Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Individualized computational models recapitulate directly-measured neural correlates of sematic and phonological impairments
Poster Session D, Saturday, October 26, 10:30 am - 12:00 pm, Great Hall 3 and 4
Ryan Staples1, Andrew T. DeMarco1, Peter E. Turkeltaub1,2; 1Georgetown University Medical Center, 2MedStar National Rehabilitation Hospital
Left-hemisphere strokes frequently result in alexia. Artificial neural networks (ANNs) provide a mechanistic model for the impaired cognitive processes in alexia. Whether ANNs are adequate models of the brain, however, is unclear. Here, we assess single-word reading in a sample of left-hemisphere stroke survivors. We impair models of reading by damaging semantics, phonology, or mixed semantics/phonology. Model damage parameters are matched to stroke survivors, and we assess how well the models reproduce patterns of word reading accuracy. We use the stroke survivor-matched model damage parameters as dependent variables in multivariate voxel-based lesion symptom mapping (VLSM) analyses. We hypothesized that the models could reproduce reading accuracies across high and low frequency and consistency words in stroke survivors, and that the model damage parameters would co-localize with directly measured semantic and phonological impairments. Fifty-two left-hemisphere stroke survivors read aloud single words (varying in frequency and consistency) and pseudowords. Participants also performed semantic (Pyramids and Palm Trees, TALSA category judgement) and phonological (pseudoword repetition, rhyme judgement) tasks. We trained five independent instantiations of a triangle model of reading. Model instantiations differed only in their randomized starting weights. Models were then lesioned by removing percentages of the connections into and out of the phonological and semantic layers 10% intervals (10-90%) to model all combinations of severity of phonological and semantic damage. Each of the 99 possible lesions were modeled 15 times and the average accuracies were recorded for each lesion on: pseudowords, all words, and words in each bin of high vs. low frequency and consistency (four word types). Stroke survivors were matched to lesioned models using Euclidean distance between two-dimensional vectors of word reading and pseudoword reading accuracy. Model accuracy on the four word types was compared to real participant performance. Support vector regression-VLSM was used to determine if specific lesion locations corresponded to the modeled ANN lesions of (1) phonological damage and (2) semantic damage. Model lesion parameters were inversely correlated with phonological (r(50)=-0.68, p<0.001) and semantic (r(50)=-0.29, p=0.047) task accuracy. Stroke survivors’ reading of the high/low frequency/consistency words was well-matched by the models (correlation between model and stroke survivors’ reading: high-frequency consistent words, r(50)=0.96, p<0.001; high-frequency inconsistent words, r(50)=0.941, p<0.001; low-frequency consistent words, r(50)=0.968), p<0.001, low-frequency inconsistent words, r(50)=0.85, p<0.001). VLSM showed that the phonological lesion parameter is related to supramarginal (SMG) and superior temporal gyrus (STG) lesions. Semantic model lesions localized to left angular and middle occipital gyrus (both voxelwise p<0.005, clusterwise p<0.05). Model parameter VLSM results overlapped with directly measured phonology and semantic VLSM results. Our results critically show that ANNs implementing the triangle model are models of not only cognitive processes, but also of the brain. Though matched only to reading accuracy, we show that the ANNs also capture non-reading phonological and semantic processes. Our findings show that ANNs provide a framework to parse individual variability in post-stroke alexia outcomes, potentially leading to improved therapeutic approaches. Future work will adjudicate between models implementing competing hypotheses of the neurocognitive processes in reading by comparing their fits to the brain.
Topic Areas: Disorders: Acquired, Computational Approaches