BU Researchers Use “Digital Twins” to Predict Language Recovery After Stroke

By Brendan Galvin

For people recovering from a stroke, relearning how to communicate can be one of the biggest rehabilitation challenges. 

For bilingual patients, the challenge is even greater.

Which language should their speech therapy focus on?

A recent study published in npj Digital Medicine explores how artificial intelligence-powered “digital twins” can help answer that question. 

Led by Hariri Institute Faculty Affiliate Swathi Kiran—James and Cecilia Tse Ying Professor in Neurorehabilitation at Boston University’s Sargent College of Health & Rehabilitation Sciences; Founding Director of the Center for Brain Recovery (CBR); and Director of the Aphasia Research Laboratory—the work demonstrates how computer models can predict language recovery outcomes for bilingual stroke survivors, demonstrating the future of individualized rehabilitation.

The study focused on people with aphasia, a language disorder that commonly occurs after stroke and can affect speaking, understanding, reading, and writing. While treatment decisions for monolingual patients are relatively straightforward, those for bilingual patients are more complicated.

“If you speak multiple languages and you have a stroke, then you have difficulty communicating in multiple languages,” Kiran explained. 

“People who speak multiple languages don’t always get the right kind of therapy when they’re in the hospital, because if the clinician speaks English and the patient speaks English, then they do the therapy in English,” she said. “But if the patient speaks primarily Spanish and the clinician speaks English, they’re not getting the therapy in the right language.” 

In many clinical settings, language therapy is delivered in the clinician’s language rather than in the language that may be most beneficial for the patient. 

Traditionally, determining the optimal treatment language would require studying thousands of patients.

To address this challenge, Kiran’s team spent more than a decade developing computational models capable of creating a “digital twin” of an individual patient. 

Unlike large language models that predict words based on patterns in data, these digital twins are designed to mimic the brain’s language systems and the way neurons process words and sentences.

The result is a personalized computer simulation that can estimate how a specific patient is likely to recover under different treatment conditions.

“What we’re able to do with these digital twins is simulate the outcome beforehand,” Kiran said. “The computer model can tell us what recovery might look like if a person is trained in English versus Spanish and which outcome is likely to be better.”

The study enrolled 48 Spanish-English bilingual stroke survivors who had chronic aphasia in a randomized controlled trial. Half the participants were treated in the language prescribed by the digital twin based on its a priori simulation, and the remaining half were treated in the language opposite to that recommended by the digital twin. 

At first, the findings surprised the researchers. While all participants improved with therapy, there was no clear difference between the two groups overall.

As the team analyzed the data more closely, they found the digital twins were capturing highly individualized differences, including each person’s language abilities before the stroke and the severity of their brain injury.

“We thought that, on average, the digital twin simulation would predict the accurate language to treat,” Kiran said. “But it was actually much more personalized in the way it was making predictions.”

Kiran’s research traces back to a Hariri Institute Focused Research Program (FRP) , which brought together clinicians, computer scientists, and machine learning experts to explore how digital modeling could improve diagnosis and treatment. That interdisciplinary foundation provided the collaborative infrastructure to advance the project into this larger NIH-funded study.

For Kiran, the study represents an important step toward precision rehabilitation. 

Future research could move beyond selecting a treatment language and begin personalizing therapy at an even finer level, such as determining which words, exercises, or training strategies are most effective for a particular patient.

The next generation of digital twin technology may help ensure that future patients receive rehabilitation plans tailored specifically to their individual recovery needs.