe this is an important step in that direction."

The study included 104 people with MS and nearly 13,000 days of passively collected data, which was scrubbed from participants' smartphones, including step count, location, heart rate, sleep, call time and screen time. Researchers then fed the data into a machine learning model that calculated patients' depression symptom burden and predicted their symptoms for the next two weeks.

The model was able to predict with 70-80% accuracy a participant's symptom burden for them to take to their next doctor's visit and inform treatment steps. Xia said this accuracy range is common for other tools in the field, such as breast cancer risk predictors.

Participants seemed to respond well to the technology. While everyone enrolled completed the three-month timepoint, about half of the participants agreed to extend the study out to six months.

The study shows it's possible to create a digital tool to track MS symptoms without patients having to leave their homes or go to a facility, said Bruce Bebo, executive vice president of research for the National Multiple Sclerosis Society, in an email. He was not involved in the study.

It's an early study, he said, but adds to a growing body of research focused on answering questions about multiple sclerosis "by going to the source — people living with MS — and providing them with cutting-edge ways to measure how symptoms impact them each day."

If the app ultimately rolls out to the public, clinicians may be able to spot symptoms earlier and tailor treatment accordingly — perhaps even slowing down disability progression or restoring lost function, said Bebo.

In the AI era, issues have arisen about bias in training algorithms, such as transcription services not capturing people with accents and health risk indicators incorrectly predicting outcomes for women or people and color.

It's an issue Xia and Goel considered when training and applying their own model. Because of the breakdown of participant demographics, some of that bias was inevitable: Participants were mainly women and 93% of them were white.

"While we believe we have used approaches that would stop the bias from happening too much, some racial and socioeconomic bias is going to be there that we need to evaluate," said Goel. "As researchers, we want to build systems that support the use of responsible technology and data."

What happens next depends, too, on policy changes regarding AI regulation and rules around training algorithms, he said.

The researchers also took privacy into consideration. All data, after it was pulled, was de-identified, and location data was abstracted so specific locations couldn't be traced — just the movement from one location to another.

Xia also said they aimed to use the least amount of data from the digital sensors as possible to achieve their goal, to increase patient data safety and reduce the risk of privacy violations.

Moving forward, the team wants to optimize the model to apply to different populations. Goel is in contact with researchers at colleges across the U.S. about applying the model to a different demographic to expand its accuracy.

Goel has already looked at how the model works for breast cancer survivors and teens with suicidal ideation, and wants to expand their breadth.

"The model we built adapts to these other populations pretty well," he said.

Goel said he and Xia also plan to examine different kinds of depression to further promote precision medicine. For instance, some people get agitated while depressed, whereas others isolate.

"We want to collect more data to understand this better, because that is important for us to figure out what interventions should be done."

_____

© 2025 the Pittsburgh Post-Gazette. Visit www.post-gazette.com. Distributed by Tribune Content Agency, LLC.