Decoding speech from human motor cortex using an intracortical brain computer interface


Clinical Trials

Poster Number: 40


Daniel Rubin, MD, PhD, Center for Neurotechnology & Neurorecovery, Mass. General Hospital; Harvard Medical School, Tommy Hosman, MS, Brown University, Anastasia Kapitonava, BA, Center for Neurotechnology & Neurorecovery, Mass. General Hospital, John Simeral, PhD, Center for Neurorestoration and Neurotechnology, Providence VA Medical Center; Brown University, Sydney Cash, MD, PhD, Center for Neurotechnology & Neurorecovery, Mass. General Hospital; Harvard Medical School, Leigh Hochberg, MD, PhD, Massachusetts General Hospital, Brown University, and Providence VA Medical Center

Background: For people with muscular dystrophy, ALS, locked-in syndrome, and other neurologic conditions causing quadriparesis and severe dysarthria/anarthria, the loss of fluent communication is highly disabling. Previously, intracortical brain computer interfaces (iBCI), including the BrainGate Neural Interface System, have allowed people with paralysis to type and write individual letters and decode phonemes. Here we describe efforts to decode intended speech using an iBCI. Understanding how the semantic information of language is translated into motor signals within cortex will enable the development of more accurate algorithms for decoding speech directly from neural activity.
Methods/Objectives: Research is conducted with permission under an IDE from US FDA and the MGH IRB. A 37-year-old with quadriplegia from a spinal cord injury enrolled in the BrainGate clinical trial had two 96-channel microelectrode arrays placed chronically in dominant precentral gyrus. During recording sessions in his home, the participant read a series of pseudo-randomly presented words while we recorded neural activity. To vary context, word cues were presented as either text, pictures, or both, and another cue type indicated whether the word was to be pronounced with a normal or prolonged enunciation.
Results: Principal component analysis of the 192-channel neural activity identified the 35-dimensional feature space capturing the greatest variance across the task. We trained a support vector machine (SVM)-based decoder to distinguish between each spoken word and epochs of silence. In cross-validated analyses, the SVM correctly distinguished speech from silence in 96.6% of 2440ms epochs spanning 1176 spoken words over 3220 seconds of word reading. Individual articulatory elements of speech were well discriminated, with clustering of words observed most strongly on initial consonant sound. Decoding performance was enhanced in the elongated compared to normal enunciation context and in the combined picture and text compared to either picture or text alone contexts, suggesting that task context influences the discriminability of the neural representation of speech in motor cortex.
Conclusions: In this preliminary study, iBCIs can be used to detect intended speech directly from motor cortex. Ongoing work is focused on refining decoding algorithms to restore fluent speech-based communication to people with severe dysarthria and anarthria.