On Extracting Digitized Spiral Dynamics' Representations: A Study on Transfer Learning for Early Alzheimer's...

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Early detection of Alzheimer's disease (AD) is crucial for timely intervention.
• Traditional diagnostic methods can be invasive or costly.
• Dynamic gesture analysis offers a potential non-invasive biomarker.

Purpose of the Study:

• To develop and validate a decision-aid tool for early Alzheimer's disease detection.
• To assess the efficacy of using Archimedes spiral drawing tasks for AD diagnosis.
• To optimize feature extraction and classification methodologies using transfer learning.

Main Methods:

• Utilized a Wacom digitizer to record Archimedes spiral trajectories.
• Extracted kinematic time functions (pressure, altitude, velocity) from spiral data.
• Applied transfer learning for automatic feature extraction from trajectory images.
• Employed deep learning models and decision fusion techniques for classification.
• Experimented on a cohort of 30 Alzheimer's disease patients and 45 healthy controls.

Main Results:

• Extracted kinematic features significantly improved sensitivity and accuracy over raw images.
• Intermediate-level features from deep networks demonstrated the highest discriminant capabilities.
• Decision fusion of classifiers trained on optimal features achieved 84% sensitivity and 81.5% accuracy.
• Demonstrated an absolute improvement of 22% in sensitivity and 7% in accuracy compared to baseline methods.

Conclusions:

• The Archimedes spiral drawing task shows significant potential for early Alzheimer's disease detection.
• Transfer learning and kinematic feature analysis provide an effective methodology for AD diagnosis.
• The proposed decision-aid tool offers a promising, non-invasive approach for clinical application.