To test a low-cost system for obtaining continuous quantitative measurements of movements of people with Parkinson’s disease (PD), investigators attached accelerometers to the upper and lower extremities of twenty patients with Parkinson’s disease, one patient with multiple system atrophy (MSA) (participant 4), and eight age- and sex-matched healthy controls (HC) who performed twelve tasks modified from the Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS). Study participants were connected to four accelerometers and underwent twelve videotaped procedures. Videotapes were rated live by a trained examiner and scored according to the Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS). At least one week after the initial test, ten patients with Parkinson's disease and eight healthy controls underwent a retest. The patient with MSA completed only a test session. Participant 14 completed two re-tests due to an incomplete initial test: the first retest served as the test session and second retest served as the retest. Participant 16 did not complete any testing. The dataset includes the following for each participant: age and health status, the coding form with MDS-UPDRS scores, raw files, converted spreadsheets, and the fast Fourier and continuous wavelet transforms.

Data analysis has focused the five repetitive tasks (finger tapping, hand movements, pronation-supination movements of the hands, toe tapping, and leg agility). This data was also utilized for a deep learning application (image classification) using GoogLeNet in MATLAB with the Deep Learning Toolbox. The images of fast Fourier transforms and continuous wavelet transforms were divided into two classes: low (corresponding to MDS-UPDRS scores of 0 and 1) and high (corresponding to scores of 3 and 4). An equal number of images was selected randomly from each category for classification and split 80-20 into training and testing sets. To reduce overfitting, the investigators incorporated a dropout layer. The network was trained with stochastic gradient descent with momentum optimizer and executed in a CPU environment. The network’s ability to correctly classify the test images was used to determine an accuracy score.