Bag of Samplings for computer-assisted Parkinson's disease diagnosis based on Recurrent Neural Networks

Abstract

Parkinson's Disease (PD) is a clinical syndrome that affects millions of people worldwide. Although considered as a non-lethal disease, PD shortens the life expectancy of the patients. Many studies have been dedicated to evaluating methods for early-stage PD detection, which includes machine learning techniques that employ, in most cases, motor dysfunctions, such as tremor. This work explores the time dependency in tremor signals collected from handwriting exams. To learn such temporal information, we propose a model based on Bidirectional Gated Recurrent Units along with an attention mechanism. We also introduce the concept of “Bag of Samplings” that computes multiple compact representations of the signals. Experimental results have shown the proposed model is a promising technique with results comparable to some state-of-the-art approaches in the literature.