Classification of healthy subjects and Alzheimer's disease patients with dementia from cortical sources of resting state EEG rhythms: A study using artificial neural networks

Previous evidence showed a 75.5% best accuracy in the classification of 120 Alzheimer’s disease (AD) patients with dementia and 100 matched normal elderly (Nold) subjects based on cortical source current density and linear lagged connectivity estimated by eLORETA freeware from resting state eyes-closed electroencephalographic (rsEEG) rhythms (Babiloni et al., 2016). Specifically, that accuracy was reached using the ratio between occipital delta and alpha1 current density for a linear univariate classifier (receiver operating characteristic curves). Here we tested an innovative approach based on an artificial neural network (ANN) classifier from the same database of rsEEG markers. Frequency bands of interest were delta (2-4 Hz), theta (4-8 Hz Hz), alpha1 (8-10.5 Hz), and alpha2 (10.5-13 Hz). ANN classification showed an accuracy of 77% using the most 4 discriminative rsEEG markers of source current density (delta/alpha1 and theta/alpha1 ratios in posterior cortical lobes). It also showed an accuracy of 72% using the most 4 discriminative rsEEG markers of source lagged linear connectivity (alphas between posterior cortical lobes). With these 8 markers combined, an accuracy of at least 76% was reached. Although these linear rsEEG markers of cortical activity and connectivity unveil different relevant neurophysiological mechanisms underpinning cortical arousal and vigilance in AD patients, they provide quite redundant information for classification purposes. Future AD studies should use ANNs combining the present markers with other linear (i.e. directed transfer function) and nonlinear rsEEG markers to improve the classification accuracy.