Machine learning in the clinical and language characterisation of primary progressive aphasia variants

Primary progressive aphasia (PPA) is a clinical syndrome of neurodegenerative origin with 3 main variants: non-fluent, semantic, and logopenic. However, there is some controversy about the existence of additional subtypes. Our aim was to study the language and cognitive features associated with a new proposed classification for PPA.

Sixty-eight patients with PPA in early stages of the disease and 20 healthy controls were assessed with a comprehensive language and cognitive protocol. They were also evaluated with ¹⁸F-FDG positron emision tomography (PET). Patients were classified according to FDG PET regional metabolism, using our previously developed algorithm based on a hierarchical agglomerative cluster analysis with Ward's linkage method. Five variants were found, with both the non-fluent and logopenic variants being split into 2 subtypes. Machine learning techniques were used to predict each variant according to language assessment results.

Non-fluent type 1 was associated with poorer performance in repetition of sentences and reading of irregular words than non-fluent type 2. Conversely, the second group showed a higher degree of apraxia of speech. Patients with logopenic variant type 1 performed more poorly on action naming than patients with logopenic type 2. Language assessments were predictive of PET-based subtypes in 86%-89% of cases using clustering analysis and principal components analysis.

Our study supports the existence of 5 variants of PPA. These variants show some differences in language and FDG PET imaging characteristics. Machine learning algorithms using language test data were able to predict each of the 5 PPA variants with a relatively high degree of accuracy, and enable the possibility of automated, machine-aided diagnosis of PPA variants.