Efficient Pause Extraction and Encode Strategy for Alzheimer's Disease Detection Using Only Acoustic Features from...

Area of Science:

• Neurology
• Computational Linguistics
• Machine Learning

Background:

• Clinical studies indicate speech pauses correlate with cognitive differences in Alzheimer's Disease (AD) patients.
• The diagnostic potential of speech pause information for AD detection remains underexplored.

Purpose of the Study:

• To propose and validate a novel speech pause feature extraction and encoding strategy for acoustic-signal-based AD detection.
• To evaluate the efficacy of pause features compared to traditional acoustic features in classifying AD.

Main Methods:

• Developed a voice activity detection (VAD) method to extract and encode binary pause sequences from speech.
• Employed an ensemble machine learning approach utilizing VAD pause features alongside ComParE and eGeMAPS acoustic features.
• Validated the method on English (ADReSS, ADReSSo) and Chinese datasets.

Main Results:

• The VAD Pause feature demonstrated superior performance for AD classification compared to extensive ComParE and eGeMAPS feature sets.
• The ensemble method enhanced classification accuracy by over 5% compared to baseline approaches (e.g., 8% on ADReSS).
• The pause-sequence-based method achieved 80% accuracy on a local Chinese dataset.

Conclusions:

• Speech pause information holds significant potential for accessible and generalizable AD detection.
• The proposed VAD pause feature extraction and encoding strategy offers a promising avenue for early AD diagnosis.