This study was approved by the Ethics Committee of Beijing Tiantan Hospital that was affiliated to the Capital Medical University of the People’s Republic of China. The study was conducted in accordance with the Declaration of Helsinki, and all patients provided informed consent for the use of their medical records.

Patients were consecutively recruited and retrospectively analyzed from October 2014 to June 2021 at the Department of Neurology, Beijing Tiantan Hospital, Capital Medical University. The inclusion criteria in this study were as follows. First, patients had definite autoimmune encephalitis; the clinical diagnosis of definite AE should meet the following criteria: (a) subacute onset (rapid progression of <3 months) of working memory deficits, seizures, or psychiatric symptoms suggesting involvement of the limbic system; (b) bilateral brain abnormalities on MRI highly restricted to the medial temporal lobes; (c) cerebrospinal fluid (CSF) pleocytosis (white blood cell count of more than five cells per mm³); and (d) electroencephalogram (EEG) with epileptic or slow-wave activity involving the temporal lobes. The diagnosis can be made when all four criteria have been met, and final diagnosis was confirmed by the detection of serum or CSF positive for specific neuronal autoantibodies, including classical paraneoplastic antibodies (Hu, Yo, Ri, Ma2, CV2, and Amphiphysin), N-methyl-D-aspartate receptor (NMDAR), leucin-rich glioma-inactivated-1 (LGI1), contactin-associated protein-2 (CASPR2), γ-aminobutyric acid type B receptor (GABA_BR), α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptor (AMPAR), and glutamic acid decarboxylase 65 (GAD65). Serum and CSF samples were tested using both cell-based assays (Euroimmun, Lübeck, Germany) and immunohistochemical analyses in the Neuroimmunology Laboratory of the Peking Union Medical College Hospital (Beijing, China). Second, patients should be first admitted in our department from symptom onset. The exclusion criteria included (1) patients did not fulfill sleep assessment or received medication treatments before sleep assessment; (2) patients missed ¹⁸F-FDG-PET data; (3) patients had concurrent autoantibodies (two or more); and (4) patients had history of sleep disorder before disease onset.

Clinical information on demographics, medical history, smoking, drinking, days from onset to diagnosis, comorbid symptoms, neuropsychological and psychiatric test, accompanied tumor, in-hospital laboratory test, electroencephalogram (EEG) and imaging examination, and treatments, were obtained from electronic medical records at first admission. Body mass index (BMI) was calculated as weight (kg)/height (m²). Neuropsychological test contained Montreal Cognitive Assessment (MoCA) (17). Depression and anxiety were evaluated using the Hamilton Depression scale (HAMD) and the Hamilton Anxiety scale (HAMA). Laboratory test included specific antibodies; white blood cell counts, proteins, and oligoclonal bands in CSF; homocysteine (hyperhomocysteinemia was defined as homocysteine concentration > 15 μmol/L); neutrophil-to-lymphocyte ratio (NLR) was calculated by dividing the absolute neutrophil count by the absolute lymphocyte count; neuron-specific enolase (NSE); and serum sodium. Imaging examination included magnetic resonance imaging (MRI) and ¹⁸F-FDG-PET.

Sleep assessment was performed through face-to-face or telephone interview from all patients or their caregivers. First, sleep complaints should be obtained, including insomnia, parasomnia, sleep breathing disorders, sleep-related behaviors or movements, and hypersomnolence. In addition, sleep quality was assessed with the standardized Pittsburgh Sleep Quality Index (PSQI) questionnaire, which evaluates multiple dimensions of sleep over a 1-month period (18). It is a self-reported questionnaire that has seven components, including subjective sleep quality, sleep latency, sleep duration, habitual sleep efficiency, sleep disturbances, use of sleeping medications, and daytime dysfunction. Each of the seven components is scored according to levels 0–3, and each score is eventually accumulated to yield a total PSQI score ranging from 0 to 21, where the higher the score, the worse the sleep quality. In this study, PSQI scores >5 indicated that patients had SD. Then, included AE patients were further divided into two subgroups (AE with SD and AE without SD) according to their PSQI results.

Only 28 patients underwent nocturnal video-polysomnography (PSG) examinations. PSG was conducted using an integrated digital recording system (Greal, Compumedics, Melbourne, Australia) that contained data acquisition, storage, and sleep analysis at the Sleep Center of Beijing Tiantan Hospital, Capital Medical University, China. Every PSG recording contained multiple channels, mainly including electrooculogram (EOG), electromyogram (EMG), electrocardiogram (ECG), respiratory signals, and EEG. EEG channels covered three brain regions (frontal, central, and occipital). Equipment placement, sleep staging, and event scoring were completed by registered polysomnographic technician. Subsequently, the following PSG parameters were collected: total time in bed; total sleep time (TST); sleep latency; sleep efficiency; wakefulness after sleep onset; percentage of stage N1, N2, N3, and rapid eye movement (REM); arousal index (AI) in whole sleep, non-rapid eye movement (NREM) sleep, and REM sleep; apnea hypopnea index (AHI); limb movements index (LMI); and periodic limb movements index (PLMI). In addition, 18 age- and gender-matched controls were recorded using the same procedures. The above variables were measured by two experienced sleep specialists blinded to clinical information of the conditions of either patients or controls, and in case of obvious discordance in their initial evaluations, an informed consensus statement was reached. The Kappa coefficient of two specialists was 0.82. REM sleep behavior disorder (RBD) was defined as a parasomnia with dream-enactment behaviors occurring during REM sleep and associated with the lack of the physiological REM sleep muscle atonia. Obstructive sleep apnea (OSA) was defined as an AHI of five or more events per hour.

All raw data of PET images were collected from PET/CT workstation; a detailed ¹⁸F-FDG-PET procedure has been published elsewhere using a PET/CT scanner (Elite Discovery, GE HealthCare, USA) (19). Before examination, patients did not receive neuroleptic drugs, fasted for at least 6 h, and controlled fasting blood glucose levels <8 mmol/L. Then, patients were injected with ¹⁸F-FDG at a dose of 3.7–5.0 MBq/kg in a quietly dedicated room. After a 45–60-min uptake period, brain PET scan was performed in the 3D-time-of-flight mode for 10 min, and whole-body PET scan was performed for approximately 20–30 min. The brain imaging data were reconstructed using ordered subset expectation maximization methods, with four iterations and eight subsets, and smoothing with a 5-mm full-width at half-maximum filter. Statistical parametric mapping (SPM) 12 (Wellcome Centre for Human Neuroimaging, London, UK) implanted in a MATLAB 2018a environment (MathWorks, Inc., USA) was used for imaging reprocessing analysis. The preprocessing steps were as follows: first, the PET images were spatially normalized into a common Montreal Neurological Institute (MNI) atlas anatomical space following a 12-parameter affine transformation and nonlinear transformations, yielding images composed of 2 mm × 2 mm × 2 mm voxels. Then, default SPM smoothing was applied using a 14-mm Gaussian kernel to increase the signal-to-noise ratio.

For the analysis on brain metabolic mechanism of sleep disorders in AE, we carried out group-level comparison between AE with SD and without SD by two-sample t-test model of SPM12 with age and gender as the nuisance variables. Significant results were viewed at the height threshold (p < 0.001) and corrected for multiple comparisons (FDR, corrected, p < 0.05). If no cluster of significant difference was found, the more liberal threshold at p < 0.001 uncorrected was considered to perform further exploratory analyses (20). Finally, we used the xjView SPM extension (Cui & Li, Human Neuroimaging Lab, Baylor College of Medicine) to visualize the corresponding anatomic locations of each peak MNI of these significant different clusters.

To further explore homogeneity of regional brain glucose metabolic changes between groups in particular spatial pattern, we performed a volumetric region of interest (ROI) analysis in the significant clusters based on aforementioned SPM results. First, the following 10 ROIs were identified: medial temporal lobe (MTL, mainly including amygdala and hippocampus), posterior cingulate cortex (PCC), basal ganglia (BG, including globus pallidus, putamen, and caudate), brainstem (including midbrain, pons, and medulla), superior frontal gyrus (SFG) and medial frontal gyrus (MFG), thalamus, hypothalamus, cerebellum anterior lobe (CAL), and cerebellum posterior lobe (CPL). These ROIs were defined based on the group-level results obtained from SPM analysis. The brainstem, hypothalamus, and CAL and CPL regions of interest were generated using the WFU-Pickatlas toolbox for SPM12 based on Talairach Daemon lobars, and the remaining ROIs were derived and summarized from the Anatomical Automatic Labeling (AAL) atlas. Subsequently, the ratio of the standardized uptake value (SUVR) representing ¹⁸F-FDG uptake calculates in those selected ROIs was obtained from each individual. Briefly, an SUVR was derived for ¹⁸F-FDG-PET from the voxel size weighted median uptake in the regions of interest normalized to the whole brain.

Continuous variables with a normal distribution were presented as the mean ± standard deviation, and non-normal variables were reported as the median [interquartile range (IQR)]. The normality of the data was performed by the Shapiro–Wilk test. Categorical variables were showed as frequency with corresponding percentage. We compared groups using t-tests for continuous variables that were normally distributed, Mann–Whitney U tests for non-parametric data, and χ² tests or Fisher’s exact tests for categorical variables.

Clinical variables were comprehensively collected for possible inclusion into the risk model. We used binary logistic regression to assess independent risk factors associated with AE with SD; all variables with p < 0.2 in the univariate analysis were included in multivariable logistic regression model. Subsequently, the likelihood ratio test was used in a backwards elimination process (p < 0.05 to retain, p > 0.1 to remove) to select the final set of independent risk factors for retention into the model. Odds ratio (OR) with 95% CI was presented for logistic regression model. The correlation between the SUVR values of selected ROIs and PSG parameters was conducted using Spearman test. A two-sided p < 0.05 was considered to indicate statistical significance. SPSS 22.0 software package (IBM Corp., Armonk, New York, USA) and Prism 8 (GraphPad Software, CA, USA) were used for statistical analyses.

The flowchart of patient inclusion is shown in Figure 1 . A total of 187 patients with definite and antibody-confirmed AE were eligible to participate in this study, of whom 121 patients were ultimately identified in the current study after exclusion, including 19 patients with NMDAR antibodies, 71 patients with LGI1 antibodies, 4 patients with CASPR2 antibodies, 21 patients with GABA_BR antibodies, and 6 patients with GAD65 antibodies. Among these cases, 52.9% of AE patients (n = 64) had SD (36.8% NMDAR, 59.2% LGI1, 75% CASPR2, 47.6% GABA_BR, and 33.3% GAD65; Figure 2A ); the median PSQI score in the group of SD was 9 (IQR, 7–14), which was obviously higher than those without SD [median, 9 (IQR, 7–14) vs. 3 (IQR, 2–3), p < 0.001). In addition, patients with CASPR2 and LGI1 had higher PSQI score than those with NMDAR antibodies (p < 0.05, Figure 2B ).

There is no significant difference regarding the median time from diagnosis to the sleep assessment [9 (IQR, 6–20) days vs. 9 (IQR, 6–17) days, p = 0.892) between patients with SD and without SD. The baseline characteristics of patients based on sleep quality are summarized in Table 1 . Patients with SD were older than those without SD (median age, 55.4 vs. 49.0 years, p =0.001). Hypertension, smoking, cognitive deficits, psychosis, and abnormal behaviors showed statistically differences between patients with SD and without SD (p < 0.05 for all). Compared with patients without SD, patients with SD had higher levels of MoCA score (p =0.001) and NSE (p < 0.001) and had lower levels of serum sodium (p =0.001). Eleven of these patients (9%) were associated with tumors: 8 lung cancers (2 with LGI1 antibodies, 7 with GABAB antibodies), 1 colorectal adenoma (1 with LGI1 antibodies), 1 ovarian teratoma (1 with NMDAR antibodies), and 1 thymoma (1 with LGI1 antibodies). Tumor removal was performed in three patients, and chemotherapy was performed in eight patients. In a univariate analysis, seven patients (10.9%) presented tumors in the group of sleep disorders, and four patients (7%) with tumor were shown in the group of patients without sleep disorders. There was no statistical significance between the two groups (p = 0.454). All patients received first-line immunotherapy [20 IV methylprednisolone (IVMP) alone, 21 IV immunoglobulins (IVIg) alone, and 80 IVMP + IVIg]. Twelve patients additionally received second-line immunotherapy (four rituximab and eight mycophenolate mofetil). There was no significant difference between patients with SD and without SD regarding relevant treatments (p = 0.363). More details of corresponding descriptions are reported in Table 1 .

A subgroup of 28 patients completed PSG recordings. These 28 patients were a representative sample of the whole cohort of 121 patients because there were no significant differences across clinical variables between patients with PSG and the whole cohort ( Supplementary Table S1 ).

The sleep parameters of PSG are presented in Table 2 . The median TST was 292.8 (160–489.4) min. Sleep efficiency was approximately 56.1%. Sleep onset latency was 37 (1.5–189) min. Wakefulness after sleep onset was 145.8 (26.5–264.5) min. For percent of sleep stages, the stage N1, N2, N3, and REM comprised 25.2% (4.3%–56.8%), 51.4% (15.6%–85.6%), 9.4% (0%–29.8%), and 12.5% (0.7%–32.5%) of the TST, respectively. REM sleep latency was 139 (24–287.5) min. Nine patients had dream enhancement behaviors, and the final diagnosis of RBD was confirmed on PSG in five patients (5/28, 17.9%), of whom three had LGI1 antibodies. Sixteen of 28 patients (57.1%) developed OSA with median AHI 10/hour. AI in whole sleep, NREM, and REM were 8.6 (0–33.5), 17 (0.5–33), and 14 (0.9–30), respectively. The median LMI and PLMI were 16/h (1–360/h) and 12/h (0–25/h), respectively.

There were no significant differences between patients with SD and without SD regarding age, gender, and body mass index ( Table 2 ). Compared to controls, AE patients had less TST, less sleep efficiency, longer sleep latency, more wake, higher percent of stage N1, lower percent of stage N3 and REM, and more AI in NREM sleep (p < 0.05 for all). Other sleep variables demonstrated no significant differences between the two groups.

Results of multivariate logistic analysis of the final model are shown in Table 3 . We selected variables with p < 0.2 in the univariate analysis, including age, male, hypertension, smoking, drinking, cognitive deficits, psychosis and abnormal behaviors, MoCA score, HAMD score, hyperhomocysteinemia, NLR, NSE, serum sodium, and initial MRI. After adjusting for all included clinical variables, the logistic regression analysis showed that smoking [OR, 6.774 (95% CI, 1.238–37.082); p = 0.027], increased HAMD score [OR, 1.074 (95% CI, 1.002–1.152); p = 0.045], hyperhomocysteinemia [OR, 2.815 (95% CI, 1.057–7.496); p = 0.038), elevated NSE level [OR, 1.069 (95% CI, 1.007–1.135); p = 0.03] were independently correlated with higher risk of SD in AE patients. Contrastingly, high MoCA score [OR, 0.821 (95% CI, 0.752–0.896); p < 0.001] was associated with lower risk of SD in AE subjects.

Figure 3 shows the significant metabolic differences of group comparisons based on voxel-based analysis between patients with and without SD, no significant difference was noted between the two groups with regard to the median time from diagnosis to the initial scan [6 (IQR, 2–18) days vs. 5 (IQR, 2–16) days, p = 0.698]. Patients with SD demonstrated relatively increased metabolism in the MTL, BG (medial globus pallidus, putamen, and caudate), extending to the cerebellum (cerebellum anterior lobe, cerebellum posterior lobe, dentate, cerebellar tonsil, and culmen), brainstem (midbrain and pons), thalamus, and hypothalamus compared to patients without SD (p < 0.05, FDR corrected; Figure 3 and Table 4 ). Furthermore, patients with SD also showed hypometabolism in the frontal lobe (MFG and SFG) and PCC (p < 0.001, uncorrected; Figure 3 and Table 4 ). Cluster extent, peak MNI coordinates, corrected p values, maximum Z value, corresponding brain regions, and voxel size for each region are detailly reported in Table 4 .

To investigate and validate whether the above significant clusters containing different brain regions were truly related to sleep quality, we further conducted an ROI-based correlation analysis between patients with SD and without SD. The comparison of SUVR value between the two groups is shown in Figure 4 . Compared with patients without SD, patients with SD showed relatively higher metabolism in the MTL (0.99 ± 0.17 vs. 0.93 ± 0.09, p = 0.008), BG (1.2 ± 0.12 vs. 1.14 ± 0.1, p = 0.002), brainstem (0.87 ± 0.07 vs. 0.83 ± 0.07, p = 0.003), hypothalamus (0.89 ± 0.11 vs. 0.83 ± 0.1, p = 0.003), CAL (1.1 ± 0.09 vs. 1.06 ± 0.09, p = 0.033). Patients with SD also demonstrated more increased metabolism than those without SD in thalamus (1.1 ± 0.07 vs. 1.09 ± 0.07, p = 0.251) and CPL (0.93 ± 0.08 vs. 0.91 ± 0.07, p = 0.269), but there was no significant difference between the two groups. In contrast, decreased metabolism in the SFG (0.91 ± 0.07 vs. 0.95 ± 0.06, p = 0.001), MFG (0.93 ± 0.08 vs. 0.98 ± 0.08, p = 0.003), and PCC (1.04 ± 0.12 vs. 1.1 ± 0.11, p = 0.002) were observed in patients with SD in comparison with those without SD.

In order to explore the relationship between PET and sleep quality in more detail, we further performed a more fine-grained PSG-based correlation analysis ( Figure 5 ). Wakefulness after sleep onset was associated with abnormal metabolism in the brainstem (r = 0.52, p =0.005), SFG (r = −0.54, p =0.003), MFG (r = −0.51, p = 0.005), and CAL (r = 0.51, p = 0.006). Stage N1 correlated with cerebral metabolism predominantly in CAL (r = 0.62, p <0.001) and CPL (r = 0.62, p <0.001). Stage N2 was associated with metabolism in the hypothalamus (r = 0.47, p = 0.012). Stage N3 was related to glucose metabolism in PCC (r = 0.42, p =0.027), brainstem (r = −0.42, p = 0.026), MFG (r = 0.4, p = 0.033), CAL (r = −0.42, p = 0.028), CPL (r = −0.41, p = 0.029). Stage REM correlated with cerebral metabolism in MTL (r = −0.39, p = 0.041), brainstem (r = −0.28, p =0.04), thalamus (r = −0.38, p =0.049), and hypothalamus (r = -0.42, p =0.027). REM latency was associated brain metabolism in thalamus (r = -0.4, p =0.035). AI in NREM sleep was related to glucose metabolism in brainstem (r = 0.49, p =0.009), SFG (r = −0.39, p =0.042), MFG (r = −0.43, p = 0.021), CAL (r = 0.43, p =0.024). AI in REM sleep correlated with cerebral metabolism mainly in BG (r = 0.44, p = 0.021). AHI was associated with abnormal metabolism in MFG (r = −0.39, p = 0.041). There was no statistically significant correlation between other PSG variables and regional brain metabolism on FDG-PET.