Twenty-one elite swimmers (ES) from Shanghai Swimming Management Center (Supplementary Tables 1–4) and 23 college students with no history of specialized swimming or other professional sports training (control group, CG) were recruited. All participants read and signed informed consent forms. The study was approved by the Ethics Committee of the Fudan University. Participants were all right-handed. The handedness was determined by self-reports and verified by the observation of their hand use in writing and performing the task. None of them had reported a history of mental illness. The EEG and behavioral data of 5 elite swimmers and 7 controls were excluded from the analysis because their EEG signals contained too many instabilities or drifts. The final sample consisted of 16 ES and 16 CG (n = 32, 15 Female). Sixteen swimmers (13 master-level) were distributed to 5 specialties (6 Free style; 4 Backstroke; 3 Breaststroke; 2 Individual Medley; 1 Butterfly) (for participant demographics, see Supplementary Tables 1, 2).

Participants were seated comfortably in an armchair in a soundproof room. The experiment started with a 2-min waking eyes-closed (EC) period, followed by a 2-min eyes-open (EO) period with eyes fixed on a screen with a crosshair. Then, participants were required to perform the CR task while keeping the head still, followed by another EC (2 min) and EO (2 min) states. After each state, there was a 1-min short break. EEG signals were recorded throughout the process. Prior to perform the CR task, participants were instructed to practice several trails till they successfully completed one trail by themselves. In the CR task there were 8 trails and each trial involved 6 different blocks connected by directional arrows, and every block contained 4 balls connected by lines. The participant was required to use left/right finger or foot to indicate the location of the ball which appeared in one of the four corners of the screen (i.e., left finger = upper left, right finger = upper right, left foot = lower left, and right foot = lower right). If the participant responded correctly, the ball would disappear. Participants were required to respond as quickly and accurately as possible while keeping the head stable. The trial-by-trial reaction time and the total number of errors were documented. The experimental procedure is illustrated in detail in Figure 1.

EEG signals were recorded from an 8-channel EEG system (eego, ANT Neuro, Berlin, Germany) and digitized at a sampling rate of 1,000 Hz. The reference electrode was placed between the Cz and Pz channels, and others (Fpz, Fz, F3, F4, Cz, C3, C4, and Pz) were distributed around the frontal and parietal areas according to the extended 10–20 international system. The impedance of all electrodes was kept below 10 kΩ.

EEG data were preprocessed in EEGLAB v.13.0.0.b, a MATLAB-based open toolbox (Delorme and Makeig, 2004). Segments with a duration of 1 min (from 30 to 90 s of the collected 120 s data) were selected from the EC and EO resting-state EEG, respectively, and all data associated with the CR task were imported for preprocessing. Raw data were re-referenced to the common average reference and filtered to a frequency range of 0.5–30 Hz. After running eye blink recognition, artifacts associated with eye movements and blinks were removed by using the AAR1.3 toolbox plugin, which performed automatic electrooculogram (EOG) artifact correction using blind source separation (BSS) and identified the EOG components using fractal analysis (Gomez-Herrero et al., 2006). After that, the resting states and CR task data were all segmented into epochs with a duration of 2 s each. Subsequently, the epochs with abnormal values beyond the upper limit of 75 μV were rejected (Collin et al., 2012). Overall, 10.1% epochs from the resting states and 16.2% from the CR task were excluded due to artifact contaminations.

It's well-known that the alpha blocking phenomenon appears during relax wakefulness and conceptualized as desynchronized neural population activity during active stimuli. Previous study has demonstrated that the alpha blocking phenomena could reflect wakefulness-to-sleepiness levels (Jiao and Lu, 2017). They estimated the degree of falling in sleepiness from drivers' wakefulness by calculating the alpha blocking rate from EEG wave. In present study, we calculated the alpha blocking rate (α_blockingrate) to reflect a subject's switch from eye-close (EC) to eye-open (EO) states as an estimate of degree of wakefulness as well as stability of subjects during experiment. We used the rate of change of the alpha blocking rate (α_blockingrate) effect of the EC and EO resting states before and after the CR task to monitor the stability of the experimental recording process and verify the validity of EEG data after preprocessing (Zheng et al., 2018). We set 20% as the stability threshold of the rate of change of the alpha blocking rate based on our long-term observations in experimental study. We observed that once subject's α_blockingrate decreased its value above 20% change, it was very likely that the subject became sleepy in long-term experiment. On the contrary, if it increased its value above 20% change, it was very likely the intrinsic brain behavior state had changed, which introduced some unexpected noise to the experiment, and affected the interpretability of the data. After the CR task, the α_blockingrate values of the 2 groups were both reduced by <20% in comparison to those before the CR task (Table 1, Supplementary Figure 1, and Supplementary Method 1).

EEG network connectivity such as functional connectivity generally refers to the statistical relationship of EEG signals between electrodes (or brain areas) (Fingelkurts et al., 2005). To avoid the effect of volume conduction and the field diffusion on multiple-recording channels, we used the phase lag index (PLI) based on phase synchronization to evaluate the brain EEG functional connectivity (Stam et al., 2007). The PLI value was calculated with the open source toolbox HERMES based on MATLAB (Niso et al., 2013). The range of PLI values is generally between 0 and 1, where a value of 1 means that the 2 EEG signals have strict phase locking at a constant non-zero phase lag and a value of 0 means no coupling (or coupling with a relative phase that encircles 0 modπ, which is likely to result from volume conduction) (Zheng et al., 2018). Thus, the larger the PLI value indicates the stronger the non-zero phase synchronization and the stronger the connectivity (Stam et al., 2007). We first applied an approach called network-based statistics (NBS) (Zalesky et al., 2010) to analysis brain functional connectivity based on PLI and to find the significant connectivity edge (SCE). In terms of the energy related to the wiring cost. The weaker or less extensive the connectivity is, the less active synaptic connections there are. The less active synaptic connections will cost less energy. That is, network wiring cost (C_w) is proportional to the wiring number of connections (W_NC) (Achard and Bullmore, 2007; Zheng et al., 2021). There were up to 28 edges among the 8 electrodes. Some PLI values were very low which were around noise level. There should be an optimal baseline and threshold to reduce the noise interference. Hence, we set different thresholds e.g., from 1/10, 1/9, 1/8, 1/7, 1/6, 1/5, 1/4, 1/3, 1/2–1 maximum of PLI value of 32 participants (Max) in order to compare those relatively stronger functional connections for both ES and CG in the CR task state. There was a significance between 2 groups for some threshold (1/4 Max, 1/8 Max, and 1/9 Max) (Figure 2C and Supplementary Figure 10). Here, as an example case, 14 Max was used as the threshold for distinguishing the talented swimmer group from the control. Absolute threshold was better than the relative threshold of each subject for the comparability between ES and CG. Therefore, we set 1/4 of the maximum PLI value of 32 participants in 2 groups as the threshold in each frequency band. If the PLI of two channels was greater than the threshold value, one W_NC was calculated. The W_NC of all participants were calculated and analyzed during the CR task.

Brain signaling activity involves not only a wiring cost for network connectivity but also an activity cost for neuronal discharge. Studies have shown that a higher mean frequency (MF) of EEG reflected higher levels of cerebral blood flow and metabolism (Ingvar, 1971; Hyder et al., 2013), and MF was confirmed to have a positive correlation with these physiological variables (Ingvar et al., 1976; Zheng et al., 2021). That is, the high MF suggests higher frequency of brain electronic activity that will cost more metabolic energy (Hyder et al., 2013; Yu et al., 2018). Since the MF can indirectly reflect the energy cost of neural electrical activity in the brain, we used the MF of the CR task state as the index for the energy cost of brain activity (C_activity). MF will be higher or lower when the condition is changed. In the present study, we calculated and compared the MF of ES and CG groups at the same condition.

Calculation of MF: Each subject was selected 50 s signals of EC state, 50 s signals of EO state, and 70 s signals of CR task (selection from the beginning recording signals of the CR task) to be preprocessed. The mean frequency of 25 segments or 35 segments EEG data from 8 channels were calculated in 2 s. It is known that electromyogram (EMG) artifacts have a higher amplitude than the EEG signals and can be removed by using independent component analysis (ICA) technique (Chen et al., 2014; Frolich and Dowding, 2018). But in our present study, there were only 8 electrodes and they were not enough for using ICA to detect the EMG artifacts. According to the characteristics of EMG, such as distributing relatively higher frequency with a higher amplitude, we used the ratio in the formula (1) for the evaluation of the muscular content. The electromyography (EMG) artifacts were removed according to the ratio of high-frequency bands power over low-frequency bands power due to the relatively higher frequency of myoelectric. The calculation formula of the ratio was as follows:

Where i represents an integer frequency ranging from 1 to 30 Hz, and P(fi) means the power value at a certain integer frequency. When the value of Ratio of one segment from one channel of a subject was >1, this segment was removed because it belongs very likely to the EMG artifacts (Supplementary Figure 14 and Supplementary Method 4). According to the above methods, about 2.81‰ segments of EC, 8.59‰ segments of EO, and 7.03‰ segments of the CR task were removed before calculating the mean frequency of each channel of every participant in 3 conditions (EO, EC, and the CR task).

Averaged power spectra were computed across segments of different states in each participant. The power values were calculated for 5 frequency bands (δ: 1–4 Hz; θ: 4–8 Hz; α: 8–13 Hz; β_lower: 13–20 Hz; β_upper: 20–30 Hz). It has been well-established that low-frequency signaling activities, such as δ or θ, are related to sleep or the resting state of the brain, while high-frequency signaling activities, such as α and β, are related to the cognitive function (Kumar and Bhuvaneswari, 2012). The power ratio of the high-frequency band to the low-frequency band can reflect the degree of brain activation (Cheron et al., 2016). Therefore, we calculated the power ratio of the upper β frequency band to the θ frequency band of each channel (upper beta/theta ratio, or UBTR), to represent the activation rate of eight brain areas (Arns et al., 2013; Vollebregt et al., 2015). The formula of UBTR was as follows:

Note: In our study, there were no electrodes in the temporal or occipital areas of the brain. When we drew the spectrum power and mean frequency topologic maps, we added an additional 18 electrodes and set the values to 0 to avoid the influence of the frontal and parietal signals on the periphery. The added electrodes were Fp1, Fp2, F5, F6, C5, C6, P5, P6, Po3, Po4, F7, F8, O1, O2, T7, T8, Po1, and Po2.

The electrical potential of eye blinks is required along with the brain rhythm signals by the electrodes in EEG and shows a higher intensity in the frontal electrodes and possesses higher amplitude than the brain rhythms (Sovierzoski et al., 2008; Al-gawwam and Benaissa, 2018). In present study, blink recognition was performed based on the EEG data of each participant after high- and low-pass filtering. The processes of algorithm identification were: (1) to find all the peaks that may be blinks in the Fpz channel using the function findpeak in Matlab (seen Supplementary Method 3); (2) to select 250 ms time series before and after each peak and calculate the amplitude of the peak to the trough on the left and right sides, respectively, and then to compare the averaged left and right amplitudes with the peak threshold. It would be kept when the averaged amplitude was larger than the peak threshold; (3) to meet the condition that there were at least in 3 other channels the amplitudes of the trough were greater than the trough threshold at the same period; (4) to remove the local maximum peak with the amplitude difference between 2 consecutive peaks less than one third of the maximum peak in the Fpz channel. The peak threshold was set to 2/5 of the maximum peak amplitude of the subject and the trough threshold was set to one third of the lowest trough amplitude of the subject in the Fpz channel during the complex reaction task. One blink was recognized after the above 4 conditions were all met.

According to the above-mentioned blink recognition algorithm, all the blinks of each subject were identified during the complex reaction task. The peaks of blinks were aligned to 0 ms. The mean amplitude from 250 to 150 ms before the blink peak was set as the baseline amplitude. Task-evoked blink potentials of each subjects were plotted after normalized by the Z-score method.

Blink rate or the frequency at which the eyelids open and close has been proposed and used to study cognitive control, learning, working memory, and decision making (Eckstein et al., 2017). In the manuscript, the instantaneous blink rate of each subject was calculated over time from the beginning to 60 ms of the complex reaction task, and was normalized with the average blink frequency of his/her group in 60 s. In the calculation process, subjects who blinked <3 times in the first 60 s should be removed.

Previous study noticed that the interblink intervals were quite variable between subjects (Ponder and Kennedy, 1928). According to the percentages of different blink intervals to the total number of blink intervals of each participant, the interblink interval histogram (IBIH) of each person in the process of CR tasks was calculated for his/her total CR tasks continuous time sequence s selected from each subject in the 2 groups. The total number of blink intervals of all subjects in each group was calculated with the time bin of 1 s. Most interblink interval durations of these subjects were distributed in <20 s while a very few interval durations (>20 s) distributed sparsely with maximal value reaching 91 s.

Brain functional connectivity based on PLI was analyzed mainly by an approach called network-based statistics using the NBS v1.2 toolbox, based on MATLAB (Zalesky et al., 2010). After 5,000 permutation tests, if there was a significant difference (p < 0.05) between ES and CG groups, it was marked with a line as a SCE in the topology diagram (Figures 2H,I). Other index data were analyzed by SPSS 20.0. An independent t-test was used between the 2 groups in one state such as the speed and accuracy of the CR task, WNC at different frequency bands or different threshold, the lateralization index in the resting or task state, mean frequency in the resting or task state, spectrum power at different frequency bands etc. A paired t-test was used between 2 states within the same group, e.g., mean frequency of ES between EC and EO state or between EO and the CR task state; the significance threshold was set at p < 0.05. Pearson correlation analysis was used for individual EEG and CR task performance.

The averaged reaction time and accuracy values of the CR task were shown in Table 1. The ES responded significantly faster (Figure 2A) and more accurate (Figure 2B) than the CG. This result is consistent with reports from other sports (Mori et al., 2002; Williams et al., 2002; Kida et al., 2005).

Alpha activity is greatly reduced by the increase in light input from the resting EC state to the EO state or blocked by other attention-related signals. The alpha-blocking phenomenon is conceptualized as desynchronized neural population activity during active stimuli, and alpha blocking rate (α_{blockingrate)} between the EC and EO states is used to monitor the stability of the experimental recording process (Method seen Supplementary Method 1). It is generally believed that a data recording process is relatively stable if the α_blockingrate is <20% before and after the tasks (Bazanova and Vernon, 2014). In the current study, the mean α_blockingrate of the elite swimmers after the CR tasks was 6.4% lower than the rate before the tasks. In the control group, the α_blockingrate after the CR task was reduced about 15% (Table 1 and Supplementary Figure 1). This relatively higher and stable α_blockingrate values in elite athletes might reflect that they could maintain sustained attention for relatively longer period than controls.

Based on network-based statistical analysis, there were 2 strengthened functional connectivity edges in the θ frequency band (4–8 Hz) in the ES compared with the CG during the CR task (Figure 2G). These edges were distributed in the left frontal area, connecting the left frontal area (F3) to the central parietal area (Cz) and to the right side of the parietal area (C4). In the upper β frequency band (20–30 Hz), compared to the college student controls, master swimmers had 6 less correlated functional connectivity edges (Figure 2H), which were mainly distributed in the frontal, temporal and parietal regions of the brain. The connection strengths of all recording sites for other frequency bands were not significantly different between the 2 groups.

In the upper β frequency band, the frontoparietal network W_NC of ES was significantly lower than that of CG (p < 0.05; Figure 2D) and was positively correlated with swimming training years within the ES group (Figure 2G). As shown in the Supplementary Figure 2, there were positive linear relations between training years and CR accuracy or speed. However, in the other 4 frequency bands, W_NC was not different between ES and CG (Supplementary Figure 3). Within the group either ES or CG, W_NC was also positively correlated with CR speed (Figure 2E) or accuracy (Figure 2F). However, the 2 intercepts of linear relations of ES were both greater than controls.

A gradual decrease was shown in the mean frequency of the frontoparietal area from the EC state to the EO state and then to the CR task (Figures 3A,D,E and Supplementary Figure 4). This result was consistent with our group's previous findings on the mean frequency of the normal population in the frontoparietal area (Supplementary Figure 5). However, during the CR task, there was no difference in the mean frequencies between ES and CG (p > 0.05; Figure 3B), nor was the change rate of the mean frequency from the EO state to the CR task (p > 0.05; Figure 3C).

Although the absolute power values of ES in the 5 frequency bands were not significantly different from those of CG during the CR task (p > 0.05), there was an upward trend in the ES in the frontoparietal region of the left hemisphere while it was absent in the CG (Supplementary Figure 6).

Comparing to the EO state, ES showed increased activities in the prefrontal region at each frequency band during the CR task state, while CG showed more spatially distributed across multiple regions (Figures 4A–E and Supplementary Figures 6, 7). In terms of the absolute value of the power change, at the δ frequency band (1–4 Hz), ES had significantly smaller power changes in the left frontal area (F3) than CG (Figure 4A); in the θ frequency band (4–8 Hz), there was a significant difference in the center of the parietal area (Cz), and the power change of the ES was significantly smaller than that of the CG (Figure 4B); at the lower β frequency band (13–20 Hz), ES had a significantly smaller change in the right frontal area (F4) (Figure 4D); and in the upper β frequency band (20–30 Hz), an increased power change was found in the left parietal lobe (C3) of ES (Figure 4E).

From the EO state to the CR task state, the UBTR of ES decreased significantly less than that of CG in the frontal Fz channel (Figure 5). There were 5 channels (F3, F4, Cz, C3, and C4) in which the UBTR changes of the 2 groups interacted. That is, there was an increasing trend in the UBTR change of the ES, while there was a decreasing trend in that of the CG. However, only in the Cz channel was there a significant difference (Figure 5).

Supplementary Table 2 and Supplementary Figure 8A showed the frontal lateralization index (LI_f) (Method seen Supplementary Method 2). There were no significant differences between ES and CG either in the resting state (EC and EO) or in the CR task state (p < 0.05). However, the parietal lateralization index (LI_p) of ES was significantly different from that of CG in the CR task state (p < 0.05). During the CR task, the LIp value of ES was positive and close to zero, while that of CG was negative and far from 0. This result suggested that top swimming athletes had more balanced left and right cerebral hemisphere activities and more strengthened activation in the right hemisphere than college students (Table 1 and Supplementary Figure 8B).

The EEG-based blink amplitude and blink frequency of ES were both lower than those of CG, although the absolute values of the 2 groups were not significantly different during the CR task (Figures 6A–C). The interblink interval histogram (IBIH) also revealed that the blink frequency of ES in the short interval group was lower than that of CG (Figure 6E). From the beginning of the tasks to the following 60 s, the task-related blink rate of ES showed a regular periodic concentrated blink pattern with an interval of ~15–16 s. However, in the CG, there were continuous blinks in short periods but no obvious concentration pattern. Meanwhile, the blink frequency of CG was almost higher at each time point than that of ES during the first 1 min (Figure 6D). Interestingly, there were some burst blinks (more than or equal to 2 blinks in 1 min) in the 2 groups during the CR task. However, ES had more burst blinks than CG in both the burst blink rate (Figure 6F) and the proportion of subjects (Figure 6I).

To test whether the performance of the CR task was affected by the blinks, we computed Pearson correlation coefficients to evaluate the relationships between the complex reaction accuracy or speed and blink frequency. As shown in Figure 6G, there was a negative correlation between log₁₀ blink frequency and complex reaction accuracy in both groups. However, a smaller slope was displayed in the ES than in the CG, which indicated that the task performance of master swimmers was less influenced by the blinks. Consistent with the reaction accuracy, the complex reaction speed also exhibited a negative linear relationship with log blink frequency but was less affected than the reaction accuracy in terms of the slope difference (Figure 6H).