Mild cognitive impairment detection from facial video interviews by applying spatial-to-temporal attention module

Area of Science:

• Neurology
• Artificial Intelligence
• Gerontology

Background:

• Early detection of Mild Cognitive Impairment (MCI) is crucial for timely interventions to prevent dementia progression.
• Deep Learning (DL) offers potential for non-invasive, low-cost MCI detection using accessible data like video recordings.
• Facial and interaction features in video can serve as biomarkers for cognitive status.

Purpose of the Study:

• To develop and evaluate a DL framework for detecting MCI in older adults using only facial features from home video recordings.
• To assess the efficacy of combining spatial and temporal facial information with interaction features for MCI detection.
• To determine the discriminating power of spatiotemporal facial features for MCI identification.

Main Methods:

• Utilized video-recorded conversations from the I-CONECT study (NCT02871921) involving socially isolated older adults.
• Developed a DL framework employing convolutional autoencoders for spatial facial features and transformers for temporal information.
• Introduced the Spatial-to-Temporal Attention Module (STAM) to integrate facial and interaction features for cognitive condition detection (MCI vs. Normal Cognition).

Main Results:

• The combined DL model incorporating facial and interaction features achieved a detection accuracy of 88% for MCI.
• Excluding segment and sequence information of facial features within themed video segments reduced accuracy to 84%.
• Spatiotemporal facial features demonstrated significant discriminating power for detecting MCI in older adults.

Conclusions:

• DL models analyzing spatiotemporal facial and interaction features from video are effective for non-invasive MCI detection.
• The proposed STAM framework shows promise for early identification of MCI, facilitating timely interventions.
• This approach offers a low-cost, accessible method for cognitive health monitoring in older populations.