Modulation of alpha oscillations by attention is predicted by hemispheric asymmetry of subcortical regions

Evidence suggests that subcortical structures play a role in high-level cognitive functions such as the allocation of spatial attention. While there is abundant evidence in humans for posterior alpha band oscillations being modulated by spatial attention, little is known about how subcortical regions contribute to these oscillatory modulations, particularly under varying conditions of cognitive challenge. In this study, we combined MEG and structural MRI data to investigate the role of subcortical structures in controlling the allocation of attentional resources by employing a cued spatial attention paradigm with varying levels of perceptual load. We asked whether hemispheric lateralization of volumetric measures of the thalamus and basal ganglia predicted the hemispheric modulation of alpha-band power. Lateral asymmetry of the globus pallidus, caudate nucleus, and thalamus predicted attention-related modulations of posterior alpha oscillations. When the perceptual load was applied to the target and the distractor was salient caudate nucleus asymmetry predicted alpha-band modulations. Globus pallidus was predictive of alpha-band modulations when either the target had a high load, or the distractor was salient, but not both. Finally, the asymmetry of the thalamus predicted alpha band modulation when neither component of the task was perceptually demanding. In addition to delivering new insight into the subcortical circuity controlling alpha oscillations with spatial attention, our finding might also have clinical applications. We provide a framework that could be followed for detecting how structural changes in subcortical regions that are associated with neurological disorders can be reflected in the modulation of oscillatory brain activity.


Introduction
The visual world provides more sensory information than we can be aware of at any given moment. Thus, our brains must prioritise goal-relevant over distracting information. A rich body of research shows that the brain amplifies goal-relevant inputs, and suppresses non-relevant inputs by a process referred to as selective attention (1)(2)(3). There is ample evidence for topdown control of neocortical regions associated with sensory processing when information is prioritized (1,4,5). The dorsal attention network, which consists of the intraparietal sulcus/superior parietal lobule, and the frontal eye fields, is the most predominant network associated with the allocation of attention (6)(7)(8). However, although the role of neocortex for spatial attention and cognitive control has been extensively studied, the contributions of subcortical regions are less well understood. One reason, amongst many others, is that MEG and EEG are not well suited for detecting subcortical activity. Therefore, the present study aims to provide insights into the contribution of the thalamus and basal ganglia in driving top-down spatial attention.
There has been intense focus on the cortical contributions to the top-down control processes, yet there are multiple sources of evidence to suggest that subcortical structures also play an important role in cognitive control. For instance, it has been shown that the pulvinar plays an important role in the modulation of neocortical alpha oscillations associated with the allocation of attention (9). The basal ganglia have been demonstrated to be involved in various types of cognitive control, including attention (10,11) and conscious perception (12). Studies in rats and non-human primates have shown that both the thalamus and superior colliculus, are involved in the control of spatial attention by contributing to the regulation of neocortical activity ; Krauzlis et al., 2013Krauzlis et al., , 2018. Notably, when the largest nucleus of the thalamus, the pulvinar, was inactivated after muscimol infusion, the monkey's ability to detect colour changes in attended stimuli was lowered. This behavioral deficit occurred when the target was in the receptive field of V4 neurons that were connected to lesioned pulvinar (18). The basal ganglia are also involved in visuospatial attention through their connections to cortical areas such as the prefrontal cortex via thalamus. Anatomical tracing studies on selective attention and distractor suppression point to a key role of prefrontal-basal ganglia-thalamus pathway whereby sensory thalamic activity is regulated by prefrontal cortex via basal ganglia (11). Furthermore, fMRI studies in humans demonstrated increased activation in basal ganglia when covert attention was reallocated. Additionally, dynamic causal modelling has shown that the basal ganglia can modulate the top-down influence of the prefrontal cortex on the visual cortex in a task-dependent manner (19).
In terms of neuronal dynamics, power modulation of oscillatory activity in the alpha band  has been proposed to reflect resource allocation between goal-relevant and irrelevant stimuli. This has consistently been shown between studies in EEG and MEG in which attention is allocated to the left or right hemifield. Such studies typically find an alpha power decrease in the hemisphere contralateral to the attended stimuli complemented by a relative increase in alpha power in the other hemisphere associated with unattended stimuli (Okazaki et al., 2014;Thut et al., 2006;Worden et al., 2000). It is debated whether the alpha power associated with the unattended stimuli is under task-driven top-down control or rather explained by an indirect control mechanism driven by the engagement of the target (24). The latter notion is aligned with perception load theory that is defined as the perceptual demand of the task or relevant stimulus, according to which the (finite) resources are allocated (25). Indeed, a recent study demonstrated when the target stimulus has a higher perceptual load (e.g., more difficult to perceive), alpha band power increases in ipsilateral regions thus indirectly reflecting distractor suppression (26).
Based on these findings, both oscillatory activity in the alpha band and the activity of subcortical structures are involved in the allocation of attentional resources. The direct relationship between activity in subcortical regions and neocortical oscillations is poorly understood in humans, in part owing to the difficulty in detecting the activity of deep structures using MEG/EEG. One way around this is to instead investigate, the relationship between the volumetric measures of subcortical structures and oscillatory brain activity by combining MRI and electrophysiological measures such as MEG. Using this approach, it was shown that the hemispheric lateralized modulation of alpha oscillations is correlated with the volumetric hemispheric asymmetry of both the globus pallidus and the thalamus (27). The relationship between the globus pallidus and the modulation of alpha oscillations was demonstrated in the trials where the visual stimuli were associated with high-value (positive or negative) reward valence.
In this study, we aimed to identify a link between the volumetric asymmetries of subcortical structures and the modulation of alpha oscillations in the context of spatial attention without explicit reward-associations. Given the assumed contribution of the basal ganglia to rewardbased learning, it is perhaps unsurprising to find contributions of the globus pallidus in the paradigms targeting reward valence. What remains to be determined is whether these structures play a more general role in the formation of spatial attention biases. We analysed MEG and structural data from a previous study (26), in which spatial cues guided participants to covertly attend to one stimulus (target) and ignore the other (distractor). Importantly, the target load and the visual saliency of the distractor were manipulated using a noise mask. This load/salience manipulation resulted in four conditions that affect the attentional demands of target and distractor This approach allowed us to relate the hemispheric volumetric asymmetries in thalamus, caudate nucleus, and globus pallidus to the modulation of alpha oscillations when spatial attention is allocated under varying conditions of cognitive challenge.

Results
We investigated the relationship between the volumetric lateralization of subcortical structures estimated from structural MRIs and the hemispheric modulation of alpha oscillations measured by MEG in a spatially cued change detection task. We asked the participants to covertly attend to face-stimuli in the left or right visual field and indicate the direction of a subtle gaze-shift of the attended face ( Figure 1A). The influences of perceptual load and distractor salience were examined by combining noisy and clear target and distractor stimuli in a 2 x 2 design ( Figure   1B).

Modulation of alpha power with respect to left and right cues
To quantify the anticipatory change in alpha power, we analysed the modulation of power in the -850 -0 ms interval prior to the target. As expected from a previous report (26), we observed a power decrease contralateral to the cued hemifield and a relative increase ipsilaterally (i.e. an increase contralateral to the distractor, Figure 2A)  magnitude of the modulation index (MI( )) reflecting the relative difference in alpha power when attending left versus right, gradually decreased and increased over respectively the left and right hemisphere until target onset ( Figure 2B). We then identified symmetric clusters of sensors (5 over each hemisphere) that showed the highest modulation of alpha power ( Figure   2C) and focused the subsequent analyses on these sensors of interest. We then calculated the hemispheric lateralized volumes of the seven subcortical structures, as illustrated in Figure 3B (thalamus, caudate nucleus, putamen, globus pallidus, hippocampus, amygdala, and nucleus accumbens) using the FIRST algorithm on the MRI data. Thalamus (mean ± std = -0.0123 ± 0.0121, p-value < 0.000), putamen (mean ± std = -0.0149 ± 0.0285, p-value = 0.004) and nucleus accumbens (mean ± std = -0.1141 ± 0.0746, p-value < 0.000) have significantly negative LV values (i.e., left lateralization) whereas the caudate nucleus is right lateralized (mean ± std = 0.0115 ± 0.0285, p-value = 0.021) ( Figure 3B). Globus pallidus, hippocampus, and amygdala did not show any robust volume lateralization.

Relationship between subcortical regions and hemispheric alpha lateralization
To test whether the individual hemispheric asymmetries in subcortical grey matter relate to variability in HLM( ), we subjected the MEG and MRI data to a General Linear Model (GLM). In  It is worth noting that neither the behavioural nor the rapid frequency tagging measures showed significant relationships with LVs and HLM().

Association between volumetric lateralization of subcortical regions and attention related to perceptual load conditions.
To relate load and salience conditions of the task to the relationship between subcortical structures and the alpha activity, we combined low-load or high-load targets with high-saliency or low-saliency distractors to manipulate the perceptual load appointed to each trial (Method section).
We therefore applied a multivariate multiple regression

Discussion
In the current study, we sought to identify the association between the volumetric hemispheric While some MEG studies have demonstrated that it is possible to detect activity from deep structures such as the hippocampus (28)(29)(30)(31), it is questionable whether one in general can use MEG to reliably detect activity from the thalamus and basal ganglia, owing to low SNR from sources close to the centre of the head (Baillet, 2017). Given these constraints, we instead correlated MEG data with structural magnetic resonance images to uncover functional contributions of subcortical structures to spatial attention.
We evaluated the relationship between subcortical structures and cortical oscillatory activity relying on the association between structure and function. Previous research points to a link between the volume of a given brain regions and its functionality. For instance, it is well established that shrinkage (atrophy) in specific regions is a predictor of a number of neu olo i al an sy hiat i on itions in lu in Pa kinson's isease, dementia, and Huntin ton's isease. In Pa kinson's isease, at o hy in the nu leus accumbens and thalamus correlated with cognitive impairments (34). In a large-scale study on 773 participants, patients ith Alzheime 's isease have been sho n to have a si nifi antly smalle amy ala, thalamus, caudate nucleus, putamen, and nucleus accumbens than matched controls (35). Patients with symptomatic Huntin ton's isease also sho significantly smaller caudate nucleus than presym tomati a ti i ants ho e e a ie s of Huntin ton's isease ene mutation (36).
Based on these considerations, we argue that the volume of basal ganglia relates to the ability to modulate posterior brain oscillations in attention type tasks. We demonstrated this by

Thalamus, Globus Pallidus, and Caudate nucleus play different roles in various load conditions.
Our results demonstrate a shift in the contribution of the thalamus, globus pallidus, and caudate nucleus when increasing the perceptual load of the target and saliency of the distractor. While in the low load, low saliency condition, the lateralized volume of the thalamus was correlated with the interhemispheric bias in alpha modulation, in the low load, high saliency, as well as high load, low saliency conditions, globus pallidus was related to the alpha oscillatory activity. Finally, the caudate nucleus was mainly associated with the high load, high saliency condition.
This differing pattern of the thalamic and basal ganglia structures might be suggestive of their respective contributions to the control of attentional resources. Involvement of the thalamus when the task is in its simplest form can be explained by its role relaying information between the basal ganglia and the prefrontal cortex (42,43). applications in terms of early diagnosis. Our approach could also be extended to other tasks resulting in hemispheric lateralization of oscillatory brain activity, e.g. working memory tasks (52) or language tasks (53). Our results also call for more direct investigations of the relationship between subcortical regions and neocortical oscillations which is best done by intracranial recordings in non-human primates or utilizing human recording from deep-brain stimulation electrodes combined with EEG or MEG.

Methods and Materials Participants
We analysed a previously collected dataset, described in (26). 35 right-handed healthy volunteers (25 female, mean age: 24 ± 5.7) participated. All reported normal or corrected-tonormal vision. One participant did not give consent for their data to be used outside of the original study and one was removed due to poor MRI [segmentation] quality, resulting in 33 participants in total. All subjects signed an informed consent form before participation and were paid £15 per hour. The study was conducted in compliance with the Declaration of Helsinki and was approved by the Science, Technology, Engineering, and Mathematics (STEM) ethical review committee of the University of Birmingham.

Experimental design
Participants were instructed to perform a cued change detection task (2 blocks of 256 trials, 45 minutes; Fig 1A),

Visual stimuli
Stimuli e e i ula fa es that om ise 8˚ visual an le in iamete an la e ith 7˚ eccentricity from fixation and were presented in the lower hemifield. Over trials, the perceptual load of targets was manipulated using a noise mask; masked targets are harder to detect and therefore incur greater cognitive load in their detection. The saliency of distractor stimuli was also manipulated using a noise mask; masked distractor stimuli are less salient and therefore less disruptive to performance on the detection task. The noise mask was created by randomly swapping 50% of the stimulus pixels ( Figure 1B). This manipulation resulted in four targetload/distractor-saliency conditions: (1) target: low load, distractor: low saliency (i.e., clear target, noisy distractor), (2) target: high load, distractor: low saliency (i.e., noisy target, noisy distractor), (3) target: low load, distractor: high saliency (i.e., clear target, clear distractor), (4) target: high load, distractor: high saliency (i.e., noisy target, clear distractor) (Figure 1B and C).
The stimulus set consisted of eight different face identities that were randomized across trials.
On each trial, the identities of both stimuli were the same; however, to avoid visual differences between left and right the faces were mirror symmetric from the fixation point. Stimuli were projected using a VPixx PROPixx projector (VPixx technologies, Saint-Bruno, Canada) in Quad RGB mode (1440Hz) with an effective resolution of 960x540 pixels. Face stimuli were tagged with an invisible rapid-frequency-tagged flicker (for more details please refer to Gutteling et al., 2022). The distance between the participant and the projection screen was 148cm resulting in a .6˚ of visual an le s een. Regions outside of the masks were excluded from subcortical alignment (56).

Structural data acquisition
To assess hemispheric laterality for each SGM nucleus, we calculated the Lateralization Volume indices (LVs):

Equation 1
Where ℎ and represent the anatomical volume of a given subcortical structure (s) in number of voxels, in the right and left hemisphere, respectively. This equation implicitly controls for individual differences in brain volumes and has been commonly used to compute hemispheric structural asymmetries (Mazzetti et al., 2019). LVs can range between -1 and 1 where a positive LV indicates rightward asymmetry and vice versa.

MEG data acquisition
Electromagnetic data were recorded from participants while seated in upright position, using a

MEG data analysis
MEG data analysis was performed using custom scripts in MATLAB 2017a and 2019b (The MathWorks) and the FieldTrip toolbox (59). The analysis pipeline was adapted from the FLUX pipeline (60) and the scripts are available at https://github.com/tghafari/AMI_Substructures.

Preprocessing
Raw MEG data were high-pass filtered at 1Hz and demeaned. Then data were segmented in 4s epochs (-3s to 1s) relative to the target-onset (gaze shift of the face stimuli). Secondly, trials with sensors artifacts (e.g., jumps) were removed manually to prepare the data for automatic artifa t attenuation usin in e en ent om onent analysis (I A; "runica.m" in iel i ).
Components related to eye blinks/movements, heartbeat and muscle activity were rejected.
Thirdly, by visually inspecting the trials, we removed those containing clear residual artifacts such as eye blinks. We also removed trials with saccadic deviations larger than 3˚ from fixation (using EyeLink eye tracker data) during the 1.5s interval before target-onset (-1.5 -0 s) (average ± SD = 13.7% ± 8.0 trials). Sensors that were removed during preprocessing were interpolated using a weighted neighbour estimate.

Time-frequency analysis of power
To calculate the time frequency representations (TFR) of power, we used a 3-cycle fixed timewindow (e.g., 300ms for 10Hz) at each 10ms step. The data segments were multiplied by a Hanning taper to control the frequency smoothing and reduce spectral leakage. For computational efficiency, we also used a zero-padding, rounding up the length of segments to the next power of 2. Then a fast Fourier transform (FFT) was applied to the tapered segments in the 2-30 Hz frequency range in 1Hz steps and the power was estimated. The power was then summed for each gradiometer pair.
To quantify the anticipatory oscillatory activity, we focussed on the -850 to 0 ms interval before target onset. To select sensors constituting the region of interest (ROI), we calculated the 8-13 Hz alpha modulation index (MI(a)) for all sensors. TFR of power for each sensor was averaged over all trials in the -850 to 0 ms interval, for attention to right and left. Then the MI() for each participant and each sensor was calculated as:

Equation 2
Where Power( )k denotes the alpha power at sensor k in each condition. To evaluate hemisphere-specific lateralization of alpha band modulation, we applied the hemispheric lateralization modulation (HLM()) index:

Equation 3
Where = 5 represents the number of sensors in each ROI and ( ) ℎ or ( ) denote the modulation index for sensor k over the right or left hemisphere, respectively.

Generalized Linear Model
To Here, LVTh, LVCN and LVGP refer to the lateralization volume of thalamus, caudate nucleus, and globus pallidus, respectively.

Multivariate multiple regression
To simultaneously model the predictive relationship between the lateralized volume of thalamus, caudate, and globus pallidus, and all four load conditions, we used a multivariate multiple regression (MMR) (61) analysis. MMR is used to predict multiple dependent variables using multiple systematic parameters. It allows for modifying our hypothesis tests and confidence intervals for explanatory parameters and responses, respectively (62). The model was defined as: ( 1 ) + ( 2 ) + ( 3 ) + ( 4 ) ~ 0 + 1 ℎ + 2 + 3 +

Equation 5
Where HLM( ) refers to hemispheric lateralization modulation of alpha power in load conditions 1 to 4 ( Figure 1C), respectively; refers to the coefficients in the model; LVTh, LVCN and LVGP refer to the lateralization volume of thalamus, caudate nucleus, and globus pallidus, respectively.
To ensure our chosen MMR predicts meaningful variance in HLM(a) scores, we compared a full model containing LV indexes from all 7 subcortical regions to one where the key structures of interest (i.e., thalamus, caudate nucleus, and globus pallidus) had been removed, leaving putamen, nucleus accumbens, hippocampus, and amygdala as regressors. This model is referred to as the reduced model. We also compared a model containing the key regressors of interest (LVTh, LVCN, LVGP) to a null model that contained only subject intercepts as regressors.

Behavioral data analysis
To evaluate if the participants response times and accuracy was correlated with the hemispheric lateralization of alpha oscillatory activity as well as lateralized volume of subcortical structures, we calculated behavioral asymmetry (BA) as below:

Equation 6
Where / ℎ and / correspond to the behavioural asymmetric performance in accuracy or response times when the attention was toward right or left visual hemifield, respectively. We then calculated the Pearson correlation between the BA and HLM.
Finally, we ran the winning GLM model with accuracy and response times as the dependent variable and LVTh, LVCN, and LVGP as the regressors.