fMRI Insights on Word Similarity Within the Brain: Identifying Distinguishable Word Features for Speech BCI

Abstract

With its ability to investigate the similarity between brain responses to different stimuli, representational similarity analysis (RSA) may be a powerful tool to identify distinguishable linguistic features for speech BCIs. But despite RSAs potential, its adoption in speech BCI research has been somewhat limited and only moderately successful. The current study aims to bridge this gap by applying RSA to word production fMRI data – allowing the investigation of neural similarity between different words. 11 healthy subjects pronounced 28 words during fMRI image acquisition at 7T. Representational Similarity Matrices (RSMs) were computed from activity within the sensorimotor cortex, the cerebellum and the superior temporal area. The magnitude and distribution of similarity levels within these RSMs suggested inconsistency of neural responses to words. In an attempt to improve the quality of our data, various correction methods were applied. These included the reduction of general noise by regressing out white matter principal components, as well as accounting for pronunciation-related movements by excluding trial-pair comparisons with substantial head position divergence and regressing out trial means. None of these correction methods succeeded in revealing consistent neural responses to words, rendering further interpretation of word similarities inappropriate. The successful cross-validation of our RSA configurations with gesture data indicates that our implementation is fundamentally sound, and simultaneously hints towards reasons as to why our word dataset may be unsuitable for RSA. The findings of the present study are discussed with regards to fMRIs temporal resolution, pronunciation-related motion artifacts, voxel selection, and amounts of trial repetitions.