Prediction and classification of sleep quality based on phase synchronization related whole-brain dynamic
Zhongxing Zhou1, Biao Cai2, Gemeng Zhang2
1Biomedical Engineering Department, Tulane University, New Orleans, LA, United States; Tianjin University, School of Precision Instruments and Optoelectronics Engineering, Tianjin, China.
Abstract:
Recently, functional network connectivity (FNC) has been extended from static to dynamic analysis to explore the time-varying functional organization of brain networks. Nowadays, a majority of dynamic FNC (dFNC) analysis frameworks identified recurring FNC patterns with linear correlations based on the amplitude of fMRI time series. However, the brain is a complex dynamical system and phase synchronization provides more informative measures. This paper proposes a novel framework for the prediction/classification of behaviors and cognitions based on the dFNCs derived from phase locking value. When applying to the analysis of fMRI data from Human Connectome Project (HCP), four dFNC states are identified for the study of sleep quality. State 1 exhibits most intense phase synchronization across the whole brain. States 2 and 3 have low and weak connections, respectively. State 4 exhibits strong phase synchronization in intra and inter-connections of somatomotor, visual and cognitive control networks. Through the two-sample t-test, we reveal that for the group with bad sleep quality, state 4 shows decreased phase synchronization within and between networks such as subcortical, auditory, somatomotor and visual, but increased phase synchronization within cognitive control network, and between this network and somatomotor/visual/default-mode/cerebellar networks. The networks with increased phase synchronization in state 4 behave oppositely in state 2. Group differences are absent in state 3, and weak in state 1. We establish a prediction model by linear regression of FNC against sleep quality, and adopt a support vector machine approach for the classification. We compare the performance between conventional FNC and PLV-based dFNC with cross-validation. Results show that the PLV-based dFNC significantly outperforms the conventional FNC in terms of both predictive power and classification accuracy. We also observe that combining static and dynamic features does not significantly improve the classification over using dFNC features alone. Overall, the proposed approach provides a novel means to assess dFNC, which can be used as brain fingerprints to facilitate prediction and classification.
