Poster C18, Wednesday, August 21, 2019, 10:45 am – 12:30 pm, Restaurant Hall
Sparse canonical correlation analysis reveals anatomical correlates of naming errors in Primary Progressive Aphasia (PPA)
Rose Bruffaerts1,2, Jolien Schaeverbeke1, An-Sofie De Weer1, Natalie Nelissen1, Eva Dries2, Karen Van Bouwel2, Mathieu Vandenbulcke3,4, Rik Vandenberghe1,2;1Laboratory for Cognitive Neurology, Department Neurosciences, KU Leuven, 2Neurology Department, University Hospitals Leuven, 3Laboratory for Translational Neuropsychiatry, Department Neurosciences, KU Leuven, 4Psychiatry Department, University Hospitals Leuven
Primary Progressive Aphasia can manifest as a nonfluent variant (NF) in which speech production is impaired, a semantic variant (SV) with impaired comprehension and a logopenic variant (LV) with word finding difficulties. The majority of PPA patients demonstrate impaired picture naming albeit for different reasons: picture naming entails correct object identification, access to semantics, phonological retrieval and speech production. Identification of the regions implicated in generating different error types can elucidate pathological changes. Whereas univariate analysis can detect voxels that contribute significantly, multivariate analysis can be used to determine the maximal extent of regions that predict errors. Sparse canonical correlation analysis (sCCA, Avants et al., 2010) is an extension of principal component analysis aimed at finding linear relationships in a low-dimensional space derived from e.g. neuroimaging data. Here, we use sCCA to identify regions in which grey matter atrophy correlates with different types of naming errors. Audio recordings of the 60-item Boston Naming Test were scored in 41 PPA patients (19 NF, 13 SV, 9 LV; age range 49-79 y.o., 23 female). Errors were classified using the scheme developed for SV by Woollams et al. (2008), including omissions, semantic errors, superordinate errors and circumlocutions. We added speech production errors (distorted speech) as an error class. Segmentation was performed using SPM12 and CAT12 (voxel size 1.5x1.5x1.5mm³) and images were smoothed (8x8x8mm³ Gaussian kernel). The dataset was split into a training set (n=21) and a testing set (n=20). In the training dataset, sCCA (http://github.com/stnava/sccan) determined a subset of grey matter voxels, in which atrophy linearly correlated with error type frequency (cluster threshold: 200 voxels). This subset of voxels from the training dataset was then validated using the independent test dataset and the maximal sparseness at which a correlation between atrophy and errors was found (P<0.05) was determined. Scanner type, total intracranial volume and age were used as nuisance variables. Correct responses were given in 47% of trials for SV, 59% for NF and 69% for LV, with omissions in 19%; 10% and 4% respectively. Speech production errors occurred in 1% of trials for SV, 14% for NF and 4% for LV. Semantic errors were produced in 9% of trials for SV, 8% for NF and 8% for LV. Superordinate errors were generated in 5% of trials for SV, 1% for NF and 3% for LV. Circumlocutions were found in 11% of trials for SV, 2 % for NF and 6% for LV. Using sCCA, speech production errors correlated with bilateral atrophy of the premotor regions (sparseness: 0.001 meaning 0.1% of grey matter, 443 voxels). Semantic errors correlated with atrophy in the left ventral stream (sparseness: 0.008, 0.8% of grey matter, 2152 voxels). Superordinate errors correlated with bilateral frontooccipitotemporal atrophy (sparseness: 0.059, 5.9% of grey matter, 14483 voxels). Speech production errors in PPA correlated to atrophy in premotor regions. Semantic errors correlated with atrophy in the left ventral stream. Superordinate errors reflected widespread atrophy. The anatomical correlates of naming errors reflect impaired language and speech processing at different scales.
Themes: Disorders: Acquired, Meaning: Lexical Semantics
Method: Other