Individual Word Classification During Imagined Speech Using Intracranial Recordings

Abstract

In this study, we evaluated the ability to identify individual words in a binary word classification task during imagined speech, using high frequency activity (HFA; 70–150 Hz) features in the time domain. For this, we used an imagined word repetition task cued with a word perception stimulus, and followed by an overt word repetition, and compared the results across the three conditions. We used support-vector machines, and introduced a non-linear time-realignment in the classification framework—in order to deal with speech temporal irregularities. As expected, high classification accuracy was obtained in the listening (mean = 89%) and overt speech conditions (mean = 86%), where speech stimuli were directly observed. In the imagined speech condition, where speech is generated internally by the patient, results show for the first time that individual words in single trials were classified with statistically significant accuracy. Classification accuracy reached 88% in a two-class classification framework, and average classification accuracy across fifteen word-pairs was significant across five subjects (mean = 58%). The majority of electrodes carrying discriminative information were located in the superior temporal gyrus, inferior frontal gyrus and sensorimotor cortex, regions commonly associated with speech processing. These data represent a proof of concept study for basic decoding of speech imagery, and delineate a number of key challenges to usage of speech imagery neural representations for clinical applications.