Artificial intelligence classifies primary progressive aphasia from connected speech.

Area of Science:

• Neurology
• Computational Linguistics
• Artificial Intelligence

Background:

• Primary progressive aphasias (PPA) are neurodegenerative dementia syndromes.
• Diagnosis traditionally relies on cognitive profiles, with ongoing debate on PPA variant classification.
• Identifying distinctive linguistic features for PPA variant classification remains a challenge.

Purpose of the Study:

• To apply artificial intelligence (AI) and natural language processing (NLP) for unsupervised classification of PPA speech samples.
• To identify key linguistic features that differentiate PPA variants.
• To develop an accurate classifier for PPA variants and healthy controls.

Main Methods:

• Unsupervised classification of speech samples from 78 PPA patients using large language models.
• NLP analysis to identify dissociating linguistic features.
• Supervised classification incorporating identified linguistic features, including verb frequency, to classify PPA variants and controls.

Main Results:

• Large language models identified three distinct PPA clusters with 88.5% agreement with clinical diagnoses.
• Cortical atrophy patterns in data-driven clusters aligned with clinical diagnostic criteria.
• A classifier using 17 linguistic features, notably high- and low-frequency verbs, achieved 97.9% accuracy in classifying four groups (3 PPA variants, 1 control).

Conclusions:

• AI and NLP can effectively partition PPA patient speech into clinically relevant groups.
• Specific language features, particularly verb frequency, are robust indicators for PPA variant classification.
• These findings improve diagnostic accuracy and deepen the understanding of language processing in neurodegeneration.