Cortical atrophy signatures and machine learning MR‐based classification of primary progressive aphasia variants

Background Primary Progressive Aphasia (PPA) is a rare neurodegenerative disorder characterized by progressive language impairment and three clinical phenotypes: semantic variant (svPPA), logopenic variant (lvPPA) and non‐fluent variant (nfvPPA). The goals of the present study were to 1) identify the group brain atrophy “signatures” of each PPA clinical phenotype, 2) determine the degree of individual variability in the spatial extent of atrophy within each clinical phenotype, and 3) explore the ability of a machine learning algorithm to classify patients based on regional atrophy values. Method Anatomical MRIs were collected from patients with a clinical diagnosis of svPPA (N=22), lvPPA, (N=21), nfvPPA, (N=18), and age‐matched healthy controls (N=25). Group‐level atrophy “signatures” were created using whole‐brain group GLMs comparing cortical thickness of each PPA variant to the control group. Next, group percent overlap maps were created by binarizing each patient’s z‐normalized atrophy map (relative to the control group, thresholded z<‐1.5), adding those maps together, and then dividing by each group’s N. Finally, a linear support vector machine (SVM) was applied to w‐scored atrophy values (controlling for intracranial volume, age, and sex), from brain regions identified as being most frequently affected within each patient group. Result 100% of svPPA patients demonstrated atrophy in the left anterior temporal lobe. Atrophy patterns in the lvPPA and nfvPPA groups were more heterogeneous; 80% of lvPPA patients demonstrated atrophy in the left posterior superior temporal, middle temporal, inferior temporal, and fusiform gyri; and 60% of nfvPPA patients demonstrated atrophy in the left pars opercularis, caudal middle frontal, and superior frontal gyri. The linear SVM classified individual patients with a total precision (true positives/true positives+false positives) = 91.6% (svPPA=98.4%, lvPPA=88.2%, nfvPPA=78.3%) and a recall, or sensitivity (true positives/true positives+false negatives) = 90.0% (svPPA=95.45%, lvPPA=88.25%, nfvPPA=76.25%). The features identified as most important for classification were the left temporal pole, middle temporal gyrus, and insula. Conclusion Our results provide further evidence that the PPA variants have distinct neuroimaging profiles that can support diagnosis, with svPPA and lvPPA patients exhibiting fairly homogenous atrophy signatures. We further showed that a linear SVM applied to regional atrophy values can classify individual patients with high accuracy.