t-SNE applied to discriminate healthy individuals from those with Parkinson’s disease executing motor tasks detected by non-contact capacitive sensors

Abstract

The diagnosis and evaluation of Parkinson’s disease (PD) is a task that has been performed through clinical evaluation and subjective scales. Over the years several studies have reported results and technologies with the purpose of making the follow-up of PD more objective. Usually, in the objective evaluation, inertial and electromyographic sensors are employed for recording movement and muscular activation. A major challenge that exists in the area is related to the monitoring of the technological horizon, to identify and incorporate new technologies and methods that can be used for the evaluation of PD. In this perspective, it was proposed in this research the use of non-contact capacitive sensors to record four motor activities of the hand and wrist (i.e., radial deviation, ulnar deviation, flexion and extension). Another identified challenge is related to the correct classification of individuals with PD. To accomplish this, it makes necessary the use of tools for signal processing and machine learning. In this study, features related to amplitude and time of the signal were estimated and then combined by means of t-Distributed Stochastic Neighbor Embedding (t-sne), which is an innovative tool for dimensionality reduction and visualization of information. Experimental data were collected from a neurologically healthy individual and one with PD. The use of t-sne allowed for the visualization of data on a two-dimensional. The use of non-contact capacitive sensors introduces an innovative way to measure information from people with PD. Furthermore, the application of t-SNE showed to be a successful tool for the discrimination between a healthy individual from that with PD.