Automatic speech analysis for detecting cognitive decline of older adults

Area of Science:

• Neurology
• Computational Linguistics
• Gerontology

Background:

• Speech analysis shows promise for early Alzheimer's disease (AD) and mild cognitive impairment (MCI) detection.
• Acoustic and linguistic features are commonly used, but their comparative effectiveness, especially in China's aging population, is unclear.

Purpose of the Study:

• To compare the screening effectiveness of acoustic, linguistic, and combined speech features for AD and MCI detection.
• To develop and validate an automated Chinese speech analysis model for cognitive impairment diagnosis.

Main Methods:

• 92 Chinese participants (AD, MCI, healthy controls) underwent speech tasks and cognitive screening (MoCA-B).
• Acoustic (pitch, jitter, MFCCs) and linguistic (POS tags, TTR) features were extracted.
• Machine learning algorithms (SVM, RF) were employed to compare classification accuracies.

Main Results:

• Linguistic features generally yielded higher accuracy than acoustic features in differentiating cognitive states.
• Support Vector Machines (SVM) achieved 80.77% accuracy for HC vs. AD using combined features.
• Random Forest (RF) achieved 80.43% accuracy for HC vs. AD/MCI using only linguistic features.

Conclusions:

• Linguistic features are highly effective for automated cognitive impairment diagnosis in Chinese speakers.
• Automated speech analysis is a validated tool for cognitive health assessment in the Chinese population.