Using Deep Learning Neural Networks to Improve Dementia Detection: Automating Coding of the Clock-Drawing Test

Area of Science:

• Neurology
• Artificial Intelligence
• Public Health

Background:

• Alzheimer's disease and related dementias (ADRD) pose a significant public health challenge.
• The clock-drawing test (CDT) is a common screening tool for ADRD, but manual coding introduces potential bias.
• Large-scale studies require efficient and unbiased methods for analyzing CDT results.

Purpose of the Study:

• To develop and evaluate an automated system for scoring the CDT using Deep Learning Neural Networks (DLNN).
• To compare the performance of different DLNN models (ResNet101, EfficientNet, Vision Transformers - ViT) for CDT image analysis.
• To introduce and assess an ordinal scoring system for CDT, moving beyond traditional nominal classification.

Main Methods:

• Utilized a large dataset of CDT images from the National Health and Aging Trends Study (NHATS) (2011-2019).
• Implemented and compared three DLNN architectures: ResNet101, EfficientNet, and Vision Transformers (ViT).
• Developed an ordinal scoring system (0-5) and compared DLNN-based coding against manual coding standards.

Main Results:

• Vision Transformers (ViT) demonstrated superior performance compared to ResNet101 and EfficientNet in CDT scoring.
• The ViT-based system outperformed manual coding in accuracy and consistency.
• The ordinal coding system effectively minimized under- or over-estimation errors in scoring.

Conclusions:

• An AI-powered system using ViT offers an accurate and efficient method for automated CDT scoring.
• This automated approach reduces bias and improves reliability in ADRD screening through CDT analysis.
• The developed ViT-coding system is now integrated into the NHATS study for annual CDT coding, enhancing ADRD research.