Discrimination of Healthy Controls and Patients with Mild Cognitive Impairment based on Time‐Varying Functional Connectivity Analysis of Resting‐State EEG

Abstract

AbstractBackgroundDeveloping economically viable assessment tools that are highly sensitive to cognitive decline and neural dysfunction is critical for the study of neurodegenerative mechanisms and interventions to promote cognitive resiliency. Due to its high time resolution, accessibility, and affordability, EEG‐based detection of cognitive impairment is attracting considerable attention from the research community.MethodOur research is based on resting‐state EEG (eye‐closed, 64‐channel), and the current dataset includes 137 consensus‐diagnosed, community‐dwelling African Americans (ages 60‐90 years, 84 Normal Cognition, NC; 53 Mild Cognitive Impairment, MCI) recruited through the Wayne State Institute of Gerontology Healthy Black Elders Center and the Michigan Alzheimer’s Disease Research Center with subjective cognitive complaints.We first conducted joint time‐frequency‐spatial analysis on the time‐varying functional connectivity of all the possible pairs between the selected EEG regions of interest and generated the feature vectors for each subject. The selected features were then fed into a machine learning algorithm for the discrimination of NC and MCI. By tuning the observing window size (in time‐domain) and the number of features used, we obtained a series of different configurations of the discrimination model between NC and MCI, each with its own accuracy. The overall result is a combined output of the model under different configurations.ResultThe accuracy of the discrimination model was calculated under different window sizes and numbers of features. Our result shows that there are 11 window‐size‐and‐feature configurations which can achieve an accuracy in the range of 80.29%∼86.86%. We selected 21 configurations as voters and get the final discrimination result through majority voting (+1 for NC and ‐1 for MCI). Moreover, we obtained a soft output (i.e., a likelihood score) for each subject by combining all the votes which were weighted using the accuracy of each voter. For the 137‐sample dataset we are working with, the 10‐fold cross‐validation accuracy of the model is 91.97%.ConclusionOur analysis indicates that EEG‐based time‐varying function connectivity analysis is a promising technique for early detection of MCI. Potentially, the soft output we obtained for each subject might be used for prediction of possible cognitive decline.