Stroke, ALS, and other neurodegenerative and acute brain disorders can rob people of the ability to speak. Brain-to-computer interfaces can restore it by translating brain activity into machine-generated speech. But they rely on patient-by-patient data collection, a major barrier to making these devices practical and widely available.
Duke researchers are making these systems faster and easier to deploy. In a study published in Nature Communications, researchers demonstrated — for the first time in humans — that pooling data across patients can work, rather than relying on fully individualized training. The team was led by Greg Cogan, PhD, associate professor in neurology, and Jonathan Viventi, PhD, Hawkes Family Associate Professor in Biomedical Engineering.
By using high-resolution recordings (micro-ECoG recordings) from awake neurosurgical patients, they aligned brain activity across individuals to train a shared decoder, improving performance while reducing the amount of data needed from each patient.
“In our study, a working decoder could be built from as little as five minutes of the new patient's own recordings, supplemented by aligned data from others,” Cogan said.
It’s an encouraging step toward making these systems more scalable and accessible.
Other Duke authors: Zac Spalding, Suseendrakumar Duraivel, Shervin Rahimpour, Charles Wang, Katrina Barth, Ceci Schmitz, Shivanand P. Lad, Allan H. Friedman. Derek G. Southwell.
Funding: The National Institutes of Health, the National Science Foundation, and a Duke Institute for Brain Sciences Incubator Award.