In this study, 51 subjects (25 healthy participants and 26 depressed patients) were enrolled. The patients were recruited from Guangdong 999 Brain Hospital and were diagnosed by a psychiatrist, according to the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV). The participants with depression were evaluated using the Hamilton Depression Scale (HAMD) and the Self-Rating Depression Scale (SDS). Depressed patients with drug abuse, suicide risk, pregnancy, current or history of head injuries, seizures, or epilepsy were excluded. All the healthy participants (approximately 20 years old) were undergraduates at Sun Yat-Sen University and were free from sleep disorders such as insomnia, depression, and sleep apnea. The male/female in the control and depression groups was 11/14 and 15/11, respectively. All the participants were accustomed to using the right hand. All the procedures were performed in compliance with the 1964 Helsinki declaration and its subsequent amendments. This study was approved by the Ethics Committee of Guangdong 999 Brain Hospital (approval number: 2020-010-059), and the research was carried out here. The subjects were informed of the experimental procedures in advance and told not to drink caffeinated or alcoholic beverages and not to overeat for 24 h before arriving at the sleep center of the hospital. Informed consent was obtained from all the individual participants.

All the participants underwent overnight polysomnography (PSG) examination, which lasted for 9–10 h. Recordings were performed by using a Compumedics Profusion EEG Recording System with Neuvo Amplifier. PSG recordings included 16 EEG channels [regarding six brain areas: frontal poles (Fp1/M2 and Fp2/M1), frontal lobe (F3/M2, F4/M1, F7/M2, and F8/M1), central (C3/M2 and C4/M1), occipital lobes (O1/M2 and O2/M1), parietal lobes (P3/M2 and P4/M1), and temporal lobes (T3/M2, T4/M1, T5/M2, and T6/M1)] based on the standard 10–20 system, electrooculography (EOG), electrocardiography (ECG), electromyography (EMG), oral and nasal respiratory airflow, chest and abdomen breathing movement, blood oxygen saturation, snoring, leg movement, and body position. The sampling frequency of the EEG, EOG, and EMG signals was 500 Hz, while that of the respiration signal and oxygen saturation were set to 100 and 10 Hz, respectively. Sleep stages were scored by experienced sleep technicians, according to the AASM criteria as Wake, N1, N2, N3, and REM (Saper et al., 2010). The information of the subjects is shown in Table 1.

Segments containing obvious artifacts were removed by visual inspection. A total of 25194/19150 segments of depressed patients/healthy participants were obtained (1467/567 W epochs, 4692/2945 R epochs, 1030/545 N1 epochs, 11608/8138 N2 epochs, and 6397/6955 N3 epochs).

The dFC analysis focused on the sleep stages’ switching periods. The first 30-s epochs at the beginning of each sleep stage were calculated. These 30-s segments were divided into three 10-s segments for calculating the dFC. In total, 2739/1936 segments of depressed patients/healthy participants were obtained for the dFC analysis (313/202 W epochs, 164/258 R epochs, 427/442 N1 epochs, 1169/774 N2 epochs, and 666/260 N3 epochs).

A fourth-order zero phase shift Butterworth bandpass filter with a passband of 0.5–50 Hz was applied for signal preprocessing. The EEG data were estimated and processed in a moving window of 3 s (with an overlap of 2 s).

The analysis was performed in the five frequency bands of δ (0.5–4 Hz), θ (4–8 Hz), α (8–12 Hz), σ (12–16 Hz), and β (16–32 Hz).

The WPLI, explained by Vinck et al. and improved by Liao et al., was calculated to measure the phase synchronization index for characterizing the FC (Liao et al., 2019). The WPLI introduces the imaginary part of the cross power spectrum to weigh the signal for improving the immunity to noise sources; thus, it can reflect the cortical oscillation relationship between different parts of the same brain area or between different brain areas. The WPLI is defined as follows: WPL I i , j , τ = | E { sin ( Δ φ i,j,τ ) } | E { | sin ( Δ φ i , j,τ ) | } , where 0 ≤ W P L I i , j , τ ≤ 1, and E () is the expected value operator. Δ φ i , j , τ , representing the phase difference between a pair of nodes i and j, was calculated using the formula Δ φ i , j , τ = φ i ( τ ) − φ j ( τ ) .

The instantaneous phase φ ( τ ) is extracted by using a combined Morlet wavelet, which has a stable passband and a narrow transition band. A single Morlet complex wavelet is defined as follows: φ ( t ) = 1 π f b e 2 iπt f c e − t 2 f b , where f c is the central frequency of the wavelet, and f b is the bandwidth parameter. The combined Morlet wavelet is obtained by superimposing multiple Morlet wavelets with different center frequencies f M . The central frequency of the Morlet wavelet is thus defined as follows: f M = f L + M × Δ f , M = 0 … N − 1 , where Δ f , which is the central frequency spacing of the wavelet, is 0.05 in this study. The corresponding f L and N are defined according to different frequency bands, such as f L = 0.5 and N = 70 for δ and f L = 4 N = 80 for θ. The Morlet combination is defined as the following formula: ψ c ( t ) = 1 C ∑ M = 0 N − 1 σ f M ( t ) = 1 C π f b e − x 2 f b ∑ M = 0 N − 1 e 2 iπ f M t , where C is the correction coefficient that makes the amplitude–frequency characteristic passband of the combined wavelet to be 1. The phase information within a fixed frequency band will be more accurate because of the high concentration of the Morlet wavelet in the time and frequency domains. The wavelet coefficient of the EEG signal S ( t ) , which is obtained from the electrode at time τ , is defined as follows: W S ( τ ) = ∫ − ∞ + ∞ S ( t ) ψ C ∗ ( t − τ ) dt = A ( τ ) e i φ ( τ ) .

“∗” represents conjugation. The process and calculation formula of Morlet wavelet phase extraction can be obtained in the manuscript of Wang et al. (2011).

The weighted undirected network was based on the WPLI with rows and columns corresponding to two electrodes. To evaluate the characteristics of the weighted undirected network, some graph theory metrics were used by the Brain Connectivity Toolbox (BCT) (Rubinov and Sporns, 2010; Rossini, 2021). GE (global efficiency) and node strength were calculated to compare the functional networks in healthy participants and depressed patients. GE can be used to evaluate the separation and randomness of the cortical network. With respect to node strength, the sum of link weights around the node can exhibit the hub of the network. We built a binary network using a weighting matrix. According to previous research (Achard and Bullmore, 2007; Jalili, 2017; Kabbara et al., 2018), we fixed the density value at 0.30 for sparsity thresholding to construct a sparse matrix.

Rubinov and Sporns (2010) introduced the calculation formula and representative meaning of the parameters in graph theory. GE reflects exchanged efficient information in the network, and the measures quantified the network efficiency. GE is defined as follows: E = 1 n ∑ i ∈ N E i = 1 n ∑ i ∈ N ∑ j ∈ N , j ≠ i d ij W − 1 n − 1 . E i is the efficiency of node i. d i j is the distance between nodes i and j, and d i j is defined as d ij W =   ∑ a ij ∈ gi ↔ j W f ( W uv ) , where f is a map from the weight to length, and g i ↔ j is the shortest weighted path between i and j. W u v is the link value in the 16∗16 matrix of the WPLI. In this study, we fixed the density value at 0.30 for sparsity thresholding to construct a sparse matrix. Consequently, when the WPLI > the threshold, the map f indicates that the data are not changed. When the WPLI < the threshold, the map f represents 0. The node strength, which can exhibit the hub of the network, is the sum of the link weights connected to the node in the weighted network. The node strength is defined as follows: k i W = ∑ j ∈ N , j ≠ i W ij , where W i j is the link weight between nodes i and j.

The statistical tests were carried out on five frequency bands in depressed patients and controls to access the cortical networks in depressed patients. The Kolmogorov–Smirnov test was used to verify the normality of the obtained data. Since GE, node strength, and WPLI did not satisfy the normal distribution or the variance homogeneity test, a nonparametric test (Mann–Whitney U test) was applied. To evaluate the laterality of the network, a paired T-test was used for the left and right hemispheres in the depressed patients and controls. The significance of the differences between the sleep stages for each frequency band was analyzed by the Kruskal–Wallis test followed by the post hoc Dunn–Bonferroni test. All the analyses were performed by IBM statistics software SPSS (version 22.0).

The GE values and significant differences between the groups are shown in Figure 2. The depressed patients had significantly higher GE values than the controls in all the frequency bands and sleep stages studied. In addition, the differences between the patients and controls were greater than those between different sleep stages within each group.

Dynamic analysis of GE during switching of the sleep stages was also performed. The average global of the network and changes in the standard deviation of the two groups over these epochs are depicted in Figure 3. The standard deviation at the onset of the N3 stage became smaller in θ, α, σ, and β bands in the patient group, while that in the control group remained at a relatively low level.

Within the depressed and control groups, the differences across the sleep stages manifested differently in some frequency bands. We observed different results across Wake, N1, and REM between the groups. The depressed group had significant differences between Wake and N1 in all the researched bands, but the controls only exhibited a significant difference in the σ band, and this comparison result is opposite to that of the patients. However, the depressed group did not indicate any significant difference between REM and N1, but such differences were shown in most bands for the control group. With the deepening of N-REM sleep, which is from N1 to N2 and to N3, GE was increasing for both the groups in general, but the changes in θ and β bands were not so significant. Exceptions were as follows, in control: the GE of N1 was larger than N2 in the δ and β bands. Meanwhile, less significant differences were observed in θ for the two groups.

To have a better understanding of the network hubs, the node strength was studied to analyze the differences between the two groups in the network link weight. Figure 4 and Figure 5 illustrate the average node strength by color and volume of every electrode, respectively. Compared with the control group, the depressed patients were significantly higher in the node strength. All the comparisons of the node strength between the patients and controls showed significant differences.

Notably, the point at which the highest value appears varied with the sleep stage and frequency component. In both the depressed and control groups, the highest node strength was recorded in the N3 stage for δ and σ bands, during W for α and β bands and in the N2 stage for θ band.

There was no significant difference in the dominance between the depressed and control groups. In almost all the sleep stages and frequency bands, the node strength of the central node was significantly higher than that of the other nodes.

The undirected connectivity graphs characterized significant changes in the WPLI between the depressed and control groups (Figures 6, 7). The weight of each link was the average of weights of that link in the networks of all the participants in every sleep stage. These figures showed the top 30% connectivity in both the groups of each band and each sleep stage.

The link patterns in the two groups were also different. In the patient group, there were links between the hemispheres in each frequency band and sleep stage. On the contrary, the link of the control group is only observed within each hemisphere, except for each sleep stage in δ and Wake in θ. But there are many long-range links inside the hemisphere in the control group.

Furthermore, in most frequency bands and sleep stages, the anterior cortex was obviously denser than the posterior cortex in the depressed patients. For the control group, this anterior–posterior difference was attenuated, and no clear anterior dominance was noted in the high-frequency bands (σ and β bands).

The asymmetry of the cortical network between the hemispheres was studied for the depressed and control groups, respectively. A paired t-test was performed to analyze the node strength between the left and right hemisphere belonging to the same sleep stage. The average of node strength and significant differences between the hemispheres are indicated in Table 2. No significant differences for each frequency band and sleep stage were observed in the control group. In the patients group also, no significant differences were noted except for the δ band, where the node strength of the patients’ right hemisphere was significantly higher than that of the left hemisphere in the W, N2, and R stage.