TOWARDS INTERPRETABILITY OF SPEECH PAUSE IN DEMENTIA DETECTION USING ADVERSARIAL LEARNING.

Area of Science:

• Computational linguistics
• Neurology
• Artificial Intelligence

Background:

• Speech pauses are emerging biomarkers for dementia detection.
• Deep learning models show high accuracy but lack interpretability regarding pause impact.

Purpose of the Study:

• To investigate the influence of speech pause position and length on dementia detection.
• To enhance the interpretability of deep learning models for Alzheimer's Disease (AD) detection.

Main Methods:

• Utilized adversarial attack and training approaches to perturb speech pauses in testing and training samples.
• Examined model accuracy, pause context (e.g., proximity to specific verbs), and pause length variations.

Main Results:

• Identified specific speech pauses (e.g., near the verb "is") as highly sensitive to dementia detection.
• Increasing sensitive pause length or presence shifted model inference towards AD.
• Decreasing sensitive pause length or absence shifted inference towards non-AD.

Conclusions:

• Speech pause characteristics are crucial for accurate and interpretable dementia detection.
• Adversarial learning effectively reveals dementia-sensitive pause features.
• Findings can improve AI-driven diagnostic tools for Alzheimer's Disease.