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

Poster B32, Wednesday, November 8, 3:00 – 4:15 pm, Harborview and Loch Raven Ballrooms

Predicting Western Aphasia Battery Subscores from the Spatial Distributions of Localized Brain Lesions

Grant Walker1, Gregory Hickok1, Julius Fridriksson2;1University of California, Irvine, 2University of South Carolina

Neuropsychology provides insight into the neurobiology of language by identifying critical brain networks that result in predictable deficits when damaged. We examined the relationship between the spatial distributions of stroke-induced brain lesions and performance on a comprehensive test battery of language function. We used archived, de-identified, neuroimaging and behavioral assessments from 162 participants enrolled in an ongoing study of left-hemisphere ischemic stroke at the University of South Carolina. Binary lesion maps were manually segmented by a neurologist from structural MR images and warped to the MNI space. The Western Aphasia Battery (WAB; Kertesz, 1982) includes behavioral tests to assess spontaneous speech fluency and information content (Speech), auditory verbal comprehension of single words and commands (Comp), repetition of phrases from single words to sentences (Rep), and naming objects or producing single-word responses to verbal prompts (Naming). These subscores combine into a composite measure of aphasia severity (AQ), and also indicate a diagnostic category. The average AQ in our sample was 62.7 (range: [5, 98.6]); the sample diagnoses included Broca’s (31%), Anomia (18%), Resolved (14%), Conduction (11%), Global (10%), Wernicke’s (7%), and a single Transcortical Mixed case. 3-D lesion maps were transformed in two ways: 1) total lesion volume (TLV) was calculated, and 2) each map was vectorized and normalized to have unit length, effectively controlling for TLV. All behavioral measures were normalized with z-scores. We constructed partial least-squares (PLS) regression models to predict each WAB subscore and the AQ from the lesion vectors, and we used simple linear regression to predict the same measures from TLV. We used permutation tests (n=1,000; one-tail, =.05) to identify voxels that were associated with a significant decrease in test scores. We constructed bootstrap estimates (1,000 samples) for the in-sample variance accounted for (VAF) and cross-validation estimates (1,000 pseudorandom 80/20 splits) for the out-of-sample VAF. We regressed TLV against the behavioral score residuals after removing voxel-based in-sample VAF, to determine if the transformations provided unique contributions to linear predictions. All behavioral measures yielded significant voxel associations. 18,472 Speech voxels (61% in-sample VAF) clustered in precentral and postcentral gyri, and inferior and middle frontal gyri. 22,969 Comp voxels (51% in-sample VAF) clustered in the temporal lobe. 8,837 Rep voxels (56% in-sample VAF ) clustered in posterior temporal lobe, supramarginal gyrus, and postcentral gyrus. 6,688 Naming voxels (51% in-sample VAF) clustered in the temporal lobe, including the pole, and the supramarginal gyrus. 611,236 AQ voxels (59% in-sample VAF) covered nearly 2/3 of the left lateral cerebrum. For all measures, voxel-based regression had ~10% more in-sample VAF than TLV; TLV accounted for an additional 2-3% variance in the residuals, either significant (p<.05) or trending (p<.10). The out-of-sample VAF was ~8.5% less than the in-sample VAF; TLV and PLS regression had similar out-of-sample VAF. The spatial distributions of lesions provided information about behavioral deficits beyond TLV. The PLS regression models help identify critical neural substrates of speech-language functions.

Topic Area: Language Disorders

Back to Poster Schedule