When stroke patients with the communication disorder aphasia undergo language therapy, results often vary. Some survivors recover their language skills almost completely; others show only partial improvement. The reasons for these vastly different responses are often not readily apparent.
Anne Billot, a PhD student in behavioral neuroscience at the School of Medicine, wanted to find out what factors contributed most significantly to language recovery and if a patient’s response to therapy treatments could be predicted from the outset. In 2021, she teamed up with Swathi Kiran, director of Sargent’s Aphasia Research Laboratory and the James and Cecilia Tse Ying Professor in Neurorehabilitation, and an interdisciplinary team of researchers from BU, Northwestern University, and Johns Hopkins University, to study uses of machine learning for predicting language recovery success in aphasia survivors.
“This is one of the first studies to use this approach in studying people with aphasia,” says Billot (MED’23). “Knowing in advance how people will recover is very important for patients themselves and their family, so that they can plan how their life will be affected. It’s also important for clinicians to choose the best treatments that will be most beneficial for the patients.”
Aphasia is a communication disorder that results from damage to the parts of the brain that control language formulation and processing. For the initial part of the study, 55 stroke survivors with aphasia—who were all at least six months post-stroke—received treatment at BU, Northwestern, or Johns Hopkins for two hours a week over 12 weeks. The researchers had previously conducted behavioral assessments of the participants, who, Billot says, exhibited a wide range of aphasia severity. The therapy targeted areas including naming, syntax processing, and spelling—all of which can become impaired following a stroke. The researchers also took fMRI scans of the patients’ brains to gather sets of neuroimaging data. As expected, at the end of the treatment, some participants nearly fully recovered their language skills. Others did not recover as well.
Developing Models
After the course of treatment, Billot met with colleagues from BU’s computer science department and the Rafik B. Hariri Institute for Computing and Computational Science & Engineering—which provided funding for the project. They got to work figuring out how all the data gathered from the neuroimaging and behavioral assessments, combined with additional background on the patients, such as demographic information, could be used to develop machine learning models to predict how the patients would ultimately recover. Billot and her fellow researchers also wanted to see what types of information would be most important for a model to make the most accurate prediction.
“This sort of work is a game changer for the clinical field,” says Kiran. “It’s not a new approach for the computer science or data science fields. They’ve already developed these models and they’re applying them to these different problems, like weather prediction and climate sustainability. But when you apply these models to stroke rehabilitation, you now have a way to make very concrete predictions of what’s most important to recovery.”
This sort of work is a game changer for the clinical field.
—Swathi Kiran
The researchers found that the best combination of data for predicting language recovery after a stroke included the aphasia severity, where the lesion from the stroke is and its size, demographics such as age and education level, and resting-state connectivity—or how different areas of the brain work together when at rest. And how well that connectivity has been preserved post-stroke is the key predictor in how fully survivors will recover—a breakthrough from the study, according to Billot and Kiran.
“When you have a stroke, you have severe damage to one or more parts of the language regions of the brain,” says Kiran. “What’s remaining, how areas of the brain are in sync with each other, seems to predict how much somebody can recover with intervention. I think this is a big deal because it says what’s remaining is what’s working. And what’s working predicts how much somebody will improve from treatment.”
Billot sees opportunities to expand the study—namely, using machine learning to prescribe the most promising treatment for an aphasia patient. “With this study, we were more looking at which type of data could help us predict how someone will respond overall to language treatments,” Billot says. “The next step is knowing which data we need to use, to look at each type of treatment, and then really reach this kind of precision medicine that is the end goal.”