Domain generalization for voice-based cognitive impairment detection.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Computational Linguistics

Background:

• Voice biomarkers show promise for early detection of cognitive disorders.
• Variations in recording environments challenge AI model accuracy for cognitive impairment diagnosis.
• Developing generalizable AI models is crucial for reliable detection across diverse datasets.

Purpose of the Study:

• To develop a robust and generalizable AI model for diagnosing cognitive impairments.
• To overcome challenges posed by varied recording conditions in voice biomarker data.
• To enhance the reliability of AI-driven cognitive disorder detection.

Main Methods:

• Implemented a domain generalization approach using an adapted Deep Domain-Adversarial Image Generation (DDAIG) framework.
• Transformed input data to minimize center-specific characteristics and emphasize domain-invariant features.
• Focused on features indicative of cognitive impairment for improved model generalization.

Main Results:

• Cognitive impairment (CI) classification accuracy was 0.90 after domain generalization.
• Center classification accuracy dropped from 0.96 to 0.64, indicating reduced dependence on site-specific data.
• The reduction in center classification metrics demonstrates effective domain generalization.

Conclusions:

• The adapted DDAIG framework successfully reduced center-specific learning in AI models.
• Enhanced generalization of cognitive impairment classification across different data centers was achieved.
• Domain generalization is vital for creating reliable AI diagnostic tools for cognitive disorders.