Advancing Dementia Screening Through Handwriting Analysis and Data Augmentation

Early detection of dementia, particularly Alzheimer’s Disease (AD) and Mild Cognitive Impairment (MCI), remains a significant challenge in healthcare. This dissertation investigates handwriting-based cognitive assessments as a viable alternative, leveraging off-line (scanned images) and on-line (digitally captured) drawing tasks to enhance classification accuracy through deep learning (DL), data augmentation, and transfer learning.

One of the central contributions of this work is the study of dataset size requirements for AI-based dementia screening (The Magic Number), demonstrating that EfficientNet achieves reliable performance with only half of the available data. These findings challenge the assumption that large-scale datasets are indispensable for robust screening. Another key aspect involves a deep feature concatenation (DFC) framework (Better Together), which integrates multiple handwriting sources—pentagon drawings, sentences, and signatures—leading to a classification improvement from 60% (single-source) to 80% (multi-source with augmentation).

A comparative assessment of drawing modalities (Blueprint of Tomorrow) establishes that off-line handwriting analysis provides better results than on-line methods, with EfficientNet achieving 90% accuracy in binary classification. This investigation extends to two studies focused on the House Drawing Test (HDT), where one examines off-line and on-line representations independently, and the other introduces an approach that converts and refines on-line data into off-line format. This method yields 82% accuracy (86% AUC), reinforcing the suitability of static image-based models for automated analysis.

This investigation extends to two studies focused on the House Drawing Test (HDT), where one examines off-line and on-line representations independently (Predicting Alzheimer’s Disease and Mild Cognitive Impairment), and the other introduces an approach that converts and refines on-line data into off-line format (Screening of Alzheimer's Disease). The latter achieves 82% accuracy (86% AUC), reinforcing the suitability of static image-based models for automated analysis.

This dissertation also presents the first evaluation of data augmentation techniques for handwriting-based dementia detection (Ink of Insight). The effects of augmentation strategies are analyzed across classical machine learning (SVM, RF, k-NN) and DL architectures, with EfficientNet achieving 87% accuracy (91% AUC). Further investigation (Efficient Automatic Data Augmentation) explores automated augmentation methods, demonstrating that non-learnable techniques (TrivialAugment, UniformAugment) improve generalization by up to 15% with minimal computational demands.

These studies establish handwriting analysis as a scalable, non-invasive, and cost-effective tool for dementia screening. By refining dataset requirements, exploring augmentation strategies, and integrating multiple handwriting sources, this research advances AI applications in neurodegenerative disease diagnostics and lays the groundwork for further developments in computational healthcare.