Machine learning for EEG-based biomarkers in Parkinson’s disease

Parkinson's disease (PD) is a neurological disorder characterized by motor deficits as a result of progressive degeneration of dopaminergic neurons. Patients often report visual impairments and exhibit deficits of visual acuity, contrast sensitivity and color vision, suggesting abnormal processing of visual information. Although electrophysiological studies indicate amplitude and latency differences in patients, one of the major challenges is the lack of a conclusive, accurate, non-invasive biomarker for diagnosis and monitoring of therapeutic efficacy. In this study, we used machine learning to model and identify the most relevant signs from the EEG spectra over the span of visual stimulation. First, we used human electroencephalography (EEG) to measure visual evoked responses during steady state (ssVEP) in a visual surround suppression paradigm, in which measures of gain control and temporal aspects of the visual response are simultaneously tracked. We found an excessive response gain profile of the contrast response function, and an abnormally greater background noise level in PD patients compared to healthy controls. Second, performance of three learning models was tested: logistic regression, decision tree and extra tree. Together, these provide powerful insights about the most relevant features that allow classifying patients from controls. These models contribute with new hints on the interpretation of functional abnormalities in PD based on spectral EEG during visual stimulation. Machine learning provides a promising tool for accurate and robust prediction in PD diagnosis, based on EEG signatures during visual information processing.