Ink of Insight: Data Augmentation for Dementia Screening through Handwriting Analysis

Abstract

We investigate the use of handwriting data as a means of predicting early symptoms of Alzheimer's disease (AD). Thirty-six subjects were classified based on the standardized pentagon drawing test (PDT) using deep learning (DL) models. We also compare and contrast classic machine learning (ML) models with DL by employing different data augmentation (DA) techniques. Our findings indicate that DA greatly improves the performance of all models, but the DL-based ones are the ones that achieve the best and highest results. The best model (EfficientNet) achieved a classification accuracy of 87% and an area under the receiver operating characteristic curve (AUC) of 91% for binary classification (healthy or AD patients), whereas for multiclass classification (healthy, mild AD, or moderate AD) accuracy was 76% and AUC was 77%. These results underscore the potential of DA as a simple, cost-effective approach to aid practitioners in screening AD in larger populations, suggesting DL models are capable of analyzing handwriting data with a high degree of accuracy, which may lead to better and earlier detection of AD.tempate