Graph-BrainConvNet: A One-class GCN-based approach for MCI detection from source-level MEG

Mild cognitive impairment (MCI) is a neurocognitive disorder that increases the risk of dementia in older age, making early detection crucial for subsequent risk reduction strategies. Existing studies with advanced neuroimaging techniques mostly focused on finding MCI-related neuromarkers either during resting state (without evaluating their cognitive ability) or during complex cognitive tasks, which can be challenging for non-MCI geriatrics as well. This paper introduces a novel feasibility experiment to identify neural correlates related to attention from MCI patients' while watching a silent movie with random auditory distractions. A novel one-class graph convolutional neural network (GCN)-based model is proposed to detect MCI neuromarkers using source-level functional brain connectivity and spectral-temporal features, achieving a classification accuracy of 92.63±0.96% between control participants (not diagnosed with MCI) and MCI for the gamma rhythm. MCI patients showed higher responses to auditory distractors than control participants (CPs), indicating cognitive decline. Our experiment reveals potential auditory gating deficits in MCI, which the proposed GCN model can capture for early diagnosis.