Neuroanatomical mapping of artificial intelligence‐based classification of language in PPA

Abstract

AbstractBackgroundPatients with Primary Progressive Aphasia (PPA) are usually subtyped into one of the three canonical subtypes (nfvPPA, svPPA, lvPPA) based on a neurological and cognitive assessment including their language characteristics typically measured from a battery of confrontational tests. While widely used, this classification system has been criticized, and also the approach makes assumptions about the features of language that are important to measure. Here we use methods from Artificial Intelligence to measure features of speech from naturalistic connected speech samples with the goal of determining how well this data‐driven approach matches independent clinical subtypes and how its results relate to neuroanatomical abnormalities measured from MRI.MethodLanguage data was obtained from 78 PPA patients (28 nfvPPA, 26 lvPPA, 24 svPPA) describing the WAB Picnic Scene. A transformer model, RoBERTa, was used to measure similarities in language features, IVIS to perform dimensionality reduction, and nested k‐means to cluster language samples. We then examined the cortical atrophy patterns of these clusters of PPA patients versus healthy controls (N=25).ResultSeven PPA clusters were identified with 88% agreement with the classic classification system. Individuals in Clusters 1 and 2 (mainly nfvPPA) exhibited speech dysfluency and reduced clausal complexity with atrophy in the left pars opercularis, superior and caudal middle frontal gyri. Those in Clusters 3 and 4 (predominantly lvPPA) exhibited difficulties in subject‐verb agreement, demonstratives and tense and shared atrophy in the superior and middle temporal and inferior parietal gyri. Individuals in Clusters 5‐7 (mainly svPPA) exhibited deficits in nouns/verbs access with atrophy in the left temporal pole and inferior and middle temporal gyri.ConclusionData‐driven Artificial Intelligence methods applied to naturalistic speech samples from PPA patients identify clusters of patients that match well to clinical subtypes, and that exhibit cortical atrophy patterns typical of those subtypes. This suggests that PPA subtypes are natural kinds and that computational analysis of simple speech samples can be used to identify them.