Abnormal discourse in semantic dementia: A data-driven approach

Abstract

Structural and content-related deficits occur in connected discourse of patients with semantic dementia (SD). We used principal components analysis (PCA) to characterise the sources of variation in word usage during picture description by controls and SD patients. This data-driven approach allowed: comparison of the distance between individuals in the two-dimensional space; correlational analyses between principal component (PC) values and performance on other tests; identification of words whose variance contributed most to the definition of the PCs.Transcripts of Cookie Theft picture descriptions from 21 patients with SD and 21 controls were used to generate frequencies of all word types (n = 557) across participants. Frequency values of words with ≥10 occurrences (n= 81) were entered into a PCA. Values of emergent dimensions were correlated with performance on tests of single word meaning. The first PC accounted for 59% of the variance, the second for a further 10%. Patients and controls showed good separation within the resulting space. Factor loading scores indicated that control performance was characterised by function (factor 1) and content (factor 2) word usage, while patients showed a greater tendency to use pronouns, deictic and generic words. Knowledge of single word meaning correlated with factor 1 but not with factor 2. Differences in word usage can differentiate connected speech of SD patients from controls using a rapid, automated, data-driven algorithm. The distinction between groups, loadings on the two components, and their differential correlations with semantic tasks raise the possibility of independent differences in syntax and lexical content.

Machine learning approaches to diagnosis and laterality effects in semantic dementia discourse

“…As an indicator of the relative importance of a feature to one class or the other, we also show the frequency of occurrence of each feature in the transcripts belonging to the two classes of interest. Consistent with earlier studies of discourse (Bird and Lambon Ralph, 2000; Garrard and Forsyth, 2010), the selected features that occur more often in NC transcripts consist of content words that appear in relatively low written and spoken frequency in published corpora (e.g., ‘picnic’, ‘blanket’, ‘sailing’ and ‘pail’), while those occurring more often in SD scripts include generic terms (e.g., ‘thing’, ‘something’) and components of metanarrative statements (e.g., ‘you’, ‘know’ and ‘remember’). The features found to be relevant to the L > R versus R > L distinction were fewer in number, but generally more indicative of one category over the other, from which it would be predicted that using selected features would result in a greater improvement in the classifiers’ performance on this distinction.…”

“…The analyses reported in this paper support the idea that clinically relevant distinctions can be derived from samples of discourse using nothing more than the lexical frequency data inherent in the transcripts, as previously proposed by Garrard and Forsyth (2010).…”

“…Garrard and Forsyth (2010) hypothesized that the language of SD patients and controls would be distinguishable on the basis of the lexical frequency data from transcripts of connected speech. They used principal components analysis (PCA) to identify two latent variables in a high dimensional ‘discourse space’, the values of which distinguished between connected speech samples obtained from patients with SD and those produced by controls.…”

Logopenic aphasia in Alzheimer's disease: clinical variant or clinical feature?

“…For exactly these reasons, the numbers per group included in the studies of Wilson et al 9 and others5 6 38 39 were of similar magnitude. Second, if an excess of semantic errors had been documented, then the profile obtained for the patients in our sample would have remained distinct from any of the PPA profiles.…”

Abnormalities of connected speech in semantic dementia vs Alzheimer's disease

“…This reductionist approach is then used to draw inferences about the nature of connected speech. For instance, in the case of semantic dementia (SD) anomia is readily identifiable by formal testing (Adlam et al, 2006;Gorno-Tempini et al, 2004;Hodges & Patterson, 2007;Hodges, Patterson, Oxbury, & Funnell, 1992;Neary et al, 1998), and this is often construed to mean that connected speech will be characterised by word-finding difficulty and will therefore contain many circumlocutions and/or word-finding pauses (Garrard & Forsyth, 2010;Harciarek & Kertesz, 2011;Knibb & Hodges, 2005;Ratnavalli, 2010;Rogalski et al, 2011;Seelaar, Rohrer, Pijnenburg, Fox, & van Swieten, 2011;Wilson et al, 2010). Connected speech is typically elicited during clinical consultations but not subjected to any formal analysis.…”