Towards Cognitive Impairment Screening in Elderly Communities with Audio-Visual Modal Disentangled Representation...

Area of Science:

• Gerontology
• Artificial Intelligence
• Neuroscience

Background:

• Alzheimer's disease (AD) poses a global health challenge, with early diagnosis crucial for effective intervention.
• Conventional screening methods are resource-intensive and impractical for widespread community use.
• AI-driven behavioral analysis, including speech and facial recognition, shows promise for non-invasive cognitive assessment.

Purpose of the Study:

• To develop a community-oriented intelligent screening system for cognitive impairment in elderly populations.
• To introduce CIR-AV, a large-scale Mandarin multimodal dataset for cognitive impairment recognition.
• To propose DiVA, a novel audio-visual fusion framework for enhanced cognitive assessment.

Main Methods:

• Development of CIR-AV, a multimodal dataset with facial and speech data from 574 Chinese older adults.
• Proposal of the DiVA framework, employing disentangled representation learning and cross-modal attention for audio-visual fusion.
• Implementation of a trajectory-constrained mechanism and a cross-modal attention-based dynamic fusion (CMF) module.

Main Results:

• DiVA achieved an Area Under the Curve (AUC) of 78.66% at the segment level.
• DiVA demonstrated an accuracy of 79.46% at the subject level.
• The proposed system significantly outperformed existing state-of-the-art methods in cognitive impairment recognition.

Conclusions:

• The DiVA framework provides a cost-efficient and scalable solution for large-scale community screening of cognitive impairment.
• This AI-driven approach offers a practical method for early dementia detection, especially in resource-limited settings.
• The system holds significant social and economic value for proactive public health initiatives.