Shannon wavelet entropy-based machine learning applications in Parkinson's disease diagnosis with videonystagmography

Abstract

Parkinson's disease (PD) patients exhibit alterations in saccadic eye movements due to dysfunction in the basal ganglia-thalamocortical circuitry, which regulates movement control. This study examines the usability of the videonystagmograph (VNG) as an alternative non-invasive diagnostic tool in the early stage detection of PD with the approach of machine learning methods. In this study, a database consisting of 100 participants (50 PD and 50 healthy controls (HC)) was created with a balanced distribution in terms of age range, average age and gender. Participants were asked to follow a stimulus that appeared randomly on screen of VNG system for approximately 42 seconds. To eliminate this randomness in the time series patterns, a new data set was created by taking the difference between the reference signal of the stimulus signal and the participants' eye movement signals during the tests. Shannon wavelet entropy was employed to calculate the signal entropy. Leave-one-subject-out (LOSO) cross-validation scheme is used with k-nearest neighbors (kNN) classifiers. The kNN classifier with k=3 and Chebyshev distance achieved the acceptable patient-healthy classification performance with 84.0% accuracy, 82.7% sensitivity, 84.30% F-score and 0.68 Matthew's Correlation Coefficient scores. It can be concluded that the intersection of VNG and ML holds great promise for advancing the early detection of PD.