BackgroundPrimary progressive aphasia (PPA) defines a group of neurodegenerative disorders characterised by language decline. Three PPA variants correlate with distinct underlying pathologies: semantic variant PPA (svPPA) with transactive response DNA-binding protein of 43 kD (TDP-43) proteinopathy, agrammatic variant PPA (agPPA) with tau deposition and logopenic variant PPA (lvPPA) with Alzheimer’s disease (AD). Our objectives were to differentiate PPA variants using clinical and neuroimaging features, assess progression and evaluate structural MRI and a novel 18-F fluorodeoxyglucose positron emission tomography (FDG-PET) image decomposition machine learning algorithm for neuropathology prediction.MethodsWe analysed 82 autopsied patients diagnosed with PPA from 1998 to 2022. Clinical histories, language characteristics, neuropsychological results and brain imaging were reviewed. A machine learning framework using ak-nearest neighbours classifier assessed FDG-PET scans from 45 patients compared with a large reference database.ResultsPPA variant distribution: 35 lvPPA (80% AD), 28 agPPA (89% tauopathy) and 18 svPPA (72% frontotemporal lobar degeneration-TAR DNA-binding protein (FTLD-TDP)). Apraxia of speech was associated with 4R-tauopathy in agPPA, while pure agrammatic PPA without apraxia was linked to 3R-tauopathy. Longitudinal data revealed language dysfunction remained the predominant deficit for patients with lvPPA, agPPA evolved to corticobasal or progressive supranuclear palsy syndrome (64%) and svPPA progressed to behavioural variant frontotemporal dementia (44%). agPPA-4R-tauopathy exhibited limited pre-supplementary motor area atrophy, lvPPA-AD displayed temporal atrophy extending to the superior temporal sulcus and svPPA-FTLD-TDP had severe temporal pole atrophy. The FDG-PET-based machine learning algorithm accurately predicted clinical diagnoses and underlying pathologies.ConclusionsDistinguishing 3R-taupathy and 4R-tauopathy in agPPA may rely on apraxia of speech presence. Additional linguistic and clinical features can aid neuropathology prediction. Our data-driven brain metabolism decomposition approach effectively predicts underlying neuropathology.