Generation of Synthetic Drawing Samples to Diagnose Parkinson’s Disease

The state-of-the-art artificial intelligence tools for automatic diagnosis of Parkinson’s disease from handwriting require a lot of training samples from both healthy subjects and patients to exhibit impressive performance. Publicly available datasets include very few samples drawn by a small number of individuals and that limits the use of deep learning architectures. In this paper, we evaluate if the performance of a Convolutional Neural Network that recognizes the handwriting of Parkinson’s disease patients can be improved by adding synthetic samples to the training set. In the experimentation, we synthetically generated dynamic signals of spirals and meanders through the use of a Recurrent Neural Network. The performance of the system was evaluated on the NewHandPD dataset and the results showed that the use of synthetic samples increases the recognition accuracy of the convolutional neural network.