Forty-three healthy, sedentary young adults (age range: 19–34 years, mean age: 25.33 ± 3.62 years, 23 females) were recruited for the study; one participant dropped out of the study, and two participants were excluded because of technical issues with the EEG recording. All subjects reported no signs of neurological or psychiatric illness and had a normal or corrected-to-normal vision. After providing informed consent, participants were pseudorandomly assigned either to an aerobic exercise group (n = 18) or a control group (n = 22), which were balanced in terms of age, sex, and fitness level (Table 1). Subjects received monetary compensation for their participation, and the experiment was carried out in accordance with the guidelines of the ethics committee of the Faculty of Medicine from the Otto von Guericke University Magdeburg (OVGU).

We assessed the consumption of oxygen (CPET Quark, COSMED, Italy) at the respiratory compensation point (VO₂-RC) by graded maximal exercise testing on a treadmill ergometer. The initial speed of 3 km/h was increased every 2 min to a maximum of 6.5 km/h. During this time, the slope of the treadmill also increased from 0 to 18%. This testing occurred until a respiratory exchange ratio (RER) of 1:1 was reached, indicating exhaustion near the limit of the cardiorespiratory system. Here, the oxygen exhaustion criterion was defined by the uptake at respiratory compensation (VO₂-RC) to better control for volitional effects. To assess lactate levels (Biosen C-Line, EKF Diagnostic, Magdeburg, Germany), capillary blood samples were taken from the earlobe during the resting state, at 2-min intervals during the fitness test, and 2 min after the maximum intensity. This fitness assessment was repeated after 4 weeks and again after 16 weeks, at the end of the intervention.

Electroencephalography recordings were acquired a day before and 2–3 days after the end of the 4-month intervention, while participants performed 120 trials of a DMS WM task (Figure 1A) and 60 trials of a VAS task Figure 1B). The trials from both tasks were divided into 4 and 2 blocks, respectively, which were presented in a random sequence. During the DMS task, a sample stimulus and a probe stimulus were presented in succession for 3 s, separated by a 5-s delay (i.e., the maintenance phase). Subjects were instructed to answer as correct and quickly as possible whether the sample stimulus was a public or a private place and to spend the remaining time memorizing the image. The maintenance phase is critical for sustaining the WM item and is susceptible to interference from other external factors. Afterward, participants had to decide whether the probe image was identical to (i.e., repeat) or a modification (i.e., lure) of the sample image. All of the DMS stimuli were computer-generated indoor scenes, including 50% lure images. During the VAS task, participants were shown a similar indoor scene and were instructed to detect whether a target (i.e., spherical object) was present. The background image for the VAS remained the same throughout the different blocks, so neither encoding nor maintenance was required for this task.

Continuous electroencephalogram (EEG) was acquired from 32 active electrodes mounted in an elastic cap (Brain Products GmbH, Germany) with a bandpass filter of 0.1–250 Hz and digitized at a rate of 500 Hz. We placed the electrodes according to the international 10–20 system. From the 32 electrodes, 2 electrodes were positioned at the external ocular canthi of each eye, and a vertical electrooculography (EOG) was placed below the left eye to measure horizontal and vertical eye movement, respectively. The left mastoid was used as the online reference, and all electrode impedances were kept below 5 kΩ. The skin under the electrodes was slightly abraded with a blunt needle, which was used to fill each electrode with electrolyte gel. MATLAB (MATLAB and Statistics Toolbox Release, 2012b, MathWorks, Inc., Natick, MA, United States) and its EEGLAB toolbox (Delorme and Makeig, 2004) were used for offline EEG data processing. The continuous EEG data were high-pass filtered at 0.1 Hz, low-pass filtered at 50 Hz, and referenced to the right mastoid. The data were segmented into 13-s epochs, including 1-s pre-stimulus as the baseline and 1-s post trial. Next, independent component analysis (ICA) was performed to correct for eye-related artifacts and excessive muscle activity. In addition, all trials were visually inspected, and those containing additional artifacts were excluded from further analysis. Participants included in the data analysis had less than 20% of their trials excluded after artifact correction.

Event-related spectral perturbation (ERSP) was computed using algorithms from EEGLAB (Delorme and Makeig, 2004) and custom MATLAB scripts. Artifact-free data comprising 13-s segments were used for time-frequency analysis (TFA). Trial-by-trial event-related spectral power was calculated using Hamming window tapering with five cycles and 100 logarithmically spaced frequencies ranging from 3 to 20 Hz. For each frequency, event-related spectrum power at each time-frequency point was divided by the average spectral power in the pre-stimulus baseline period at the same frequency. These measures were normalized by taking the log value of the percentage of baseline activity (ERSP_%) (Grandchamp and Delorme, 2011). By definition, the unit of ERSP_log is the decibel (dB), which is commonly used in the literature (Makeig, 1993; Cohen and Donner, 2013; Jiang et al., 2015).

The mean ERSPs were calculated for two frequency bands of interest: theta (5–7 Hz) and alpha (9–12 Hz). These bands were chosen because of their involvement in visuospatial attention with the frontoparietal network (Wang et al., 2015). Because of frequency smearing after time-frequency decomposition, we used a width slightly different from the typically defined theta (4–8 Hz) and alpha (8–12 Hz) frequency bands (Klimesch, 1996). Narrower frequency bands were chosen to better characterize the changes in neural oscillations at those frequency bands (Hsieh et al., 2011). Additionally, electrodes were grouped into two different clusters with three electrodes each: frontal cluster (F3, Fz, and F4) and posterior cluster (P3, Pz, and P4). To measure if there is an exercise-induced effect, we focused on the sample probe. That is, power estimates from the frontal and posterior clusters were averaged across a section spanning the frequency band (see Figures 2A,B) and the time period of interest (0–3000 ms), and then subjected to statistical analysis.

All statistical analyses were performed using the statistical software SPSS (IBM Corps., IBM SPSS Statistics, V24, Armonk, NY, United States, 2016). First, we ran independent-samples t-tests to assess whether there were any differences between the groups at baseline (preintervention). To compare differences in categorical variables such as sex and age, we used the Chi-square test. There were no differences observed between the groups at baseline (Table 1).

Furthermore, to examine the possible effect of training on cognition, the behavior data were submitted to separate rmANOVAs (one ANOVA per condition), where the within factor was time (pre, post), and the between factor was group (exercise, control). As for the EEG data, these were submitted to two separate repeated measures ANOVAs (one rmANOVA per frequency band of interest). The factors in the rmANOVAs included task (DMS, VAS), cluster (frontal, posterior), and time (pre, post intervention) as within-subject factors and group (exercise, control) as a between-subjects factor. We did not include any variables as covariates since groups did not differ in demographics. When appropriate, the Greenhouse–Geisser correction was used to adjust the degrees of freedom when the sphericity assumption was violated. All alpha levels for significance were set at 0.05. When significant interactions were found with rmANOVA, post hoc independent t-tests were conducted.

To assess the possible relationship between exercise and cognition, we first calculated the intervention-related changes in performance, spectral power values, and fitness. In particular, accuracy, RT, and EEG spectral power differences were obtained by subtracting pre from post measures of EEG power. For a better estimation of aerobic fitness increase, a composite fitness score was calculated separately for pre- and postintervention fitness test data as a mean of the inverse z-scores of blood lactate level changes (%) and the z-score of VO₂-RC changes (%). The bivariate (Pearson) correlations were calculated between changes in ERSP values, fitness scores, and behavioral performance to examine how fitness changes affect cognitive functioning after 4 months of exercise training.

Demographics and fitness levels for the exercise and control groups are reported in Table 1. As shown in Table 1, the groups were matched for age and sex and did not differ at baseline in body mass index (BMI), average blood lactate, and VO₂-RC [F(1,38) < 1.35; p > 0.198].

To measure exercise-induced changes in fitness, we performed rmANOVAs with time (pre- and postintervention measures) as the within-subject factor and group (exercise and control) as the between-subjects factor. As seen in Table 2, rmANOVA revealed a time × group interaction for lactate [F(1,38) = 36.93; p < 0.001] as well as for VO₂-RC [F(1,38) = 86.50; p < 0.001]. Post hoc paired t-tests showed an increase in fitness represented as a decrease in lactate [t (17) = −9.94; p < 0.001] and an increase in VO₂-RC [t (17) = 2.27; p < 0.001] for the exercise group but not the control group [t (21) = −0.25; p = 0.809 and t (21) = −1.62; p = 0.121, respectively].

Table 3 illustrates the pre- and postintervention measures for accuracy and RT for each group. We considered performance values that were greater than 2.2 SDs from the mean as outliers, and these values were excluded from the statistical analysis. rmANOVAs revealed a main effect of time for all behavioral performance measures, for the exception of the correct response [F(1,36) = 2.00; p = 0.166; see Table 3]. Although the time × group interactions were all not significant (F < 1.96; p > 0.172), meaning that the difference in behavioral performance between the pre and post measures was not due to our intervention. Figures 3A–D further illustrates this point by showing the similar exercise-induced before and after performances between the exercise and the control group.

To test the fundamental relationship between spectral EEG power and behavioral performance, Pearson’s correlation analyses were performed on baseline measures across participants. For the DMS task, we did not find a correlation between FMT and behavioral performance during the encoding phase although there was a negative correlation between FMT and accuracy (corrected hit rate = hit−false alarm) during the maintenance phase [r (36) = −0.344; p = 0.040; Figure 4A], where better performance was associated with a greater decrease in theta power relative to baseline. Also, FMT correlated with RT, where faster detection of a lure was associated with greater decrease in theta power [r (39) = 0.431; p = 0.006; Figure 4B]. Notably, Figure 4B seems to have a high leverage point on the far right, although after obtaining the Mahalanobis distance, it was not excluded. Together, these correlations indicate that a greater decrease in theta during the maintenance conveys a more accurate and faster RT. As for the VAS task, correlations were performed between behavioral performance and posterior alpha power, which is where alpha power is most prominent and known to be related to attention (Frey et al., 2015). Our results showed a positive correlation between alpha power and RT during target absent trials [r (38) = 0.333; p = 0.041] and a marginal positive correlation for RT during target present trials [r (38) = 0.284; p = 0.084] but not for accuracy [r (40) = 0.020; p = 0.902].

Repeated measures ANOVAs performed for the sample stimuli presentations (0–3 s) for theta power revealed a main effect of task [F(1,38) = 18.87; p < 0.001], no main effect of cluster [F(1,38) = 1.15; p = 0.290], and no main effect of time [F(1,38) = 0.01; p = 0.921]. The interaction between time and group was not significant [F(1,38) = 0.51; p = 0.480], as well as the task × cluster × time × group interaction [F(1,38) = 0.22; p = 0.645]. For the alpha band during the sample stimuli presentations (0–3 s), there was a main effect of task [F(1,38) = 25.09; p < 0.001], a main effect of cluster [F(1,38) = 17.76; p = 0.000], and no main effect of time [F(1,38) = 2.45; p = 0.126]. There was however, a significant task × cluster × time × group interaction [F(1,38) = 4.26; p = 0.046]. Post hoc independent t-tests showed no group difference for the posterior cluster [t (38) = 32; p = 0.751], while the frontal cluster showed a significant group difference [t (38) = 2.34; p = 0.025] during the VAS task, with increased alpha power after the intervention (see Table 4). This effect was specific to attention, as it appeared during the VAS task but not the DMS task [t (38) < 1.52; p > 0.136]. This EEG finding is further illustrated in Figures 5A,B in a histogram and a topographic map.

Furthermore, given the significant post hoc results, we sought to examine a direct link between anterior alpha power changes with aerobic fitness changes and behavior performance changes (during the VAS task). Such correlations were performed irrespective of the group (i.e., across all subjects) since fitness changes were also present in the control group. Specifically, there was a positive correlation between changes in frontal alpha power and changes in aerobic fitness [r (40) = 0.379; p = 0.016; see Figure 6A]. Additionally, changes in frontal alpha power were also positively correlated with changes in accuracy [VAS task – see Figure 6B, target absent, r (37) = 0.336, p = 0.042; target present, r (39) = 0.302, p = 0.066]. Although a correlation between changes in aerobic fitness and changes in accuracy (p > 0.423) was absent. However, changes in aerobic fitness did correlate negatively with RT for the target present condition [r (38) = −0.360; p = 0.026] and marginally for the target absent condition [r (37) = −0.297; p = 0.074], yet failed to correlate with changes in alpha power (p > 0.345).

A cross-sectional correlation analysis derived from baseline measures revealed that greater event-related desynchronization (ERD) in theta power was negatively correlated with memory performance (in the DMS task; Figure 4). Thus, stronger theta desynchronization in the maintenance phase was associated with better mnemonic discrimination and faster RT (in lure trials), which is in line with previous findings in the literature (Greenberg et al., 2015). Notably, the correlation between frontal theta power and performance was specific to lures (RT), potentially indicating that a decrease in frontal theta power associates with trials that require greater brain resources. In addition, the time-frequency representations (TFRs) for alpha power are in accordance with previous studies stating that alpha ERD occurs during a broad range of cognitive tasks, where harder tasks elicit a more negative ERD (Klimesch, 1999). Also, our findings revealed a positive correlation between alpha power and RT during target absent trials, where faster RT had greater ERD at posterior sites, further illustrating the importance of alpha desynchronization in attention processing. And, even though alpha power is most prominent in the parietal region, alpha ERD has also been observed in the frontal cortex during the performance of a selective attention task (i.e., Stroop task; Chang et al., 2015), indicating that alpha ERD reflects a top-down process.

We observed an exercise-related increase in alpha power (i.e., weaker ERD), predominantly at the frontal electrodes during the VAS task (see Figure 5). Note that an increase in alpha power stems from a reduction in alpha ERD in the postintervention session. That is, overall subjects had similar performance levels, but the exercise group was characterized by reduced alpha desynchronization (i.e., less cortical activity) to accomplish the same task. Tentatively, we attribute this increase in alpha power to neural efficiency, specifically to an enhanced capacity for resource allocation involving visual attention. Our findings are compatible with the Ludyga et al. (2016) study in which they measured activity before and after a month of a cycling intervention, and also observed improvements in fitness performance and decreased cortical activity during exercise (Ludyga et al., 2016).

Furthermore, neuroimaging studies have demonstrated that subjects with higher WM and spatial skills have weaker frontoparietal activation during cognitive tasks (Rypma and D’Esposito, 1999). Similar results have been found in EEG studies where some have indicated that elite athletes (e.g., shooters, karatekas, and gymnasts) require “less” cortical activation (weaker ERD) in task-relevant brain areas than novices during sport-specific tasks (Ludyga et al., 2016) and a judgment-allocating task (Del Percio et al., 2009; Babiloni et al., 2010). This observation has been coined “neural efficiency” (Del Percio et al., 2010; Maffei et al., 2017). Studies that have taken aerobic fitness into consideration have come to similar conclusions (Hogan et al., 2015; Ludyga et al., 2016). For instance, a cross-sectional study performed by Hogan et al. (2013) found that preadolescents with lower fitness require more neural resources for a given task, which was reflected by an increased EEG coherence in comparison to the fit subjects.

Moreover, our results showed that participants with greater improvements in fitness had a larger increase in alpha power in the anterior regions (Figure 6A). Other cross-sectional studies have also found comparable correlations (Weinstein et al., 2012; Chu et al., 2016; Hyodo et al., 2016). An exercise intervention study in older adults found that higher aerobic fitness derived from the walking program was associated with greater changes in white matter integrity in the frontal and temporal lobes (Voss et al., 2013). Given these previous findings, it is congruent that we observed an increase in alpha power reflected in the anterior electrodes, which also correlated with changes in accuracy (Figure 6B). Notably, such correlations (alpha-fitness and alpha-performance) were existent across groups, implying that alpha power modulations may also be sensitive to other physiological changes. Fitness changes, for instance, might not only be driven by the intervention, but also other parameters such as daily routine activities, adaptability in training (Bonafiglia et al., 2016), and a genetic disposition to physical activity could also be playing a role (Bouchard, 2012). Nonetheless, only our exercise group showed less cortical activation to performed the VAS task compared to the control group, as shown by the increased alpha power after the intervention, which we attribute to neural efficiency.

Our study has some limitations. First, we had a limited sample size of 18 subjects comprising the exercise group. Further intervention studies will have to confirm our findings in larger samples, and it would be beneficial to lengthen the intervention duration to determine whether the training benefits transfer to hippocampal-dependent tasks and oscillations in young adults. On the other hand, our sedentary group was matched in age, sex, and baseline fitness level. Second, we did not have a tracker to measure outside activity; however, we encouraged all participants to avoid changes to their lifestyle outside the laboratory for the time of the intervention. The sedentary group of young adults who underwent our vigorous exercise training showed improvements from our 4-month training intervention, which we measured with not one but two fitness markers providing a more precise indication of the aerobic fitness change. Hence, the negative findings that we observed here were not due to an ineffective intervention.

Taken together, our study provides tentative evidence in support of cardiovascular exercise modulating oscillatory brain activity. Also, our findings support the possibility that aerobic training of sedentary young adults may influence neural dynamics underlying visual attention. However, we did not confirm our hypothesis that aerobic exercise would enhance theta oscillations and improve WM performance in a mnemonic discrimination task.