All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the Ethics Committee of the National Center of Neurology and Psychiatry (approval number: A2011-071).

As health screening prior to the study, 16 healthy right-handed adult men aged 23.4 ± 2.4 (mean ± SD) years underwent overnight polysomnography (PSG), structural MRI, and hematological testing and completed a questionnaire survey and medical interview with a physician.

Exclusion criteria were as follows: some form of sleep disorder, mental disorder, visual impairment including color blindness, or severe physical complication; current use of medication or substances that might affect the study outcome (e.g., sleep aids, antihistamines, other sleep-inducing drugs, and steroids); a metallic implant such as a pacemaker; travel to a foreign country with a ≥6-h time difference within the last 3 months; shift work; and smoking. One of the 16 subjects dropped out of the study for personal reasons and was therefore excluded from the analysis; thus, all analyses were performed using data from the remaining 15 subjects (aged 23.3 ± 2.1 years). These subjects were originally recruited to study individual PSD (16), and they completed the same interventional protocol. In this article, however, we focused on completely different aspects of the data; that is, we newly adopted a neuroimaging technique to visualize the mood changes and neural responses before and after ES, and the analyses reported here do not overlap those previously published.

The experimental protocol used is shown in Figure 1. After a habitual sleep monitoring period at home (HS period, 2 weeks), the study period consisted of four sessions: a 2-day baseline (BL) period, 9-day ES period, one TSD night, and one recovery sleep (RS) night. On day 1 of the BL period (BL1), on days 4 and 9 of the ES period (ES4 and ES9), and after the TSD, subjective mood was evaluated and ASL MRI was performed.

Subjects recorded bedtime and wake-up time during the HS period. Using the algorithms developed by Sadeh et al. (42), habitual sleep duration (HSD) was calculated from data obtained by a wrist actigraph (Micro-Mini Motionlogger Actigraph, Ambulatory Monitoring, Inc.) worn on the non-dominant hand.

The experiment was conducted in the sleep laboratory at the National Center of Neurology and Psychiatry. This facility includes six isolation units for sleep and a shared living room with electromagnetic shielding for electroencephalography and a homeostatic environment of illuminance, humidity, and temperature. After the HS period, subjects went into the sleep laboratory and were monitored for 14 days. The schedule for each participant was established relative to his at-home records. The average bedtime calculated for each participant was set as 0 (24) h. In the BL period, 8-h time in bed (lights off) started at the mean bedtime observed in the HS period. In the 9-day ES period, bedtime and wake-up time were advanced and delayed, respectively, by 2 h each to make the total time in bed 12 h. At the wake-up time on ES9, TSD began and lasted for 39 h. Then, subjects had RS for 12 h starting 1 h after their habitual bedtime.

During the scheduled wake period, subjects stayed in a living room (approximately 100 lx, 8.7 m × 7.8 m) inside the isolation unit and were allowed to read and write, enjoy music and videos, play videogames, and engage in conversation with a researcher. When falling asleep, subjects were verbally awakened by attendant staff. During the scheduled sleep period, subjects were instructed to sleep in a sleeping room with the lights off (approximately 0 lx) next to the living room and were prohibited from going outside the room except for going to the lavatory. Subjects had three meals per day that met their body mass index and were restricted from unplanned meals and drinks. Ambient temperature and relative humidity in the laboratory were maintained at 25 ± 0.5°C and 50 ± 5%, respectively.

Throughout the study period, overnight PSG was performed in a soundproof sleeping room using the Neurofax EEG-1200 system (Nihon Kohden Corporation, Tokyo, Japan) and the following settings: sampling rate of 200 Hz, time constant of hardware high-pass filter of 0.3 s, and hardware low-pass filter of 30 Hz. In addition to electrocardiography, electrooculography, and chin electromyography, PSG data were recorded through Ag-AgCl electrodes placed at Fz, F3, F4, Cz, C3, C4, Pz, O1, and O2, with both earlobes as references, in accordance with the International 10–20 System. Two technicians, who were blinded to the experimental conditions, visually inspected PSG data divided into 30-s epochs and determined sleep stages according to the international standards for sleep classification (43).

Sleep parameters including total sleep time (TST: total time of NREM and REM sleep observed in TIB) were calculated from PSG data obtained every night. The exponential curve was fitted to the TST of all 15 participants, and individual OSD was derived from the lower bound of the curve, as detailed previously (16).

Subjective mood was evaluated using the State-Trait Anxiety Inventory-State Anxiety scale (STAI-S) (44) and the Positive and Negative Affect Scale-Negative Affect score (PANAS-NA) (45), and resting ASL imaging was performed 9 h after the midpoint of BL1, ES4, ES9 sleep, and TSD, that is, 5 h after the BL1 wake-up time, 3 h after the ES4 and ES9 wake-up time, and 27 h after the start of TSD.

Subjects had the same breakfast (approximately 350 kcal) 1 h before the MRI and moved into the MRI room. Immediately before imaging analysis, subjects completed the STAI-S and PANAS-NA in the behavioral laboratory next to the MRI room in order to rate current subjective mood. Then, ASL imaging was performed for 4 min. Subjects were instructed to keep their eyes open and look straight ahead to prevent them from sleeping during imaging. We measured subjects’ sleepiness during ASL scanning using the Karolinska Sleepiness Scale (KSS) (46). The KSS consists of a nine-point one-dimensional scale of sleepiness. Immediately after each scan, the subjects were asked about their sleepiness during the scan.

We obtained ASL imaging between 1800 and 2000 hours on BL1, ES4, ES9, and TSD using a Magnetom Verio 3T MRI System (Siemens). Whole-brain perfusion imaging was performed using a pulsed ASL technique called quantitative imaging of perfusion using a single subtraction, second version with thin-slice TI1 periodic saturation (Q2TIPS). In ASL, arterial blood is magnetically labeled by radiofrequency inversion pulses just below the intracranial area, including the carotid artery. By measuring the difference between the labeled and unlabeled control images, we measured arterial blood perfusion that is not affected by other brain areas and calculated rCBF/100 g/min based on the perfusion value. We obtained 45 time series whole-brain images that were already contrasted (labeled minus unlabeled). The imaging parameters were as follows: repetition time = 2.8 s, echo time = 13 ms, bolus duration (TI1) = 0.7 s, inversion time (TI2) = 2.0 s, number of slices = 12, slice thickness = 8 mm, total labeled scan volume = 45, total unlabeled scan volume = 45, total scan time = 4 min 13 s, and units = mL/100 g/min.

We also obtained a structural image at the time of health screening in each subject using a T1-weighted magnetic resonance rapid gradient-echo (MPRAGE) sequence, with the following parameters: repetition time/echo time = 1,900/2.52 ms, voxel size = 1 mm × 1 mm × 1 mm, flip angle = 9°, and field of view = 256 × 192. The image was used as a spatial reference in the analyses of functional data. Two subjects’ ASL data from ES4 were lost for technical reasons.

Preprocessing and statistical analysis of ASL data were performed using the ASLtbx tool (47) for Statistical Parametric Mapping 12 analysis software (Wellcome Department of Imaging Neuroscience, University College, London, UK¹). After correction for body movement and slice timing, co-registration to the reference MPRAGE image, and Gaussian smoothing using a kernel of 8 mm full width at half-maximum, rCBF maps were generated by measuring the difference between control and labeled images. Then, the Montreal Neurological Institute template was used to perform spatial normalization. Forty-five rCBF maps thus obtained were averaged to create a mean rCBF map for each imaging analysis.

The ROI was established a priori, and the mean intensity within the ROI was calculated for the MRI data obtained at each time point (BL1, ES4, ES9, and TSD). By use of the MarsBaR ROI toolbox for SPM 12, the peak of the ROI (7 mm in radius) was set at the coordinates x = 26, y = −4, z = −20 in the right amygdala and x = −20, y = −6, z = −18 in the left amygdala (Figure 3A). These coordinates were extracted from 639 images stored in the online database at http://Neurosynth.org (48) by using “emotion” as a search keyword. For statistical analysis, the mean blood flow in the voxels in the amygdala ROI was calculated and used in the statistical analyses.

In ASL-FC analysis, a cluster of voxels in the amygdala ROI was set as a seed region bilaterally, and the time series of 45 rCBF maps were analyzed using the CONN toolbox for SPM (Alfonso Nieto-Castanon²). A general linear model was applied to the individual time series of signals in each voxel across the whole brain, for four different conditions in each, to estimate a temporal correlation with the rCBF signals in the bilateral amygdala ROIs. The analyses generated ASL-FC beta images representing voxel-wise estimates (correlation coefficients) across the whole brain. To remove noise generated by body movements or physiological processes other than brain activities, the time series of signals of the white matter and cerebrospinal fluid was used as a regressor (the component-based noise correction method aCompCor) (49) and a bandpass filter was applied with the range of 0.008–0.12 Hz. The beta images were entered into the second level, namely, between-subject analyses.

Based on the hypothesis that interaction between the prefrontal cortex and amygdala is involved in mood fluctuations occurring during recovery from sleep debt, the prefrontal cortex was set as the ROI on the ASL-FC beta images under the four different sleep conditions. To extract brain areas with significant connectivity to the amygdala irrespective of the different conditions, the F test was performed to analyze connectivities in the beta images created for all four sleep conditions from all subjects (with the two subjects excluded for only ES4 due to loss of data) using the contrast weights vector in the SPM software: [1 0 0 0; 0 1 0 0; 0 0 1 0; 0 0 0 1]. This procedure extracted voxels with functional connectivity significantly larger (positive) or smaller (negative) than 0. Statistical significance was set at p < 0.05 (peak level, family wise error corrected) and k > 10 contiguous voxels.

The results of the analysis showed that, except for the cluster adjacent to the amygdala, the MPFC was the only brain area with functional connectivity with the amygdala (FC_amg–MPFC; Figure 4A). The mean value within the cluster in each subject was calculated using MarsBaR, which was used for statistical analysis.

Based on the hypothesis that the amygdala rCBF and FC_amg–MPFC explain negative mood during sleep debt (31, 36, 37), we performed multiple regression analyses using a structural equation modeling with all data obtained under the four conditions (n = 58) to evaluate the following models: (1) the FC_amg–MPFC predicts the amygdala rCBF and the amygdala rCBF predicts the PANAS-NA and STAI-S scores; (2) the amygdala rCBF predicts the FC_amg–MPFC and the FC_amg–MPFC predicts the PANAS-NA and STAI-S scores; and (3) the amygdala rCBF and the FC_amg–MPFC independently predict the PANAS-NA and STAI-S scores. To validate the model fit, we calculated several goodness-of-fit (GFI) indices such as GFI, adjusted goodness-of-fit (AGFI), the comparative fit index (CFI), Tucker–Lewis index (TLI) (50), and Akaike’s information criteria (AIC). The global fit indices were also supported by a GFI > 0.85 (51), an AGFI > 0.80 (52), a CFI > 0.90, and a TLI > 0.90 (51, 53). The AIC is a relative index to compare the multiple model fits, with the smaller value indicating a better fit (54). We modified the model in a stepwise manner based on the Modification Index, which indicates the path after the modification to improve the model fit and effect size, calculated by AMOS software. For each fixed and constrained parameter, the Modification Index is defined as (N/2) times the ratio between the squared first-order derivative and the second-order derivative (55).

Our research question was twofold, namely, the intervention effects of (1) ES and (2) TSD on participants’ subjective reports on their feelings and their brain activities regarding emotional processes. To investigate the effect of ES, two paired t tests, between BL1 and ES4 (with two subjects excluded for only ES4 due to loss of data) and between BL1 and ES9, were performed for each behavior and MRI data with Bonferroni correction (adjusted p values (56) are shown). Two paired t tests, between BL1 and TSD and between ES9 and TSD, were performed with Bonferroni correction to determine whether the TSD would erase the effect of sleep extension and whether the subjective and neural indices after TSD would be comparable to those at the BL level (i.e., before ES).

Because PANAS-NA scores did not follow a normal distribution in the Kolmogorov–Smirnov test (p = 0.20), a non-parametric test (Wilcoxon signed-rank test) was used for an analysis involving the PANAS-NA. Structural equation modeling was performed using SPSS AMOS 23.0.0 and other statistical analyses were performed using SPSS PASW Statistics 18, with the statistical significance set at p < 0.05. Descriptive statistics are expressed as means ± SEM.

On average, the subjects’ in-lab sleep duration at the BL visit did not differ from their in-home HSD. The mean HSD calculated from Actigraph data recorded at home was 7.37 ± 0.27 h, whereas the mean TST calculated from PSG data recorded in the isolation unit on BL1 was 7.35 ± 0.08 h. No significant difference found between the TST on BL1 and the HSD (t[14] = 0.048, p = 0.962). However, once they were allowed to sleep freely at the beginning of the ES period, subjects began to sleep excessively (>10 h; Table 1), suggesting a compensatory process for the PSD that had accumulated in everyday life. Sleep time gradually reduced with time, approaching their OSD between ES4 and ES9 (16). Their home HSD was significantly shorter than the estimated OSD (HSD = 7.37 ± 0.18 h, est. OSD = 8.41 ± 0.18 h, t[14] = 2.858, p = 0.0126; Table 1), which suggests the existence of PSD in the participants.

As seen in Figure 2, subjective sleepiness was significantly reduced by sleep extension, from its initial stage: the KSS score was significantly lower on ES4 and ES9 than on BL1 (t[14] = 5.527, adjusted p < 0.001 and t[14] = 6.918, adjusted p < 0.001, respectively). Negative affect was also reduced by sleep extension, with a significantly lower PANAS-NA score on ES9 than on BL1 (Z[14] = 2.41, adjusted p = 0.032), but not in the early stage (no significant difference between ES4 and BL1, t[14] = 1.122, adjusted p = 0.524). State anxiety showed the same trend, with a marginally lower STAI-S score on ES4 than on BL1 (t[14] = 2.349, adjusted p = 0.068) and no statistically significant difference between ES9 and BL1 (t[14] = 1.874, adjusted p = 0.164).

Sleepiness, which had been reduced by sleep extension, increased after TSD (significantly higher KSS score on TSD than on ES9, t[14] = 13.052, adjusted p < 0.001) and even exceeded the BL level (significantly higher KSS score on TSD than on BL1, t[14] = 5.684, adjusted p < 0.001). Negative mood, which had been reduced by sleep extension, slightly increased but the change did not reach statistical significance (no significant difference in PANAS-NA scores between TSD and ES9, Z[14] = 1.872, adjusted p = 0.122), although the negative mood became comparable to the BL level (no significant difference in PANAS-NA scores between TSD and BL1, Z[14] = 0.58, adjusted p = 1.0). In addition, once-reduced anxiety increased after TSD to a level comparable to the BL: the STAI-S score was significantly higher on TSD than on ES9 (t[14] = 2.864, adjusted p = 0.025), with no significant difference between TSD and BL1 (t[14] = 1.412, adjusted p = 0.359).

The resting-state amygdala activity, shown in Figure 3, was significantly reduced by sleep extension (significantly lower rCBF on ES9 than on BL1, t[14] = 2.56, adjusted p = 0.046), accompanied by a trend for a reduction in the initial stage (marginally lower rCBF on ES4 than on BL1, t[12] = 2.183, adjusted p = 0.099). The reduced amygdala activity tended to increase after TSD and became comparable to the BL level (marginally higher rCBF on TSD than ES9, t[14] = 2.283, adjusted p = 0.077, with no significant difference between TSD and BL1, t[14] = 0.248, adjusted p = 1.0).

The extent of negative functional connectivity between the amygdala and MPFC, shown in Figure 4, was increased by sleep extension, but not from the earlier stage of the sleep extension (larger FC_amg–MPFC on ES9 than on BL1, t[14] = 2.905, adjusted p = 0.023, with no significant difference between ES4 and BL1, t[12] = 1.35, adjusted p = 0.404). The increased negative connectivity then showed a trend for a decrease after TSD and became comparable to the BL level (marginally lower FC_amg–MPFC on TSD than on ES9, t[14] = 2.165, adjusted p = 0.096, with no significant difference between TSD and BL1, t[14] = 0.301, adjusted p = 1.0).

The fit indices at the first estimate of the three models supported model 1 (GFI = 0.86, AGFI = 0.54, CFI = 0.35, TLI = −0.3, AIC = 36.32) more than model 2 (GFI = 0.80, AGFI = 0.35, CFI = 0.09, TLI = −0.82, AIC = 44.04) and model 3 (GFI = 0.87, AGFI = −0.30, CFI = 0.35, TLI = −2.89, AIC = 38.25). Based on the Modification Index, we added a path between the residuals for the PANAS-NA and STAI-S to model 1 (considering the correlation between the two mood indices) and made the final model (Figure 5). GFI indices strongly supported the final model (GFI = 0.97, AGFI = 0.87, CFI = 0.96, TLI = 0.88, AIC = 19.24). In this model, the FC_amg–MPFC explained the variance in the amygdala rCBF (β = 0.29, p = 0.022), and the amygdala rCBF explained the variance in the PANAS-NA score (β = 0.319, p = 0.011). The amygdala rCBF did not explain the variance in the STAI-S (β = 0.197, p = 0.129).