Deep-learning-optimized microstate network analysis for early Parkinson’s disease with mild cognitive impairment

Abstract

Graph-theory-based topological impairment of the whole-brain network has been verified to be one of the characteristics of mild cognitive impairment (MCI). However, two major challenges impede the further understanding of topological features for the personalized functional connectivity network of early Parkinson’s disease (ePD) with MCI. The uncertain of characteristic frequency band reflecting the abnormality of ePD-MCI and the setting of fixed length of sliding window at a second level in the construction of conventional brain network both limit a deeper exploration of network characteristics for ePD-MCI. Thus, a convolutional neural network is constructed first and the gradient-weighted class activation mapping method is used to determine the characteristic frequency band of the ePD-MCI. It is found that 1–4 Hz is a characteristic frequency band for recognizing MCI in ePD. Then, we propose a microstate window construction method based on electroencephalography microstate sequences to build brain functional network. By exploring the graph-theory-based topological features and their clinical correlations with cognitive impairment, it is shown that the clustering coefficient, global efficiency, and local efficiency of the occipital lobe significantly decrease in ePD-MCI, which reflects the low degree of nodes interconnection, low efficiency of parallel information transmission and low communication efficiency among the nodes in the brain network of the occipital lobe may be the neural marker of ePD-MCI. The finding of personalized topological impairments of the brain network may be a potential characteristic of early PD-MCI.