Poster D56, Friday, August 17, 4:45 – 6:30 pm, Room 2000AB
Relative Contributions of Lesion Location and Lesion Size to Predictions of Varied Language Deficits in Post-Stroke Aphasia
Melissa Thye1, Daniel Mirman1;1University of Alabama at Birmingham
Introduction: The past 15 years have seen a rapid increase in the use of lesion-symptom mapping (LSM) methods to study associations between location of brain damage and language deficits, but the prediction of language deficits from lesion location remains a substantial challenge. This study examined two factors that may weaken lesion-symptom prediction: (1) use of mass-univariate voxel lesion-symptom mapping (VLSM) and (2) use of broad language deficit scores. Multivariate LSM (specifically, sparse canonical correlation analysis, SCCAN) may overcome the limitations of the mass-univariate approach and be better able to determine the brain regions that are critical for a particular deficit resulting in improved lesion-symptom predictions. In addition, prediction accuracy may improve for deficits in functional systems compared to broad measures of aphasia severity, and lesion location may be particularly informative for predicting deficits within these systems. The present study used both mass-univariate (VLSM) and multivariate (SCCAN) lesion-symptom mapping and compared prediction of broad measures of aphasia severity with prediction of more specific language deficits. Methods: The data were drawn from a large-scale study of language processing following left hemisphere stroke. Participants (N=128) completed a detailed battery of psycholinguistic tests, which included broad measures of aphasia severity (Western Aphasia Battery Aphasia Quotient, WAB AQ) and object naming ability (Philadelphia Naming Test, PNT). Participants also completed a variety of more specific tests of spoken language, verbal working/short-term memory, and semantic cognition, which were entered into a principle component analysis to calculate deficit scores in each of three functional systems: Semantics, Speech Production, and Speech Recognition. Prediction accuracy was based on 8-fold cross-validation: participants were partitioned into eight “folds” (n=16 each) and, for each fold, VLSM and SCCAN were run on the training set of participants (n=112) to generate a mass-univariate and a multivariate template for each score. For each participant in the withheld test set, the proportion of overlap between the participant’s lesion and each template was calculated. Lesion size, overlap proportion, and overlap proportion controlling for lesion size were tested as predictors of each language deficit score. All analyses were implemented in R using the lesymap package. Results: For both VLSM and SCCAN, lesion size alone was a significant predictor of PNT accuracy, WAB AQ, Semantics, and Speech Production, and explained a greater percentage of the variance for WAB AQ (31%) compared to the other measures: PNT (21%), Semantics (10%), Speech Production (8%), and Speech Recognition (0.1%). After accounting for lesion size, template overlap proportion was a significant predictor of only Speech Production and Speech Recognition deficit scores. Conclusion: General measures of language deficit (PNT and WAB AQ) were substantially predicted by overall lesion size and lesion location (overlap proportion) did not improve prediction accuracy. Only Speech Production and Speech Recognition deficit scores were better predicted by the overlap proportion than by lesion size, possibly because these systems are supported by relatively localized neural systems and may be more resistant to reorganization.
Topic Area: Language Disorders