TMemAE: Handwriting Abnormality Analysis for Parkinson Disease

Early detection is essential for the treatment of Parkinson Disease(PD). Parkinson Disease affects muscle control and habitual activity, and affects the way patients write and paint. Therefore, handwriting can be used to detect whether an individual is sick. Traditional handwritten analysis methods are limited by subjectivity and efficiency, and it is difficult to meet the needs of large-scale and high-precision disease detection. At present, the handwriting analysis methods based on deep-learning are based on shallow convolutional neural networks (CNNs) to classify the sample data. However, due to the limited negative samples (i.e., the handwriting samples of Parkinson patients), the generalization of these classification models has not been proved. Therefore, this paper turns to the field of unsupervised anomaly detection. MemAE performs well in this field, but its use of multi-layer convolution to extract features will lose the understanding of the continuity of handwritten features, and Transformer is good at extracting sequence features. Therefore, this paper combines Transformer with MemAE and proposes the Serialized Memory-augmented Module (SMM), which can better capture and reproduce dynamic patterns in the sequence data. Our method is named Transformer Memory-augmented Autoencoder (TMemAE), which can reconstruct normal handwritten samples with high quality and reconstruct handwritten samples of Parkinson patients with large errors to achieve the purpose of Parkinson Disease detection. The experimental results show that the method achieves the best performance without negative sample training.