You are viewing the SNL 2017 Archive Website. For the latest information, see the Current Website.

Poster E53, Friday, November 10, 10:00 – 11:15 am, Harborview and Loch Raven Ballrooms

Lesion mapping of syntactic and lexical features derived from Natural Language Processing of narrative speech elicited by patients with chronic post-stroke aphasia

Ezequiel Gleichgerrcht1, John Delgaizo1, Julius Fridriksson2, Dirk den Ouden2, Alexandra Basilakos2, Chris Rorden2, Leonardo Bonilha1;1Medical University of South Carolina, 2University of South Carolina

Background: Typically elicited by an open-ended question or prompt (e.g. “Tell me about your day”) or the request to describe a scene (e.g. widely used “Cookie Theft” picture), narrative speech can provide valuable clinical information about a patient’s language skills well beyond the scores of standardized language batteries. The detailed analysis of responses to such tasks, however, requires specific training and can be highly time-consuming. Method: We applied an automated computational algorithm based on Natural Language Processing (NLP) to the verbatim two-minute transcriptions of three picture description tasks (i.e., six minutes of total connected speech) elicited by 64 patients with chronic dominant-hemisphere stroke. By means of sentence boundary disambiguation, parsing, and parts-of-speech tagging, we derived 37 lexical and syntactic features, which were subjected to factor analysis through principal components. In order to identify brain areas critical to language performance, we then conducted voxel-based lesion-symptom mapping (VLSM) on factors that, together, would explain at least 50% of the variance. We employed Freedman-Lane permutation to control for both family-wise errors and for fluency (WAB spontaneous speech fluency subscore) as a potential confounder given the nature of the task. Results: Dimension reduction using Varimax rotation yielded seven factors, which altogether explained 83% of the variance. Factor 1 and Factor 2, however, were enough to explain 50.3% of the variance, so further analyses focused on these two components. Factor 1 was strongly composed of mainly syntactic features (in decreasing order of loading scores: clause width, width/height and number of verbal phrases, number of words, speech rate, number of clauses and their height, distance between noun phrase and verb phrases and so forth). Controlling for the spontaneous speech fluency subscore of the WAB, lesion mapping revealed critical voxels in the left rolandic operculum (z=-3.66). Factor 2 was strongly composed of mainly lexical features (in decreasing order of loading scores: frequency of all words, number of nouns, adverbs, and adjectives, as well as adverb, noun, verb and overall lexical variation). Controlling for fluency, lesion mapping revealed critical voxels in the precentral (z = 3.90) and globus pallidus regions (z = 3.65). Conclusion: We showed that syntactic performance beyond speech fluency might rely on areas in the rolandic operculum, while lexical performance beyond fluency may depend on areas in the precentral gyrus and subcortical structures. Our findings show that NLP applied to connected speech elicited by patients with post-stroke aphasia can shed light on the organization of language in brains with vascular damage.

Topic Area: Computational Approaches

Back to Poster Schedule