Augmenting Dementia Cognitive Assessment With Instruction-Less Eye-Tracking Tests.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Neuroscience

Background:

• Dementia screening requires innovative and sensitive tools.
• Eye-tracking technology offers potential for objective cognitive assessment.
• Current methods face challenges with small patient cohorts and disease heterogeneity.

Purpose of the Study:

• To introduce a novel method for extracting salient features from raw eye-tracking data for dementia screening.
• To apply self-supervised representation learning for dementia classification.
• To compare self-supervised features against handcrafted features.

Main Methods:

• Utilized raw eye-tracking data from dementia patients and healthy individuals during an instruction-less cognitive test.
• Employed self-supervised representation learning with a deep neural network trained on a pretext task.
• Applied Layer-wise Relevance Propagation (LRP) for explainable AI insights.
• Compared novel self-supervised features with handcrafted features for dementia discrimination.

Main Results:

• Self-supervised learning features demonstrated higher sensitivity than handcrafted features in detecting performance differences.
• Identified novel oculomotor biomarkers indicative of dementia-related cognitive dysfunction.
• Validated the efficacy of instruction-less eye-tracking tests for dementia detection.

Conclusions:

• Self-supervised representation learning is a powerful technique for biomedical applications, particularly in challenging dementia research.
• Eye-tracking data, analyzed with advanced AI, can provide sensitive biomarkers for early dementia detection.
• This approach addresses limitations of traditional methods in dementia screening.