Automated analysis of 40 linguistic and acoustic features from 1–2 minute picture-description speech recordings was evaluated in 214 participants (43 controls, 50 nonfluent PPA, 56 logopenic PPA, 65 semantic PPA) to classify PPA clinical variants, identify neuroanatomical correlates, and predict autopsy-confirmed neuropathology.
Lasso multinomial logistic regression using 4–8 features per variant achieved AUC 0.90 (95% CI 0.84–0.97) for variant classification, replicated in external validation (AUC 0.90; 95% CI 0.83–0.97), and discriminated underlying neuropathology in an autopsy-confirmed subset (n=56) with AUC 0.90 (95% CI 0.80–0.96); speech profile scores mapped onto established neuroanatomical patterns for each variant.
Cross-sectional single-referral-center design limits generalizability; autopsy-confirmed neuropathological data were available in only 64 of 171 PPA participants (56 with the most common pathology analyzed); full paper text was not accessible for review of feature-level details.
A short, automated voice recording task (~1–2 min) may reliably classify PPA variants and predict underlying pathology with high accuracy, offering a scalable alternative to specialist speech-language assessments. Clinicians in settings without specialized expertise could use this tool to support differential diagnosis and longitudinal monitoring of PPA.
Explore related topics
