Noninvasive decoding of typed sentences from human brain activity.

Brain2Qwerty achieves a 29% character error rate in decoding sentences from magnetoencephalography in healthy volunteers, outperforming electroencephalography's 65%.

• Why it matters: Restoring communication in patients who cannot speak or move often requires invasive brain devices, which carry surgical risks. Developing noninvasive alternatives can improve safety and accessibility for these patients.
• What they did: Researchers trained Brain2Qwerty, a deep learning model, to decode sentences from electro- and magnetoencephalography data while 35 healthy volunteers typed memorized sentences on a QWERTY keyboard.
• The result: The model attained an 18% character error rate in the best participants and could accurately decode various sentences outside the training set, narrowing the gap between invasive and noninvasive brain-computer interfaces.