Automated classification of primary progressive aphasia subtypes from narrative speech transcripts

Area of Science:

• Neurolinguistics
• Computational Linguistics
• Artificial Intelligence

Background:

• Early-stage neurodegenerative disorders often present subtle language deficits undetectable by standard tests.
• Connected speech analysis offers insights into language capacities but is traditionally time-intensive.

Purpose of the Study:

• To develop and validate a computational method for evaluating and classifying connected speech in primary progressive aphasia (PPA).
• To differentiate between semantic dementia (SD), progressive nonfluent aphasia (PNFA), and healthy controls using automated speech analysis.

Main Methods:

• Automated extraction of syntactic and semantic features from narrative speech transcriptions.
• Training and testing machine learning classifiers on these features for binary classification tasks.
• Analysis of linguistically significant features differentiating patient groups from controls and each other.

Main Results:

• Machine learning classifiers achieved accuracies significantly above baseline for distinguishing between SD, PNFA, and controls.
• Patients with PPA used higher-frequency words (nouns for SD, verbs for PNFA) compared to controls.
• SD patients exhibited increased word familiarity and altered use of word categories (fewer nouns, more demonstratives/adverbs); PNFA patients had slower speech and shorter words.

Conclusions:

• Computational analysis of connected speech provides an effective and efficient method for early detection and classification of PPA subtypes.
• Distinct linguistic features in connected speech can reliably differentiate between SD, PNFA, and healthy individuals.
• This approach holds promise for improving diagnostic accuracy and understanding language changes in neurodegenerative diseases.