Differentiating non-fluent and logopenic primary progressive aphasia via automated acoustic speech analysis

In primary progressive aphasia, differential diagnosis between non-fluent (nfvPPA) and logopenic (lvPPA) variants can be clinically challenging. Although linguistic features from connected speech may aid in PPA subtype classification, they depend on labor-intensive transcriptions. In contrast, acoustic features are extracted directly from the audio signal and have shown potential in capturing variant-specific patterns, but their diagnostic relevance is not well understood. In this study, we analyzed connected speech recordings from 242 individuals (40 healthy controls, 91 lvPPA, 111 nfvPPA) performing a ∼2-min picture description task. We extracted 103 acoustic features from each recording and reduced them to 9, which we finally grouped into four domain scores representing speech rate and voice quality. Domain scores were tested as predictors in classification models to detect PPA and distinguish between PPA variants. Voxel-based morphometry was used to investigate brain-behavior relationships. Classification models combining the four domain scores and MMSE yielded 88% balanced-sample accuracy for PPA-versus-controls and 73% accuracy for lvPPA-versus-nfvPPA. Pitch and speech rate were the most discriminative markers. Voxel-based morphometry associated speech rate with the frontal speech production network, pitch with the middle frontal and inferior frontal gyrus, perturbation with the insula, and spectrum with the basal ganglia and the thalamus. We showed that acoustic analysis of brief speech samples has the potential to aid the diagnostic process and contribute to more nuanced phenotyping. With further validation, such measures could be integrated into clinical workflows as low-burden, objective tools to support differential diagnosis and track disease progression in PPA.