Clinical Manifestations.

Area of Science:

• Neurolinguistics
• Computational Linguistics
• Clinical Neurology

Background:

• Distinguishing between logopenic (lvPPA) and nonfluent/agrammatic (nfvPPA) Primary Progressive Aphasia (PPA) variants is challenging, often requiring expert analysis of speech and language.
• Simple fluency/non-fluency is insufficient for diagnosis, as sentence production effort and errors occur in both variants for distinct reasons.
• Connected speech analysis offers a robust method for phenotypical classification of PPA by targeting abnormal language production patterns.

Purpose of the Study:

• To investigate the utility of computational analysis of connected speech for differentiating Italian-speaking patients with nfvPPA and lvPPA.
• To identify specific linguistic features that distinguish between the nfvPPA and lvPPA variants.
• To develop and evaluate a machine-learning model for automated classification of PPA variants based on speech data.

Main Methods:

• A cross-sectional study involving 19 Italian PPA patients (9 nfvPPA, 10 lvPPA) performing a picture description task.
• Audio recordings of speech data were analyzed using Computerized Language ANalysis (CLAN) software to extract linguistic features.
• A Mann-Whitney U test with FDR correction analyzed 45 linguistic features across four levels, and a machine-learning model was trained for classification.

Main Results:

• Ten linguistic features across phonetic, lexico-semantic, morphosyntactic, and pragmatic/discourse levels significantly differentiated nfvPPA from lvPPA.
• Key differentiating features included silent pause ratio, ratios of nouns, adverbs, articles, determiners, total utterances, utterance error ratio, total words, total morphemes, and idea density.
• The machine-learning model achieved over 90% sensitivity and specificity in classifying nfvPPA versus lvPPA.

Conclusions:

• Natural language processing and machine learning can effectively classify Italian PPA variants based on connected speech samples.
• This computational approach, applied for the first time to Italian PPA, successfully identified key linguistic markers for differentiating nfvPPA and lvPPA.
• The findings highlight the potential of automated speech analysis for improving the diagnostic accuracy and phenotypical classification of PPA.