LSTM-Based Autoencoders in Online Handwriting Data Augmentation and Preprocessing

Abstract

On-line handwriting analysis is a research field that is among others used in assessment of handwriting difficulties (HD), which can be manifestations of degenerative brain diseases such as Parkinson’s disease in the elderly, or developmental dysgraphia in children. Using advanced modelling approaches or artificial intelligence is often difficult because of the limited data availability in both demographic cohorts. In this article, a data processing approach, using LSTM-based autoencoders, is described as a way of augmenting the database with semisynthetic data or preprocessing the data to improve the performance of feature-based classification. The proposed method has led to a 3 percentage point increase in classification accuracy when compared to baseline. While the improvement is marginal, it highlights another possible area of research to improve the efficacy of automated HD assessment.