Poster B68, Thursday, August 16, 3:05 – 4:50 pm, Room 2000AB
A cognitive psychometric model of picture naming improves lesion-symptom maps
Grant Walker1, Julius Fridriksson2, Gregory Hickok1;1University of California, Irvine, 2University of South Carolina
A lesion-symptom map (LSM) can be used to infer brain function from brain damage, correlating behavioral symptom measures with measures of stroke lesions derived from modern neuroimaging. The interpretability of an LSM especially depends on good measurement models for symptoms. We recently developed a cognitive psychometric model to assess picture naming abilities in aphasia derived from 8 response type categories based on lexical status and semantic and phonological relations with the target (Walker, Hickok, & Fridriksson, 2018). We used a multinomial processing tree (MPT) to separate lexical selection errors from phonological selection errors, and we used item response modeling to separate the effects of participant abilities and item difficulties on the probability of successful selections. Here, we test the validity of the ability estimates by comparing them with 3 other ways of modeling picture naming deficits that neglect item information: 1) response type frequencies, 2) connectionist model weights (Foygel & Dell, 2000), and 3) varimax-rotated principal components of normalized error frequencies. We examined the generalizability of these measures to independent data, both behavioral and neurological. First, we examined correlations between the naming measures and scores on 3 word and picture matching tests and scores on 3 speech repetition tests from 127 people with aphasia. We found that the MPT abilities had the strongest correlations for 4 out of 6 tests (R=.64-.69), including 2 matching and 2 repetition tests. The other tests had slightly stronger correlations with the frequency of correct responses and the 1st PC of error frequencies (mainly Omission, Unrelated, and Abstruse Neologism), measures of general severity that do not distinguish between different types of mental computation. Next, we generated voxelwise LSMs for each naming measure from 81 people with aphasia using multivariate PLS regression. We used cross-validation (training n=70, testing n=11; arbitrary split) to evaluate generalizability, comparing MPT ability predictions using LSMs with predictions using lesion volume and with other naming measure predictions using LSMs. MPT abilities were better predicted by LSMs than volume (about 3% more variance accounted for with LSM) and were better predicted by LSMs than other measures of separate latent processes (i.e., connectionist weights or error frequencies). Next, we examined associations between stroke volume and MPT abilities in 61 brain regions of the JHU atlas, using multivariate OLS regression to control for damage outside of the region. Permutation of lesion-behavior pairings (n=10,000) was used to test for significance, and bootstrap resampling (n=10,000) was used to test for robustness. Lexical abilities were associated with damage in lateral temporal regions, while phonological abilities were associated with damage to the parietofrontal network, consistent with functional neuroanatomy models. Finally, we tested whether participants with apraxia of speech were influencing LSMs; excluding these participants dramatically changed the LSM for phonological ability but not lexical abilities, by reducing cross-validation prediction error and shifting associations to the posterior superior temporal gyrus. The results demonstrate that the MPT model measures cognitive abilities that depend on separable neural substrates and improves the interpretability and generalizability of measurements to independent behavioral and neurological data.
Topic Area: Computational Approaches