Near Infrared Spectroscopy based Brain Computer Interface for Decoding Speech

Abstract

People suffering from neuromuscular disorders such as tetraplegia are left in a lockedin state with preserved awareness and cognition. Brain-computer interfaces (BCI) can potentially redefine the quality of life of such individuals by allowing them to communicate their intention through modulation of localized brain activity. Near Infrared Spectroscopy (NIRS), a relatively recent BCI modality, can be used to non-invasively monitor such an activity by measuring corresponding changes in cerebral blood oxygenation. In this study, it was hypothesized that the activation of Broca’s area due to auditory imagery as conveyed by local hemodynamic activity can be harnessed to create an intuitive BCI based on NIRS. A 12-channel square template was used to cover inferior frontal gyrus and changes in hemoglobin concentration corresponding to six aloud (overtly) and silently (covertly) spoken words were collected from 8 healthy subjects. The features extracted from each of the trials using unsupervised feature learning were classified with an optimized support vector machine. The results showed large intra- and inter- subject variability. For all subjects, when considering overt and covert classes regardless of words, classification accuracy of 95.83% (±5.87%) was achieved with deoxy-hemoglobin (HHb) and 94.22% (±6.87%) with oxy-hemoglobin (O2Hb) as a chromophore. For a six-class classification problem of overtly spoken words, 66.48% (±17.07%) accuracy was achieved for HHb and 58.90% (±27.68%) for O2Hb. Similarly, for a six-class classification problem of covertly spoken words, 70.07% (±12.11%) accuracy was achieved with HHb and 65.91% (±16.89%) with O2Hb as an absorber. These results indicate that a control paradigm based on covert speech can be reliably implemented into future BCIs based on NIRS.