Twenty-nine acute anti-NMDAR encephalitis patients were enrolled in the department of neurology at the affiliated hospital of Southwest Medical University from June 2018 to May 2023. The inclusion criteria were defined in accordance with established diagnostic standards for anti-NMDAR encephalitis as previously published [27] and reported in a previous study by the authors [15]. The immunoglobulin G anti-neuronal antibodies (Abs) were tested in serum and/or cerebrospinal fluid. Auto-Abs including NMDAR, alpha-amino-3-hydroxyl-5-methyl-4-isoxazole-propionate receptor, contactin-associated protein-like 2, Leucine-rich glioma-inactivated 1, alpha-amino-3-hydroxyl-5-methyl-4-isoxazole-propionate receptor and aminobutyric acid receptor Abs as well as anti-onconeural Abs (Abs anti-Hu) in serum and/or cerebrospinal fluid were screened. Only acute AIE patients (onset time less than one month) with single positive Abs NMDAR outcomes were included. All included participants were right-handed. Exclusion criteria encompassed individuals with a prior history of severe neuropsychiatric diseases, presence of EEG artifacts, negative antibody outcomes, or those with multiple Abs positive outcome. The Modified Rankin Scale (mRS), Mini-mental State Examination (MMSE) and clinical traits were assessed for all patients. The control group was matched from our EEG database including healthy participants. For the matching criteria, we used a MATLAB function Age_gender_match (Luki, 2023, https://www.mathworks.com/matlabcentral/fileexchange/66984-age_gender_match-year_gap-group1-group2). In this function, the age and sex from healthy control (HC) group were matched randomly for 10,000 times. Thus, the matching-priority was calculated. Then, the best matched subjects were selected as control group. These sex and age matched right-handed healthy participants without neurological or mental disorders and drug addiction were included as the HC group.

This step was performed in a semi-isolated room. Before EEG acquisition, participants were required to keep their hair clean and dry. A 19-channel analog recorder (Galileo EB Neuro, Florence, Italy) was used and the leads were assigned in the scalp using a quantified ruler following the International 10–20 system released by the American Clinical Neurophysiology Society. The impedance of each electrode was reduced to lower than 10 K$\mathrm{\Omega }$. The sampling rate was 500 Hz. All participants were instructed to keep thinking of nothing in particular with eyes closed at the start of data recording (more than five minutes). The EEG recording for each participant lasted for one hour.

The EEG preprocessing steps were performed using a self-writing MATLAB (R2014a MathWorks, Natick, MA, USA) pipeline based on EEGlab (v14.1.1, http://sccn.ucsd.edu/) and are similar to a previous study [28]. The original data obtained from each participant was exported into the European Data Format. EEG data was filtered with a bandpass of 1–70 Hz. Eye movement and electromyography artifacts were removed automatically using an EEGlab plugin-in artifact removal tool (http://germangh.com). The tool employed a blind source separation method with a fast independent component analysis algorithm that automatically eliminated artifacts. Subsequently, data were re-referenced to the average lead. Bad channels were interpolated and the number of bad electrodes was limited to a single channel. Channels were then re-referenced to the average lead. Previous studies have found that EEG analysis parameters are stable and can be robustly estimated when the length of resting-state EEG for each participant is longer than 30 s [29]. EEG data with a 50 s duration for each participant gives relatively stable results. Thus, five 10-second epochs of data for each participant with eyes closed, awake were extracted and used in the follow-up analysis.

Regions of Interest (ROIs) were defined according to 84 Brodmann areas (BAs) that were divided into areas according to their structural location in Montreal Neurological Institute (MNI) space [30] (Supplementary Table 1) Source signals of ROIs were extracted using the exact LORETA (eLORETA) method with LORETA-KEY (v20190226, https://www.uzh.ch/). The eLORETA method has been described and applied in many previous reports [28, 31, 32] and at the website of Roberto D. Pascual-Marqui (https://www.uzh.ch/keyinst/loreta). In brief, eLORETA is a non-linear method and a true inverse solution for source localization with exact and zero localization errors [33]. The scalp EEG data were transformed to eLORETA files with a transformation matrix created by using the electrode location file. The signal for each ROI was then calculated using the eLORETA files and the ROI coordinates.

PLI was used to construct the FC matrix of the ROIs with the obtained signals in different bands. PLI was first described by Stam and colleagues [23]. The central idea is to discard phase locking, centered around 0 and every $\pi$ phase difference, to remove volume conduction effects. The equation is [34]:

(1)

The sign here indicates the signum function. The PLI varies between 0 and 1, where 0 indicates the absence of coupling due to volume conduction and 1 indicates true, delayed interaction. One PLI matrix was constructed for each participant in each frequency band of the six frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta 1 (13–30 Hz), beta 2 (30–45 Hz) and gamma (45–70 Hz).

Using the PLI matrices, graph parameters including small world index (sigma), global efficiency, local efficiency, clustering coefficient and shortest path length were calculated. The detailed algorithms for these parameters have been described in previous studies [35, 36]. Functional brain differentiation can be measured by the clustering coefficient and local efficiency, whereas the overall processing and transfer capacity of information, as well as the level of network integration can be assessed by the shortest path length and global efficiency. These graph parameters were obtained using a function embedded in the MATLAB based tool: Brain Connectivity Toolbox [37] (https://sites.google.com/site/bctnet/).

The power spectra for each frequency within 1–70 Hz were estimated using the toolbox EEGlab (v14.1.1). The six frequency bands were then defined as in the above FC analysis. The power in each frequency band and total power were calculated. Finally, the RBP in each band were calculated with:

(2)$$\mathrm{RBP}=100\%\times \mathit{\text{Band\_ Power/Total\_ Power}}$$

Where Band_Power indicates the frequency band power. Total_Power means the total power of the six frequency bands.

A generalized linear model was used to eliminate possible sex and age artifacts Two sample tests with false discovery rate (FDR) correction were used to compare the differences in FC and graph parameters between patients and HCs and subgroups within patients. Pearson’s correlation coefficient was calculated to detect any relations between EEG parameters and clinical scales. The effect size was estimated using Cohen’s d value. A p 0.05 after FDR correction was considered to show a significant difference. All statistical analysis was performed with the MATLAB statistics toolbox (R2014a MathWorks, Natick, MA, USA).

Twenty-nine anti-NMDAR encephalitis patients aged 29.79 $\pm$ 14.75 years and 29 HCs aged 24.58 $\pm$ 9.05 years were included in this study. No significant sex and age differences were found between the two groups. The mean MMSE and mRS in anti-NMDAR encephalitis patients scored 20.57 $\pm$ 6.43 and 2.93 $\pm$ 0.94, respectively. The mRS score at admission and discharge is 2.93 $\pm$ 0.94 and 2.00 $\pm$ 1.02, respectively. Six patients showed involuntary movements and 2 patients showed psychiatric symptoms (Table 1). Twenty-five patients had a normal intracranial pressure. Seven and 19 patients showed normal EEG and MRI patterns, respectively. Five patients had typical delta brush, eight patients showed slowing rhythm overlapping excessive beta activity and four patients showed epileptiform discharge (Table 2).

The large-scale FC was compared between anti-NMDAR encephalitis patients and HCs based on the 84 brodmann area (BA) ROIs. Few edges in delta and theta bands showed increased FC and a large number of edges in beta 1 and beta 2 bands showed increased FC in patients having anti-NMDAR encephalitis (Fig. 1). For other frequency bands, no significant differences in FC were found between the two groups. No decrease in FC was observed in participants, compared with HC.

Fig. 1.

Comparison of large-scale networks between anti-NMDAR encephalitis and HC groups${}^{\mathrm{\#}}$. ${}^{\mathrm{\#}}$PLI functional networks in delta (1–4 Hz) (A), theta (4–8 Hz) (B), beta 1 (13–30 Hz) (C) and beta 2 band (30–45 Hz) (D). Nodes were defined by Brodmann areas. The edge between paired areas indicate significant increased FC in the patient group, compared to HCs at each frequency band after FDR corrections (p 0.05). Blue text indicates the macro subnetworks of the brain. For other frequency bands, no significant differences in FC were found between the two groups. No decrease in FC was obtained in anti-NMDAR encephalitis patients, compared with the HC group. FC, Functional connectivity; HC, Healthy control; FDR, false discovery rate; PLI, Phase Lag Index; L, Left hemisphere; R, Right hemisphere.

However, results showed some differences when the FC based on scalp EEG channels were compared. The FC mainly involving frontal-parietal in alpha and beta 1 bands decreased in patients, compared with those of HCs. Few edges in theta and beta 2 bands had increased FC in patients (Supplementary Fig. 1).

Compared with the HC group, local efficiency and the clustering coefficient in delta, theta, beta 1 and beta 2 bands increased in anti-NMDAR encephalitis patients (p 0.05). Additionally, global efficiency in beta 1 and beta 2 bands increased while the shortest path length in these two bands decreased in patients (p 0.05, Fig. 2).

Fig. 2.

Comparison of graph theory parameters between AIE and HC groups${}^{\mathrm{\#}}$. ${}^{\mathrm{\#}}$ Network Property comparisons of (A) delta (1–4 Hz), (B) theta (4–8 Hz), (C) beta 1 (13–30 Hz), (D) beta 2 band (30–45 Hz) between the two groups. *p 0.05 after FDR correction. Legends are shown below the figure. d, Effect size (Cohen’s d), t value of two sample t test.

When these parameters were analyzed based on the scalp 19 $\times$ 19 PLI matrices, it was generally found that global efficiency decreased and shortest path length increased in the alpha band of patients (Supplementary Fig. 2).

Compared with the HC group, the AIE group showed increased delta and theta RBP in frontal, parietal, occipital, temporal and sub-lobar lobes. However, the RBP of alpha and beta 1 bands in parietal, occipital, temporal and sub-lobar lobes decreased in anti NMDAR encephalitis patients. Additionally, compared with the HC group, the theta band RBP increased while the alpha band RBP decreased in the limbic system of patients (Fig. 3).

Fig. 3.

Comparison of relative band power of macro subnetworks between AIE and HC groups${}^{\mathrm{\#}}$. ${}^{\mathrm{\#}}$(A–F), the relative band powers for the six frequency bands in (A) Frontal, (B) Parietal, (C) Occipital, (D) Temporal, (E) Limbic and (F) sub-lobar networks. *p 0.05, **p 0.01, ***p 0.001 after FDR correction. Legends are given below the figure. Effect size (Cohen’s d), t value of two sample t test.

When sub-groups were partitioned by EEG pattern, MRI pattern and MMSE scores, significant differences were observed in the RBP of some brain areas. Comparing the AIE group with the normal EEG pattern, AIE participants with abnormal EEG pattern showed decreased alpha RBP in the whole brain. Additionally, those AIE participants with abnormal MRI showed significantly increased theta RBP in temporal, limbic and sub-lobar systems. The alpha RBP in limbic systems also increased significantly in these participants, compared with those with a normal MRI. Moreover, compared with those participants showing higher MMSE, those with a lower MMSE showed increased beta 1 RBP in the frontal lobe and limbic system (Fig. 4). Other parameters we did not mention including both source localization-based and scalp EEG-based graph parameters showed no significant differences in these subgroups.

Fig. 4.

Subgroup comparisons and correlation analyses. RBP comparisons of subgroups divided by (A) EEG pattern, (B) MRI pattern, (C) MMSE, and (D) Correlation analysis between cortical RBP, mRS and MMSE. EEG, electroencephalogram; AbnorEEG, Abnormal EEG pattern; NorEEG, Normal EEG pattern; MRI, Magnetic Resonance Imaging; AbnorMRI, Abnormal MRI pattern; NorMRI, Normal MRI pattern; MMSE, Mini-mental State Examination; mRS, Modified Rankin Scale; RBP, Relative band power; r, Pearson’s correlation coefficient; CI, Confidential interval; d, Effect size (Cohen’s d); t value of two sample t test. *p 0.05 after FDR correction.

Significant negative correlations were found between the beta 1 RBP in the frontal lobes and mRS scores (r = –0.393, p = 0.043). The beta 1 RBP in the occipital lobe was significantly positively correlated with MMSE (r = 0.402, p = 0.031, Fig. 4). Other parameters not mentioned showed no significant correlations with MMSE or mRS.