Neuroprosthesis for Decoding Speech in a Paralyzed Person with Anarthria.
Level 4 - case-series / case-control
Single-subject interventional case study / proof of concept
PubMed 34260835 · doi:10.1056/NEJMoa2027540
What was done
A high-density subdural multielectrode array was implanted over the speech sensorimotor cortex in a single participant with anarthria and spastic quadriparesis resulting from a brain-stem stroke. Over 48 sessions (22 hours of cortical recording), cortical activity was recorded while the participant attempted to articulate words from a 50-word vocabulary set. Deep-learning algorithms were trained to detect and classify words from neural activity patterns, and a natural-language model applying next-word probabilities was integrated to decode full sentences in real time.
What was found
Sentences were decoded from cortical activity in real time at a median rate of 15.2 words per minute, with a median word error rate of 25.6%. In post hoc analyses, 98% of word attempts were detected, and words were classified with 47.1% accuracy. Signal stability was maintained across the 81-week study period.
Why it matters
The study demonstrates direct real-time decoding of intended words and sentences from cortical activity, establishing feasibility for speech neuroprostheses in individuals with lost vocal articulation.
Limits
The study is limited to a single participant (n=1) with paralysis from a brain-stem stroke, precluding generalization to other etiologies. Testing was constrained to a small predefined 50-word vocabulary, required invasive brain surgery, and exhibited a median word error rate of 25.6%.
Cited by
- supports Neuroengineering research led by Dr. Edward Chang has mapped neural activity to vocal tract control (larynx and pharynx) to decode speech and enable communication in paralyzed patients.