Technology and Dementia Preconference.

Area of Science:

• Neuroscience
• Computational Linguistics
• Biomarker Discovery

Background:

• Alzheimer's disease (AD) presents diverse clinical forms requiring tailored care.
• Logopenic variant primary progressive aphasia (lvPPA) prognosis is challenging due to varied underlying pathologies (AD or FTLD).
• Accurate diagnosis necessitates multidisciplinary approaches including CSF analysis, imaging, and psychometric assessments.

Purpose of the Study:

• To evaluate speech markers for cross-classifying AD phenotypes with Cerebro-Spinal Fluid (CSF) profiles.
• To develop a non-invasive method for early prognosis of neurodegenerative disease progression.
• To differentiate clinical phenotypes and underlying pathologies in early-stage AD.

Main Methods:

• 42 patients were classified into lvPPA- (FTLD), lvPPA+ (AD), and AD groups.
• Speech recordings from a 14-sentence repetition task were analyzed for prosodic and temporal features.
• Supervised machine learning models (Random Forest, KNN, SVM) were trained and validated for classification.

Main Results:

• Random Forest model achieved 91% cross-validation accuracy, outperforming other models.
• Speech-based classification successfully differentiated lvPPA-, lvPPA+, and AD groups.
• Findings suggest speech analysis is a valuable non-invasive tool for early prognostic identification of AD phenotypes and pathology.

Conclusions:

• Speech-based classification links clinical phenotypes with CSF profiles, offering early prognostic insights.
• Enhanced prognostic precision can improve early intervention strategies for AD.
• Larger studies are needed to validate speech classification as a reliable non-invasive tool for differential prognosis.