The procedures used in the experiment were according the Declaration of Helsinki, and all subjects gave written informed consent. The procedures were approved by the local ethics committee (Committee on- Research Involving Human Subjects, Region Arnhem-Nijmegen, The Netherlands).

Twenty-two healthy participants took part in the experiment. All participants had normal or corrected-to-normal vision and were right-handed. None of the subjects had any known neurological or psychiatric disorder. Participants gave written consent prior to the start of the experiment and were paid in accordance with the guidelines of the local ethics committee. Five participants were excluded from further analysis due to excessive eye movements to eccentric locations in more than a fourth of the trials during the experiment, which we interpreted as violation of the concept of covert attention to the target locations. Two other participants were excluded because of chance level performance in the discrimination task in the easy condition and one participant was excluded because of excessive head movements, which added artefacts to the MEG data. Therefore, a total of 14 subjects (mean age 24.8, range 19 37 years; 8 female) was taken into account for analysis. Three of these subjects had experience with covert visual spatial attention from a previous experiment.

A whole-head MEG system with 275 axial gradiometers (CTF MEG Systems, Port Coquitlam, Canada) was used to record ongoing brain activity at a sampling rate of 1200 Hz. The subject's head location relative to the MEG sensors was measured with marker coils placed at the nasion, and in the left and right ear canals. Sensors ‘MLT37’ and ‘MLF62’ could not be recorded for the first eleven subjects and sensors ‘MRF66’ and ‘MRO52’ could not be recorded for the last five subjects because of sensor dropout.

An EyeLink 1000 eye-tracker (SR Research Ltd., Kanata, Canada) pointing at the subject's right eye was used to measure eye movements with a sampling rate of 2000 Hz. At the beginning of each experiment subjects performed a calibration block to determine the gaze direction when overtly fixating. In this calibration procedure filled circles with a diameter of 1.04° were randomly shown for one second on the horizontal and vertical axes of the fixation cross at eccentricities of 3.5°, 7°, and 9°. The average calibration error was 0.2°.

Subjects were seated upright in the MEG system with their eyes 85 cm from a screen on which the stimuli were projected. The projection screen had a width of 45 cm. Stimuli were projected on the screen with an EIKI LC-XL100 projection system with a resolution of 1024×768 pixels, and a frame rate of 60 Hz. Stimuli were presented using Psychtoolbox (https://psychtoolbox.org). Gamma correction (luminance calibration) was applied to increase the brightness and contrast of the display.

Subjects were asked to perform a covert visual spatial attention task in which they had to direct their attention to one out of four different locations on the screen (near-left, near-right, far-left, far-right; see Figure 1A), while maintaining gaze at a fixation cross at the centre of the screen. At the attended location they had to perform a two-alternative forced choice Gabor patch orientation discrimination task at one of 2 difficulty levels; ‘easy’ and ‘difficult’. In the easy condition the Gabor patch had a full contrast (1∶1000) such that subjects performed the discrimination task approximately 100% correct, while in the difficult condition the patch contrast was set for each subject individually such that each subject performed the discrimination task approximately 75% correct. The contrast of the patch, defined as , was 0.048, 0.370, 0.125, 0.250, 0.083, 0.104, 0.119, 0.175, 0.106, 0.120, 0.083, 0.094, 0.125, 0.087 for our 14 subjects, respectively.

Stimuli were carefully chosen to maximize effects of eccentricity and task difficulty. Subjects attended to four spatial locations on a grey background at either 3.5° (near) or 7° (far) eccentricity at the lower left or lower right of the fixation cross (illustrated in Figure 1A). To help subjects maintain their attention at one of these locations black circles were used as placeholders. All four placeholders were shown during the entire experiment to minimize stimulus related effects. Attention was directed to the lower visual field because it has been shown that the resolution of spatial attention is better in the lower visual field than in the upper visual field [23].

The eccentricities of 3.5° and 7° and the sizes of the placeholders (1.04°, and 1.40° for the near and far location, respectively) were chosen for several reasons. Firstly, subjects should attend to visual targets outside their fovea as is standard in the covert visual attention paradigm. Secondly, the distance between the placeholders should be large enough to activate as much as possible disjunct regions in the visual cortex. Thirdly, the area within the placeholders should be such that each stimulus activates approximately equally-sized populations of neurons in the visual cortex involved in processing. The relation between the cortical magnification factor M and the retinal eccentricity E was provided by Duncan and Boyton [16] by the formula . With eccentricities of 3.5° and 7° we ensured a distance of about 1 cm between the centres of gravity of the two populations of neurons in the primary visual cortex related to the near and far stimulus [16]. Furthermore, we corrected the diameters of the placeholders by the cortical magnification factor to ensure that the neuronal populations in visual cortex that are involved in visual processing were equally large, leading to diameters of 1.04° for near placeholders and 1.40° for far placeholders.

As targets, black-and-white Gabor patches were used that were constructed using four cycle sinusoidal waves with a 2-D Gaussian mask of 52% of the placeholder's width. Thus, near and far targets had the same number of waves. The sinusoid was oriented either at −45° (left) or +45° (right) as shown in Figure 1B.

Data were analysed using FieldTrip, an open-source analysis Matlab toolbox for analysing electrophysiological data [25]. For classification of left versus right spatial attention data the Donders Machine Learning toolbox was used (https://github.com/distrep/DMLT). Next, SPSS was used for statistical analyses. Only trials with a minimum pre-target period of two seconds were used for analyses, such that all trials used contained a period of 1.5 seconds of covert attention without possible cue-related artefacts.

Figure 2 shows the behavioural accuracy (Panel 2A) and response times (Panel 2B) in the orientation discrimination task. In line with our intentions, participants acquired on average an accuracy of 96% (near) and 95% (far) in the easy condition and an accuracy of 75% (near) and 76% (far) in the difficult condition. A two-way repeated measures ANOVA demonstrated a significant difference between easy and difficult conditions in accuracy (F(1,14) = 55.763, p<0.001). No effect of eccentricity was found (F(1,14) = 0.189, p = 0.671), indicating that we properly corrected for the size of the stimuli at the near and far eccentricities by taking the magnification factor into account. There was also no significant interaction effect between task difficulty and eccentricity (F(1,14) = 0.012, p = 0.915). As the accuracy distribution was significantly non-normal in the easy near condition (D(14) = 0.307, p<0.01), we also performed a non-parametric Wilcoxon signed-rank test on the behavioural accuracy between both levels of task difficulty. This test also showed a significant effect of task difficulty (p<0.001). Furthermore, as assessed using a binomial confidence interval for each subject the accuracy in the difficult condition deviated significantly from the accuracy in the easy condition (p<0.05).

For response times, a two-way repeated measures analysis also demonstrated a significant main effect of difficulty (F(1,14) = 88.805, p<0.001). On average the response time was 0.571 s in the easy condition and 0.763 s in the difficult condition. No effect of eccentricity was found (F(1,14) = 1.256, p = 0.283) and neither was an effect of interaction (F(1,14) = 0.015, p = 0.903).

Furthermore, a significant linear correlation was found between a subject's response time and accuracy (F(1,14) = 41.702, p<0.001, partial eta-squared = 0.762), indicating that the response time increased linearly with the accuracy of the response.

For the MEG analyses, we determined for each hemisphere a ROI in which the alpha modulation, as defined by Equation (1), deviated significantly from zero over all conditions and subjects (see Figure 3). As expected, the sensors selected by this method are in the parietal and occipital cortex. The two hemispheric ROIs were used for the alpha power and alpha time-power analyses.

Our results showed that a more difficult task led to an increase of the strength of alpha lateralization. This result obtained by MEG studies in humans is in line with a previous study using single-unit recordings in monkey visual cortex (area V1) on task difficulty, which also showed that task difficulty had an effect on neuronal processing. The latter study showed that an increased task difficulty enhanced neuronal firing rate at the focus of attention and suppressed it in the surrounding regions [19].

A contra-indication is a covert attention study in humans by Cosmelli et al. [31] in which no effect of task difficulty was found. In that study task difficulty was modulated to rule out modulations of alpha band activity due to non-specific changes such as alertness. Subjects were shown two coloured squares at the attended location and they had to press a button if one of these squares matched the colour they were instructed to detect. This resulted in a difference in behavioural performance for the different colours to be detected, as certain colours are easier to distinguish than others. However no differences in alpha power over the contralateral and ipsilateral parietal sites were found between the easy and difficult condition. In the study of Cosmelli et al. 50% of the targets were shown at the cued side, while in our study 100% of the targets were shown at the cued side. Previous research has shown that spatial certainty leads to graded changes in the extent to which alpha oscillations are lateralized [32]. A potentially weaker lateralization in the Cosmelli et al. study may mean that any additional modulation due to task difficultly is lost in the relatively stronger background noise.

Our results show that in both the easy and the difficult condition subjects show an alpha lateralization shortly after cue onset (see Figure 5) and the alpha lateralization increases more after cue onset for the difficult condition than for the easy condition. These results indicate that manipulating task difficulty has an effect, but do not address the actual cause of the change in alpha lateralization. For example, the effect could be due to a change in the visual task leading to differences in attentional demand or perceptual load or the participants could have changed their strategy during the difficult task by increasing their mental effort. Future work could alternatively investigate these causes in more detail. For instance, to independently manipulate mental effort a similar experiment could be conducted in which the maximum response time is manipulated, instead of the contrast of the target stimulus.

We discovered the largest changes in alpha power compared to baseline and effects of task difficulty in the left hemisphere while attending to the right hemifield and we assume that these changes also have affected the strength of the alpha lateralization. Our finding that the left hemisphere during right hemifield attention makes a large impact on the alpha lateralization is in agreement with previous research which showed that the left hemisphere mainly supports attention shifts to the right hemifield, while the right hemisphere is involved in shifts to both hemifields [33]. Quite remarkably, as we would expect the changes in alpha power to be symmetric with respect to attended hemifield we only found significant alpha power differences compared to baseline during right hemifield attention. Recently, this bias to the right visual hemifield has also been found in ADHD patients [34].

A possible caveat in our experimental design was that attention to the cued location in the easy condition might not have been necessary as the targets in the easy condition were presented at maximal contrast. Subjects could have detected the orientation of the contrast due to a pop-up effect of the target. However, we found a significant alpha lateralization pattern in the easy condition indicating that our participants did pay covert attention to the visual stimuli in the easy condition. This lateralization pattern became significantly different from zero shortly after cue onset, as was also the case in the difficult condition.

Previous covert visual attention studies reported correlations between pre-target alpha modulation and behavioural measures as summarized in [8]. We performed similar correlation analyses as used in [35], [36], but did not find a significant correlation between the alpha modulation and either the behavioural accuracy or response times during the difficult condition. Furthermore, we have analysed the microsaccade rate between the easy and difficult condition using the algorithm proposed by [37]. Microsaccade rate decreased before target onset in both conditions indicating focused attention [38]. Additionally, the microsaccade rate was slightly, but significantly higher in the difficult condition than in the easy condition in the period after cue onset (0–1.8 seconds) but this effect disappeared near target onset.

Contrary to our expectations, we did not find an effect of eccentricity with our alpha lateralization analyses. Bahramisharif et al. [14] did find different alpha lateralization patterns for different eccentricities. We believe that the most plausible explanation for the difference with the results in our study is that we corrected for potential confounds of cortical magnification [16] and task difficulty, which were not taken into account by Bahramisharif et al. [14].

One could argue that we did not observe an effect of eccentricity, because we averaged over the ROIs such that any changes in location within the ROIs may have been obscured. However, a post-hoc within-subject cluster randomization test [28] to identify clusters of sensors in which the alpha lateralization varied between near and far conditions out of all sensors also did not reveal any significant clusters.

Based on the stronger alpha modulation in the difficult versus the easy condition, we expected to see a higher classification rate in the difficult conditions compared to the easy conditions. This higher classification rate was indeed found. This effect was found when using the alpha power of sensors from the ROIs as features. We chose to report these ROI sensor results, as these features gave on average the highest classification performance. However, we also observed the task difficulty effect when using the alpha power from all sensors as separate features for classification, indicating that the selection of sensors did not alter the observed effects. Furthermore, Figure 7 seemed to indicate that the classification rate in the easy far condition was more similar to the difficult condition than to the easy near condition, although there was no significant interaction effect. When using all the sensors as separate features for classification, the easy far condition was more similar to the easy near condition with classification rates of respectively 60.4% and 59.4%, while the difficult far and difficult near conditions had classification rates of respectively 64.7% and 65.3%. This result, using a different subset of sensors, indicates as well that the higher classification rate in the easy far condition in Figure 7 is not a robust effect. The positive correlation of 0.611 between the classification rate and the alpha lateralization showed that the classification analysis was for a large part using similar information as the lateralization analysis. While this was an MEG study and therefore not directly applicable for BCI, we expect the results to transfer to EEG since the visual covert attention paradigm seems to give similar brain signals in EEG [13]. Although the recorded brain signals will be more spread over the scalp due to increased spatial smearing in EEG, we expect the effect of difficulty to remain as it is an effect in the magnitude of the alpha power. Average classification rates in Treder's et al. [13] EEG study were higher (74.6%) than in our MEG study (63.5%), however in their study performance was improved by selecting the subject-specific pair of directions which maximised classification performance.

Surprisingly, we did not find an effect of target eccentricity on the alpha lateralization during the covert visual spatial attention task. Task difficulty did show an effect on the alpha lateralization. A more difficult task increased the alpha lateralization and the BCI classification rate. This implies that it is important to consider the influence of task difficulty in alpha oscillation studies and from a BCI perspective it seems important to have a difficult task in a covert visual spatial attention paradigm in order to maximize performance.