This study involved 26 participants (18 men and 8 women; mean age: 28 years, SD = 2.1). Participants were graduate students without any mental or neurological diseases or related family history of disease. This study was carried out in accordance with the recommendations of the local ethics committee with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the local ethics committee.

The task was designed using E-prime software (Eprime2.0, Psychology Software Tools, Inc., Sharpsburg, PA, USA). We used the classical Sternberg paradigm to design the letter WM experiment. The participants were instructed to sit in front of a computer screen, and keep their gaze at the center of the screen. The letter memory experiment protocol is shown in Figure 1. First, a symbol, “+”, appeared on the screen for 1500 ms to alert the subjects, followed by five target letters, which appeared one at a time for 500 ms each. Participants were asked to memorize all the target letters (memory coding stage). Next, a black screen appeared for 6000 ms, while participants tried to recall the letters (retention period). Test letters were presented on a black screen for 2000 ms each, and participants were asked to judge whether the letter was the same as one of the memorized target letters; they pressed “1” for “yes” and “2” for “no” (memory matching stage). The rTMS was applied at the beginning of the memory maintenance stage for 3000 ms. EEG data obtained after applications of the rTMS were used for subsequent EEG analysis.

There were six groups in the experiment and 20 trials in each group. The interval time between groups was 2 min, and that between trials was 10 s. Half of the groups received rTMS, whereas the other half received sham rTMS. All the groups were randomized so that participants did not know which treatment they received.

For the rTMS session, participants were seated comfortably and their heads were localized in space via a three-dimensional navigation system (Brainsight, Magstim, Carmarthenshire, UK). Prior to the start of the behavioral task, resting motor threshold (RMT) was determined for each participant using electromyography (Matrix, Micromed, Mogliano Veneto, Italy). The target for RMT was the left primary motor cortex, which controls the right first dorsal interosseous muscle. RMT was determined as the minimum stimulus intensity capable of eliciting muscle activity of more than 50 μV in at least 5 of 10 consecutive TMS.

We used the RMT to calibrate the stimulation intensity for each subject, starting at 110% of RMT and accounting for scalp-to-cortex distance for each targeted brain area (Stockes et al., 2006). A 70-mm figure-of-eight coil (Double 70 mm Coil, Magstim) was used during the first 3000 ms of the maintenance stage (total 15 plus) for the 5-Hz rTMS, which is similar to theta band frequency. To avoid sound interference produced by rTMS, participants were required to wear earplugs. The location of the coil was tangential to the scalp, and the hand shank was vertical to the gyrus of the target cortex. In the “sham” condition, a similar rTMS coil was applied, which was located in the same position.

Prior to determining the target location from the cortex, whole-brain images were acquired with a 3T scanner (Sigma VH/I, GE Medical Systems, Milwaukee, WI, USA). High-resolution T1-weighted images (256 axial slices, 0.5 mm × 0.5 mm × 0.8 mm) were obtained for all participants. This scan was used to reconstruct a 3-dimensional image of each participant’s head, which was then used to target rTMS to the superior parietal lobule (SPL). The SPL is located between the posterior central gyrus and the intraparietal sulcus (middle area between the P3 and P1 electrodes near Brodmann area 7).

A 32-guide EEG acquisition system (Synamps2, CompumedicsNeuroscan, Charlotte, NC, USA) with a titanium alloy electrode cap (Maglink, CompumedicsNeuroscan) to shield from rTMS interference was used to record EEG. We used the International 10–20 system to apply electrodes (FP1, FP2, F7, F3, Fz, F4, F8, T7, C3, Cz, C4, T8, TP7, CP5, CP3, CP1, CPz, CP2, CP4, CP6, TP8, P7, P5, P3, P1, Pz, P2, P4, P6, P8, O1 and O2). The ground electrode was located in front of the forehead, and reference electrode was placed in the middle of the “Cz” and “CPz” electrodes. Vertical and horizontal eye movement electrodes were also applied. All electrode impedance values were <5 kΩ. The sampling rate was 1000 Hz.

For statistical analysis of behavioral data, the paired t-test was adopted. The study detected significant differences for the “AR” and the “RT” of subjects between the two conditions, and evaluated the effects of rTMS on parietal brain regions. Bonferroni correction was used for statistical result adjustments.

Permutation test was conducted to test hypothesis as stated in the “Introduction” Section. A permutation test is usually used to assess statistical significance. As a test based on computer simulation, permutation test is especially suitable for small sample data sets, such as nonparametric tests. Moreover, permutation test provides an efficient approach to testing when the data do not conform to the distributional assumptions of the statistical method one wants to use (e.g., normality). The theory for permutation test has evolved from the works of Fisher (1935) and Pitman (1938). Pesarin (1992) developed the method, and the principle is as follows.

First, assuming two samples were assigned as X1= (x₁₁, x₁₂…x_1n1) and X2 = (x₂₁, x₂₂…x_2n2). Where n₁ and n₂ are the numbers of samples (X1 and X2). Second, the statistics of the original observations T₀ were calculated. Third, the order of the sample data is replaced by a large number of random permutations to calculate the new statistics Ti, (i=1,2,…,C=(n1+n2n1)). The P value of the statistical test is calculated by formula: (2)p=Prob(T≥T0)=1C∑i=1CI(Tiperm>T0)

In this study, the characteristics of the results from the spectral and inter-sites phase synchronization analysis were consistent with the requirements of permutation test (small sample size (26 participants), and not satisfying normal distribution). Therefore, the permutation testing was used to evaluate significant differences of EEG analysis data between the rTMS and sham conditions. The statistical results were Bonferroni corrected.

For the spectral analysis, all of the electrodes (32 electrodes) were involved in the statistical analysis. The ERSP data were averaged both in time (retention period) and frequency (theta band and alpha band) domains in each electrode. The permutation testing was used to evaluated the significant difference between the two conditions (ERSP-rTMS and ERSP-sham).

For the phase synchronization analysis, there are about 496 electrode pairs (due to the 32 electrodes) were involved in the statistical analysis. The PLV data calculated for each electrodes pair. The permutation testing also was used to evaluate the significant difference between the two conditions (PLV-rTMS and PLV-sham).

The behavioral results from the WM tasks included AR and RT values. The Pauta criterion was imposed to reject abnormal trials as follows: if the difference between the RT value from one trial and the averaged RT value for all trials was larger than three times the standard deviation (SD) of the averaged RT, then the trial was rejected. Using this criterion, 2 ± 0.8 trials were rejected in one participant.

Paired t test showed that AR values were not significantly different between the rTMS and sham groups (t₍₂₅₎ = 2.1, P > 0.05, see Figure 2). In contrast, RT values were significantly shortened in the rTMS group compared to the sham group (t₍₂₅₎ = 3.5, P < 0.05*, see Figure 2), which indicated that rTMS had effectively improved the task performances.

For the theta oscillation, ERSP results showed that there was a significant difference between the rTMS and sham groups. The ERSP values of theta oscillations (Figure 3A) were increased markedly by rTMS treatment in the left SPL area, which included P7 (P < 0.03*), P3 (P < 0.05*), P1 (P < 0.01**), and CP1 (P < 0.04*; Figure 3B). Similar result was found near the prefrontal area, except that only the ERSP values from the Fz (P < 0.03*) electrode was significantly increased with rTMS (Figure 4B).

In contrast to the theta band results, the EEG mapping (Figure 4A) for the alpha band showed that, the ERSP values in the left SPL region, which included CP3 (P < 0.04*), CP1 (P < 0.03*), CPz (P < 0.05*) and P1 (P < 0.01**) markedly declined with rTMS effect (Figure 4B).

Due to individual differences in brain capacity, performances on WM tasks differed across participants. The RT differences were significantly correlated with the EEG ERSP values differences between the rTMS and sham groups for the theta (Figure 5A) and alpha (Figure 6A) oscillations.

A positive correlation was observed at the theta frequency band, which indicated the theta oscillation ERSP values increased when the RT differences increased between the rTMS and sham groups. This phenomenon was localized mainly to the left brain regions, including the left prefrontal area, left central area, left central-parietal area and left parietal area. The electrodes that had the significant positive correlations by rTMS treatment included F7 (r = 0.425, P < 0.05*), F3 (r = 0.451, P < 0.05*), C3 (r = 0.356, P < 0.05*), CP3 (r = 0.621, P < 0.05*), CP1 (r = 0.715, P < 0.01**), CP5 (r = 0.657, P < 0.01**), P5 (r = 0.546, P < 0.05*), P3 (r = 0.553, P < 0.05*), P1 (r = 0.625, P < 0.05*), and Pz (r = 0.356, P < 0.05*). The most marked difference was observed at CP1 (Figure 5B), which was located at the left parietal area. Previous studies had shown that object memory was associated with the left-brain cortex. Thus, this result suggested that the left-brain cortex was more involved in the “retention” process of WM.

In contrast, a negative correlation was observed between RT and alpha frequency band, which indicated that the alpha oscillation ERSP values decreased when the RT differences between the rTMS and sham groups increased. This phenomenon was localized to the bilateral prefrontal area, left central area, left central-parietal area and left parietal area. The electrodes that had the significant negative correlation by rTMS treatment included F3 (r = −0.559, P < 0.01**), Fz (r = −0.726, P < 0.05*), F4 (r = −0.535, P < 0.01**), C3 (r = −0.527, P < 0.05*), CP1 (r = −0.607, P < 0.05*), CP3 (r = −0.52, P < 0.05*), CP5 (r = −0.661, P < 0.01**), P7 (r = −0.601, P < 0.05*), and P1 (r = −0.638, P < 0.01**). The most marked correlation was observed at Fz (Figure 6B), which was located at the prefrontal area. Similar brain regions exhibited significant correlations with the two frequency bands. In addition, the brain regions that correlated with the alpha frequency band covered a larger area than those that correlated with the theta frequency band.

Multiple brain areas, including the prefrontal area, central area, and parietal area are involved in the “retention” process of WM. However, it remains unclear whether connectivity exists between the parietal area and other brain areas following rTMS of the parietal lobe. Many prior studies have shown that rTMS could affect brain activities when the frequency in the brain region was similar to the stimulus frequency. In this study, 5-Hz rTMS enhanced theta-band amplitude and decreased alpha-band amplitude in different brain areas.

Via the PLV comparison between the two conditions (rTMS vs. sham), we found that phase synchronization between the parietal area and other brain areas became markedly stronger when the parietal lobe was stimulated by rTMS. For the theta oscillation, phase synchronization between the left parietal area and left prefrontal area was strengthened by rTMS (Figure 7A, P < 0.01**). A similar phase synchronism was found under alpha oscillation (Figure 7B, P < 0.01**). This finding suggested that the stimulation of the parietal brain region resulted in amplitude alterations of the prefrontal region during the “retention” period of WM.

In EEG studies of WM, brain activities in the parietal and prefrontal areas were related to coding, maintenance and processing of information. Several studies have shown that theta oscillation ERSP value in the prefrontal area increases during learning and consolidation of associative memory (Jensen and Tesche, 2002; Meltzer et al., 2008; Takehara-Nishiuchi et al., 2012; Enriquez-Geppert et al., 2014). Other findings uncovered that theta oscillation ERSP values in the parietal-occipital and temporal areas increased during the entire WM process, which is called the “gate” effect (Raghavachari et al., 2006). In addition, other studies on theta oscillation have shown that changes in ERSP value reflect correlations between different cortical areas. For cognitive task, inter-site phase coherence of theta oscillation reflects the correlation between two cortical areas, such as the parietal and prefrontal areas (Sauseng et al., 2004; Sammer et al., 2007; Wu et al., 2007; Liebe et al., 2012). In this study, theta oscillation ERSP values in both the left parietal and left prefrontal areas were positively correlated with behavioral data (e.g., RT). Further analyses showed that rTMS enhanced the phase synchronizations between these two cortical regions, which may have strengthened the connections between these two regions.

In contrast, some studies have suggested that alpha oscillation ERSP values were reduced during WM (Meyer et al., 2013; Tenke et al., 2015). In our study, the rTMS, alpha oscillation ERSP values were negatively correlated with RT at different cortical regions, including the prefrontal area, right central area and parietal area, following rTMS. This result suggested that these cortical regions were activated during the “retention” process. Furthermore, alpha oscillation synchronization between the parietal area and other cortical regions (including the prefrontal and central areas) also reflected their co-operative relationship during the “retention” process following rTMS. Meanwhile, we observed that alpha oscillation was increased on the lateral parietal area to the stimulation location, but was not significant (Figure 5B). The result could be interpreted as inhibition of the interfering information in the task-irrelevant brain areas (Klimesch et al., 2003; Hamidi et al., 2009; Benedek et al., 2014).

In this study, the oscillation ERSP value data showed variations only at the parietal area, and not the prefrontal area, for both theta and alpha oscillations. However, while establishing the correlation between the ERSP value and RT differences, the prefrontal area was also found to be activated during the process. What is the mechanism by which the activity of the prefrontal area is induced? Is this activity connected to parietal area activity? Several recent studies have reported connections between these two cortices. Kundu et al. (2015) found increased directed connectivity between the prefrontal and parietal areas during the “retention” process by using the Granger causality of the alpha oscillation; however, similar results were also observed between both the prefrontal and occipital areas, and the parietal and occipital areas during the “coding” process. In addition, Harding et al. (2015) established a brain network model for cognitive control and WM, which included the inferior frontal junction, dorsolateral prefrontal cortex (DLPFC), pre-supplementary motor area, and intra-parietal sulcus, and analyzed the causality between different cortical regions via functional magnetic resonant imaging (fMRI).

Furthermore, several studies on rTMS have explained the effects of increased correlation between the parietal and prefrontal areas on WM task performance. Wang et al. (2014) designed an associative memory experiment to test cortical-hippocampal brain networks. The target location for stimulation was the SPL. The hippocampus is associated mainly with cognitive functions such as memory and judgment, which are also associated with the prefrontal cortex and parietal area. Using fMRI data, the findings of the study revealed an association between hippocampal activations and better task performances, which suggested that cortical regions connected with the parietal area might be indirectly activated by parietal area stimulations. In our study, the PLV analysis results (Figures 7A,B) also showed there was synchronization between the parietal and prefrontal areas that was strengthened by rTMS. Moreover, the brain areas that showed synchronization were similar to those that showed correlations between oscillation ERSP value and behavioral data. These results indicated that the EEG activities in the prefrontal area were strongly correlated to the EEG activities in the parietal area during the WM task. Using rTMS, the prefrontal area may be activated indirectly by stimulating the parietal lobe. We had hypothesized that the rTMS may induce changes in EEG oscillations in regions of brain, and that the EEG phase synchronizations between these cortices would be enhanced by rTMS.

Previous studies have compared behavior performance and EEG activities of WM task participants between the rTMS and control groups. However, the original EEG data could not directly reflect whether the brain’s response was positive or negative. In our study, the positive behavioral performances (e.g., decrease in RT) were correlated with the EEG activities (theta or alpha band oscillations), which made the EEG analysis more functional. By investigating the correlation between behavior and EEG activity of different brain areas, we could find which cortical brain regions were involved in the responses of WM tasks.

In the present work, subjects achieved better performances of WM tasks with 5-Hz rTMS of the parietal area, while their theta band oscillation increased more in the left parietal and left prefrontal brain regions, and the alpha band oscillation decreased more in the left parietal and bilateral prefrontal brain regions. According to the EEG oscillation activities of the subjects, the left-brain region, including the left parietal and prefrontal areas, was affected by the action of 5-Hz rTMS. Therefore, these regions were more involved in brain activity during the “retention” period of WM, which surely increased their task performances.

In the spectral analysis, the preprocessing method contains two baseline removal procedures, which are referenced from Hamidi et al. (2008), and the method is also used in Makeig (1993) and other documents. In the first baseline removal procedure, the ERSP data during the “retention” period were obtained by ERSP analysis. In the second baseline removal procedure, the normalized ERSP data were obtained by subtracting the interval-period ERSP data from the ERSP data during the “retention” period. The main idea of this method through two processes, respectively, to remove short-term and long-term tasks irrelevant components. The traditional ERSP calculation can only remove task before and after the execution of a task independent component, but as everyone knows, brain activity will be changed greatly after a relatively long interval. Since the experiment took more than 1 h, the process which took a long time may be brought serious impact for the brain activity. In this study, the baseline removal procedure was divided the two steps, respectively, to remove irrelevant components of the task. However, this process may cause some useful components to be removed. This requires further research to evaluate and improve.

The calculation of PLV was affected by several factors. Because of the existence of the volume conductor effect, signals collected on either electrode could be considered as a synthesis of the electrical activity of all neurons. Therefore, the signals of any two electrodes had a lot of common components, which made the PLV higher than the actual value. The solution was to map the multichannel EEG back to the cerebral cortex by means of traceability analysis. At the level of the cortex, the coherence of electrical activity between the two brain regions was examined and this connection between the two brain regions could be described more accurately than that in the electrode space. However, there are some problems with this method, for example, most of the trace back algorithms take many simplifications, and the obtained electrical activity information was a statistical estimation result. In addition, different algorithms achieved inconsistent results, which add to the difficulty of the analysis. It is also important that, due to the existence of the volume conductor effect, the EEG signals actually acquired by the epidermal electrodes were convolution of many sources rather than simple linear sum. The estimation of the traceability algorithm was rarely taken into account at present, so the accuracy of traceability was still to be improved. The computation of PLV also had a problem of time window, in which the EEG is assumed to be stationary, and this will lead to some estimation errors. In conclusion, both the time-frequency features of ERSP and the phase synchronization of PLV were affected by many factors, and these results had some limitations. More comprehensive analyses should be made with reference to studies of multiple neurophysiological signals, such as functional magnetic resonance imaging (fMRI) or functional near-infrared spectroscopy (fNIRS).