Automated differentiation of non-fluent and logopenic primary progressive aphasia in Italian speakers using acoustic and linguistic speech measures.

INTRODUCTION: Differentiating the nonfluent/agrammatic and logopenic variants of primary progressive aphasia (PPA; nfvPPA and lvPPA, respectively) remains clinically challenging due to overlapping subtle speech and language impairments. We investigated whether automated connected speech analysis can support differential diagnosis. METHODS: We analyzed connected speech from 84 Italian speakers with PPA (23 nfvPPA, 23 semantic variant [svPPA], and 38 lvPPA) using a picture-description task. Prosodic, phonological, and morphosyntactic features were extracted. Machine-learning classifiers were trained for binary (lvPPA vs. nfvPPA) and multiclass (lvPPA, nfvPPA, svPPA) classification. Explainability analyses identified key features. RESULTS: The combination of speech and language features effectively discriminated among PPA variants. Binary classification reached 81.67% accuracy, while multiclass classification reached 63.86% accuracy. Noun rate, local jitter, articulation rate, and total pauses were among the most informative features. DISCUSSION: Automated analysis of connected speech provides objective linguistic biomarkers that enhance diagnostic accuracy and support reliable differentiation of PPA variants in clinical practice.