Aperiodic activity masks sensorimotor mu suppression during human action observation

Abstract

In electroencephalography (EEG), action execution (AE) is reliably associated with reductions in the mu rhythm, a periodic oscillation (8–14 Hz) over sensorimotor cortex. Similar patterns of “mu suppression” have been reported during observations of others’ actions, leading to claims that the mu rhythm indexes motor contributions to social perception. However, evidence for mu suppression during action observation (AO) is mixed, possibly due to methodological considerations (e.g. sample size, recording sites) and perceptual and attentional confounds. Moreover, measurements of periodic oscillations may be conflated with underlying aperiodic (“1/f-like”) neural activity, potentially influencing estimation of the EEG power spectrum. Here we examined the influence of aperiodic factors on mu suppression using 128-channel EEG in a large sample (N = 109), both during AE and in an AO task with appropriate visual and attentional controls. Whereas AE was consistently associated with significant mu suppression, we initially failed to find significant mu suppression for AO, suggesting that attentional and perceptual confounds may bias mu estimation during AO. However, removing the aperiodic component restored mu suppression over central electrodes. Although significant, mu suppression for AO appears less robust than originally reported, with aperiodic activity contributing to variability in the estimation of mu suppression during AO.

Introduction

Since its introduction a century ago, electroencephalography (EEG) has enabled researchers to identify neural oscillations associated with specific cognitive processes. One such periodic oscillation, the mu rhythm, is visible over central electrode sites at frequencies between 8 and 14 Hz. During tasks that require action execution (AE), such as finger tapping, this neural oscillation is robustly reported to show decreases in power, or magnitude, consistent with the idea that the mu rhythm originates from sensorimotor cortex (Hari and Salmelin 1997; Halgren et al. 2019). These decreases in mu-band power during AE are commonly thought to reflect desynchronization of large-scale resting EEG oscillations due to the selective activity of motor circuits, as evidenced by negative correlations between “mu suppression” during bimanual finger tapping and increased sensorimotor activation in functional magnetic resonance imaging (fMRI) (Ritter et al. 2009). Strikingly, a similar pattern of mu suppression has also been reported in action observation (AO) tasks, when stationary participants view videos of others’ actions (Hari et al. 1998; Babiloni et al. 2002). This has led some researchers to suggest that the same neural circuits are activated both when performing and viewing actions (Lago-Rodriguez et al. 2013), in line with the larger conceptual framework of embodied cognition. Models such as action simulation theory posit that our understanding of others’ mental states derives from internal simulation of their movements using our own sensorimotor systems (Springer et al. 2013, Wood et al. 2016). Consistent with this idea, reductions in mu suppression during AO have been reported in clinical populations with deficits in social perception, such as autism spectrum disorder (Oberman et al. 2005, Oberman and Ramachandran 2007, Dumas et al. 2014).

However, the reliability and specificity of mu suppression effects during AO have recently come under debate (Fox et al. 2016, Hobson and Bishop 2016, Bowman et al. 2017, Hobson and Bishop 2017). Some critiques of this literature arise from basic methodological considerations, such as the field’s reliance on relatively small participant samples, which increase the likelihood of false positives (Hobson and Bishop 2017). Other concerns have focused on experimental design and analysis of EEG data in AO paradigms. For example, Hobson and Bishop (2016) compared three different methods of baselining that have been used in previous studies of mu suppression (single long rest period, short between-trial rest periods, and within-trial static images). They found that the choice of baseline condition had a significant impact on the observed variability and specificity of the mu rhythm, raising questions about the extent to which measured mu suppression truly reflects a selective response to biological motion. Compounding this issue, few studies have attempted to control for attentional factors during AO, leading to potential confounds of the motoric mu rhythm with attentional suppression of the overlapping alpha rhythm (8–14 Hz). Although the cortical generators of the alpha rhythm lie posterior to the sensorimotor cortex (Halgren et al. 2019), activity from these sources can easily be conflated with the mu rhythm, particularly when only a small number of sensors are compared (Hobson and Bishop 2017).

In a recent study, we addressed several of these concerns about mu suppression through a combination of careful methodological controls and technical approaches (Siqi-Liu et al. 2018). Using a high-density 128-channel EEG array, we measured detailed scalp topographies across the entire head, allowing us to better differentiate alpha and mu rhythms. Additionally, we recorded from a more robust sample of approximately 40 participants, in line with recent recommendations for statistical power (Hobson and Bishop 2017). Last, we controlled for perceptual and attentional factors that might confound estimation of sensorimotor mu. Specifically, we used whole-body point-light displays (PLDs), which apply markers on the major joints of the body to capture patterns of motion associated with meaningful human actions (e.g. Atkinson et al. 2004). These “coherent,” biologically plausible PLDs were compared to “scrambled” versions of the same stimuli in which the starting locations of the dots were randomly displaced, controlling for low-level visual motion. To ensure that participants attended to both stimulus types, we also required them to complete an attentionally demanding continuous monitoring (one-back) task. With these controls in place, we successfully replicated previous reports of selective mu suppression for coherent versus scrambled PLDs (Ulloa and Pineda 2007).

However, since the publication of these results, there has been growing awareness in the field that estimates of oscillatory brain activity can be affected by another factor, namely aperiodic activity (Donoghue et al. 2020, Gyurkovics et al. 2021, Brake et al. 2024). The EEG signal can be divided into two major components: (i) oscillatory, or periodic, activity and (ii) non-oscillatory, aperiodic activity. Whereas periodic signals are thought to arise from the temporally synchronized firing of neural populations, the aperiodic component of the EEG signal instead reflects desynchronized neural activity. Also known as “1/f-like” activity for its characteristic pattern of decreasing power at higher frequencies, aperiodic activity can be quantified both in terms of the overall level of measured activity (i.e. intercept or offset) and the rate at which power falls off at higher frequencies (i.e. slope). Aperiodic activity has been reported to vary as a function of arousal states, including sleep versus waking (Favaro et al. 2023) and anaesthesia (Waschke et al. 2021, Maschke et al. 2023); developmental trajectory (e.g. younger versus older adults: Kałamała et al. 2024); and psychological and neurological disorders (Pani et al. 2022, Akbarian et al. 2024). Recent studies further suggest that aperiodic activity varies as a function of task demand (Gyurkovics et al. 2022, Lu et al. 2024).

Yet for researchers interested in measuring oscillations associated with cognitive processing—such as the mu rhythm—the presence of arrhythmic aperiodic activity poses a dilemma. Because aperiodic activity is predominant in the EEG power spectrum (Jacob et al. 2021), estimation of task-related oscillations can potentially be distorted by the aperiodic component. In order to isolate oscillatory brain activity of interest, spectral parameterization tools have been developed to model and remove the aperiodic component from EEG data (e.g. Donoghue et al. 2020). However, these approaches have not yet been applied to large-scale datasets to quantify the effects of removing aperiodic activity on the estimation of the mu rhythm.

A final question is how the removal of aperiodic activity affects estimation of the mu rhythm as a function of motor execution versus viewing (AE versus AO). If both executing and observing actions rely on the same neural circuits (Prinz 1997), then we would expect similar scalp topographies and response patterns for AE and AO conditions. Conversely, comparison of mu-band activity during AE and AO within the same participants could also highlight how processing differs between these conditions, particularly at electrodes beyond the commonly tested central sensor sites. Consistent with this point, the larger neuroimaging literature on AO has identified overlapping but distinct activation patterns for AE and AO (e.g. Hardwick et al. 2018). Yet, relatively few studies have compared the scalp distribution and magnitude of mu suppression elicited by AO to the pattern of motoric mu suppression during AE (Fox et al. 2016), and none, to our knowledge, have done so while controlling for differences in aperiodic activity. This concern is particularly relevant when task-related EEG is quantified relative to resting activity, such as via baseline division and log transformation (Gyurkovics et al. 2021), as is done in many studies of mu suppression.

Here we examine this question, comparing mu suppression for AO and AE in a large sample with appropriate perceptual and attentional controls. We replicate the paradigm used in our previous study (Siqi-Liu et al. 2018), with the addition of an AE task to provide a baseline measure of motoric mu suppression for comparison to AO (Fig. 1a). We collected high-density EEG data from over 100 people while they engaged in a bimanual finger tapping task and observed PLDs of whole-body movements. Finally, we compared estimations of mu suppression using a standard Fourier transform with and without aperiodic activity removed.

Figure 1.

Experimental methods. (a) Overview of the experimental paradigm. High-density 128-channel EEG was collected while participants completed two tasks: (i) AE, in which the participant alternated between resting and tapping the index and middle fingers of both hands against the respective thumbs; and (ii) AO, in which the participant viewed PLD animations of whole-body human movements while continuously monitoring for immediate stimulus repetitions (one-back task). (b) Timeline of the experiment. Participants in the Original paradigm completed 10 blocks of AE followed by two blocks of AO, then another 10 blocks of AE. Participants in the AO-first paradigm completed the AO task first, followed by all 20 blocks of AE. (c) Sample trials from the AO task, which included one block each of Coherent biologically plausible PLDs and Scrambled biologically implausible PLDs in which the starting locations of the dots were randomly displaced.

Methods

Participants

Participants (N = 139, ages 18–30 years) were recruited from the local college community for either monetary compensation or partial course credit. In the final analyses, 30 participants were excluded for the following reasons: technical issues with the EEG recording (n = 10), software/piloting issues (n = 10), excessive eyeblink or motor artefact (n = 5), or behavioural d-prime (d′) of 0 in one or more conditions (n = 5). Of the remaining participants, 55 (40 female; 51 right-handed) completed a version of the procedure (“Original”) in which the two blocks of AO were preceded and followed by 10 blocks of AE, for a total of 20 blocks of AE (Fig. 1b). The other 54 participants (20 female, 1 nonbinary; 52 right-handed) underwent a modified protocol (“AO first”) in which the AO task was presented prior to all 20 AE blocks (Fig. 1b). All experimental procedures were reviewed and approved by the Institutional Review Board of Claremont McKenna College. Informed consent was obtained in writing from all participants prior to the start of the experiment.

Stimuli

For the AO task, we used PLD videos of whole-body movements as described in Siqi-Liu et al. (2018). PLDs are an ideal stimulus to study biological motion perception, as they discard other visual cues such as form information while leaving motion cues intact. PLDs for this study were selected from a validated stimulus set developed by Atkinson et al. (2004) and consisted of 3-s video clips of various whole-body actions performed by trained male and female actors. The stimuli included bodily expressions of three emotions (happiness, anger, and sadness), as well as three affectively neutral but meaningful actions (marching, hopping, and toe touches). The effect of emotional content on mu suppression will be explored in a future study.

Another advantage of PLDs is that they can be used to construct “scrambled” stimuli in which the starting positions of the dots are randomized within the bounds of the original viewing frame, thus preserving low-level motion cues while disrupting perception of meaningful, biologically plausible actions (Atkinson et al. 2012). There were six different performances of the 3 emotions (18 videos total), as well as 2 different performances of the neutral actions (6 videos total), producing 24 coherent PLDs. The same PLDs were used to create the scrambled stimuli, resulting in 48 distinct PLDs in total.

Procedure

Participants’ brain activity was recorded with high-density 128-channel EEG while they engaged in bimanual finger tapping (AE) and AO tasks designed to elicit mu suppression (Fig. 1a). As shown in Fig. 1b, each participant experienced the same basic procedure, starting with approximately 20–30 min of set-up time in which informed consent and demographic forms were completed and the EEG cap was applied. Following the initial set-up, the participant underwent approximately 45–60 min of EEG testing, including AE and AO tasks, finishing with the collection of various surveys and questionnaires for the final 20–30 min. Thus, the entire experiment took approximately 1.5–2 h to complete.

In the original version of the experimental protocol, the two AO blocks were preceded and followed by 10-block epochs of AE (20 blocks total), with the goal of maintaining participant engagement across the experiment. Data from 55 participants collected under this protocol were included in the final analysis. However, one concern was that this protocol may have inadvertently created action-induced perceptual modulation (Schütz-Bosbach and Prinz 2007), since the prior blocks of motor execution of the AE task could potentially affect perceptual processing of the PLDs in the AO task. To address this concern, we collected data from an additional 54 participants in a modified version of the protocol. In this “AO-first” version of the procedure, participants completed the full AO task prior to the 20 AE blocks, reducing potential priming of motor representations. Subsequent statistical analyses showed no significant differences in the spectral frequency of mu suppression between the two versions of the AO protocol, enabling us to pool both datasets for a larger sample size. For all tasks, stimuli were presented in MATLAB using the Psychophysics Toolbox (Brainard 1997, Kleiner et al. 2007).

The AE task consisted of alternating periods of rest and bimanual finger tapping, or “Action.” This is a standard paradigm used to elicit the canonical motoric mu rhythm. Each condition was associated with a specific visual cue on the computer screen, either a red square (Rest) or a green circle (Action). During the Rest cue, participants were instructed to remain as still as possible, whereas during the Action cue, they were asked to tap their thumbs against their middle and index fingers at their own pace. Although individual variation in the frequency of tapping was not quantified in our study, previous research suggests that the maximal tapping rate supported by human musculoskeletal dynamics is about 6 Hz (e.g. Shourijeh et al. 2017), with self-paced tapping likely to be lower, making it unlikely to interfere with estimation of the power spectrum for the mu rhythm. In order to remove potential visual confounds from viewing their movements, participants were instructed to place their hands in their laps under a table prior to the start of the AE task, as well as maintaining visual fixation on the computer screen. Each block contained 16 s of Rest and 16 s of Action, for a total of 32 s per block. Participants either completed 20 blocks presented either in two separate runs of 10 blocks (“Original”) or as one run of 20 blocks (“AO first”).

The AO task employed the same parameters as reported in previous studies (Siqi-Liu et al. 2018, Strang et al. 2022). Participants viewed 3-s PLD videos of whole-body movements including both emotional and affectively neutral actions. At the end of each clip, the last frame would be displayed during a 2-s intertrial interval (ITI). To maintain participants’ engagement across the task, they performed an attentionally demanding continuous “one-back” task, in which they monitored the stream of videos for an immediate repetition of the same stimulus. Participants were explicitly instructed to respond only if the specific dot configuration and motion associated with a video appeared twice in a row (i.e. the same video with the same actor was repeated), and not for general repetition of the same action (e.g. two different actors doing the same action). Immediate repetitions of the same video were indicated by button press during the 2-s ITI. To ensure that the instructions were clear, all participants received 10 practice trials (5 coherent, 5 scrambled, 1 one-back trial each) before the start of experimental trials.

The AO task consisted of two separate blocks of Coherent and Scrambled PLDs, corresponding respectively to meaningful, biologically plausible movements and randomized, biologically meaningless motion as a control (Fig. 1c). The order of the two blocks was counterbalanced across participants. Each block consisted of 108 trials, with three presentations of the 18 emotional actions and nine presentations of the six neutral PLDs. In each block, 12 trials were randomly selected to be followed by one-back repetitions, resulting in 240 total trials (216 stimulus presentations plus 24 one-back trials). The order of stimulus presentation in each block was pseudo-randomly interleaved to ensure the correct number and placement of “one-back” trials, such that no stimulus would be presented more than two times in a row. Participants with poor behavioural performance (d′ ≤ 0) in either block were excluded from further analysis.

EEG data acquisition and analysis

EEG data were collected using the BioSemi ActiveTwo EEG System (BioSemi B.V., Amsterdam, The Netherlands) with 128 active electrodes inserted in fitted head caps with additional bilateral electrodes on the mastoids for reference. EEG signals were digitized continuously at a sampling rate of 512 Hz with a hardware low-pass at one-fifth of the sampling rate. All offsets were adjusted to fall between −30 and 30 prior to data collection, with no obvious slow drifts in the online data measurement.

Data preprocessing was performed offline in MATLAB (The Mathworks, Inc., Natick, MA, USA) using the EEGLAB toolbox (Delorme and Makeig 2004). Following import into MATLAB, we applied a standardized preprocessing routine implemented in the PREP pipeline plugin (Bigdely-Shamlo et al. 2015), including removal of 60 Hz electrical noise, robust re-referencing to the average signal, and detection and interpolation of bad channels. Following the PREP pipeline, we applied a 0.5 Hz high-pass filter to remove slow voltage drifts from the data.

Data were divided into epochs differently depending on the task. For the AE task, the data were epoched using an 8-s window (−9 to −1 pre-action cue for Rest, 1 to 9 s for Action). The use of the latter half of the Rest time window ensured that this condition would not be contaminated by brain activity related to the onset or initial visual processing of the Rest cue. For the AO task, data were epoched using a window from −1200 to 3200 ms post-stimulus onset. Data epochs in the AO task were baseline-corrected to the pre-stimulus period (−1200 to 0 ms). Additionally, to ensure that measured mu rhythms during AO did not reflect potential motor preparatory activity, all one-back trials and all trials where participants made a motor response were removed from data analyses.

After epoching, the data were visually inspected for motor artifacts, and excessively noisy epochs were removed. Remaining artifacts (e.g. eyeblinks, noisy electrodes) were cleaned from the data using independent components analysis (ICA) as implemented by second-order blind identification (Belouchrani et al. 1997, Tang et al. 2005). Identified ICA components of interest were cross-checked with automated classifications from the MARA (Winkler et al. 2011) plugin for EEGLAB, with task-related components then back-projected to the scalp (Jung et al. 2000).

Spectral decomposition of the EEG signal was performed in MATLAB using the FieldTrip toolbox for frequency analysis (Oostenveld et al. 2011). Power spectra were computed for frequencies from 1 to 30 Hz using the fast Fourier transform (FFT) with multi-taper method, followed by either log₁₀ transformation to normalize the frequency distribution or spectral parameterization via “fitting-­oscillations-and-one-over-f” (FOOOF; Donoghue et al. 2020), as implemented in FieldTrip. For the spectral parameterization, the aperiodic component was modelled and removed separately for each participant (10^{log10fooof–log10fooof_aperiodic}). We also compared estimation of the aperiodic-corrected mu rhythm using an additive model (10^{log10fooof—}10^{log10fooof_aperiodic}), as detailed by Gyurkovics et al. (2021). Power spectra were calculated separately for each participant and condition across the entire time window of interest. For AE, this consisted of 8-s epochs of Action and Rest. For AO, we used a 1-s window from 800 to 1800 ms after the onset of the PLD stimulus in Coherent and Scrambled conditions in order to exclude spectral activity related to stimulus onset and offset, as well as any potential motor response preparation.

Based on previous research defining the alpha-band mu rhythm as between 8 and 14 Hz (Hobson and Bishop 2017), we first performed a data-driven statistical analysis at a frequency-by-frequency level in the AE task, using dependent sample two-tailed t-tests corrected for multiple comparisons using a nonparametric cluster-based Monte Carlo permutation test (10 000 repetitions) with significance defined by an overall threshold of P = .05 (P = .025 at each tail). For the FFT data, significance was determined by comparing log₁₀-transformed values of Action versus Rest. For the FOOOF analysis, the same approach was followed but with the aperiodic component modelled and removed. To plot power spectra associated with each condition, sensors of interest (SOIs) were determined by taking the conjunction of sensors identified as significant across the entire frequency range of the significant negative cluster in each analysis described above (FFT: 8–14 Hz, FOOOF: 10–14 Hz).

We then used the frequency ranges from the AE analysis to inform the analysis of AO, averaging across the frequency range identified by each analysis. Since no significant negative clusters were found within this frequency range for the FFT approach, we used the SOIs from the AE task to plot power spectra associated with Coherent versus Scrambled conditions in this analysis. We also conducted an exploratory, data-driven analysis looking frequency-by-frequency across all frequencies (1–30 Hz) and the frequency range of interest identified from the AE analysis. Since these results did not differ with respect to mu suppression effects, the results from the 8–14 Hz frequency range are reported below.

In addition to calculating the periodic oscillations minus modelled aperiodic activity, we also inspected the slope and offset parameters for the aperiodic component itself, as obtained from the spectral parameterization procedure. Aperiodic parameters were extracted separately for AE and AO conditions. Finally, based on our results, we estimated a measure of statistical effect size (Cohen’s d) using built-in functions within the FieldTrip software package. We then entered this estimate into the G*Power software (Faul et al. 2007) to obtain the a priori required sample size to detect a significant effect at α = 0.05 and power = 0.8 for one- and two-tailed versions of the dependent-samples t-test.

Results

To understand the effects of the aperiodic component on the measured power of the mu rhythm, we examined EEG activity during AE and AO using a standard FFT analysis (aperiodic component included) versus following the removal of aperiodic activity via FOOOF. Because we had run two different versions of the AO protocol, with AE preceding AO versus AO first, we first tested whether there were any significant differences in EEG activity between the two datasets during the AO task. We compared frequency data for Coherent versus Scrambled conditions, both using FFT and FOOOF in our predefined time window of interest (0.8–1.8 s after stimulus onset). We found no significant differences between the original and AO-first datasets, either when using a frequency-by-frequency approach across all frequencies (1–30 Hz) or within specific frequency bands previously associated with mu suppression (8–14 Hz). Therefore, we combined these two datasets for further analyses.

Establishing canonical mu suppression with AE

To identify canonical mu suppression associated with motor execution, we first compared EEG activity in the AE task during periods of Action (bimanual finger tapping) versus Rest. For this analysis, we used a standard fast Fourier transform (FFT) analysis without removing aperiodic activity (Fig. 2). Statistical analysis using cluster-corrected non-parametric analysis revealed significant reductions in spectral power over central electrodes, as well as at frontal electrode sites (Fig. 2a, white dots), successfully replicating the standard finding that motor execution is associated with a measurable suppression of spectral power between 8 and 14 Hz. Notably, this mu suppression appears to be readily separable from alpha-band activity, which was visible predominantly over occipitoparietal sensors in a slightly lower range from 8 to 12 Hz (Fig. 2A, black dots). To verify that these results reflected a decrease in mu-band activity over sensorimotor cortex, we further plotted the average difference in power between Action and Rest from 8 to 14 Hz (Fig. 2b). Sensors of interest (SOIs) were defined using the conjunction of sensors showing significant effects across the entire 8–14 Hz frequency range (Fig. 2b, red dots). For these SOIs, we then plotted the average power spectrum at SOIs for both the Action and Rest conditions (Fig. 2c). As expected, this revealed a 1/f-like shape of the overall power spectrum along with a noticeable peak in the 8–14 Hz range corresponding to the mu oscillation. This peak was smaller for Action relative to Rest, matching previous reports of mu suppression.

Figure 2.

Estimation of mu suppression in AE task for spectral data using standard Fourier analysis (FFT). (a) Statistical maps showing cluster permutation-­corrected significance for Action—Rest in the a priori mu frequency band of 8–14 Hz. For this analysis, dots indicate sensors at which cluster permutation-corrected P < .05 for positive (black) and negative (white) direction of effect. (b) Scalp topography of power for Action—Rest averaged over significant frequency window of negative effects (8–14 Hz). For this analysis, dots indicate significant SOI surviving a conjunction analysis across the 8–14 Hz frequency range. (c) Average power spectrum across SOIs for Action versus Rest.

Our next question was whether this canonical mu suppression effect would be preserved when the aperiodic component of the EEG signal was removed. As shown in Fig. 3, removal of aperiodic activity via FOOOF produced a similar pattern of significant negative activity at central and frontal electrode sites, though the effects were restricted to a smaller window from 10 to 14 Hz (Fig. 3a). Likewise, comparison of power for Action versus Rest in the 10–14 Hz window showed maximal suppression at central electrode sites (Fig. 3b), overlapping with the conjunction SOIs across the 10–14 Hz window (Fig. 3b, red dots). Plotting the average power spectrum across all frequencies for these central SOIs identified a peak from 10 to 14 Hz, even with aperiodic activity removed (Fig. 3c). Similar results were found using an additive model in which aperiodic and periodic components were assumed to be independent, as suggested by other research examining resting-state EEG (Gyurkovics et al. 2021).

Figure 3.

Estimation of mu suppression in AE task for spectral data using estimation and removal of aperiodic activity via spectral parameterization (FOOOF). (a) Statistical maps showing cluster permutation-corrected significance for Action—Rest in the a priori mu frequency band of 8–14 Hz. For this analysis, dots indicate sensors at which cluster permutation-corrected P < .05 for positive (black) and negative (white) direction of effect. (b) Scalp topography of power for Action—Rest averaged over significant frequency window of negative effects (10–14 Hz). For this analysis, dots indicate significant SOI surviving a conjunction analysis across the 10–14 Hz frequency range. (c) Average power spectrum across SOIs for Action versus Rest.

Collectively, our results are in line with a large body of evidence indicating that central mu rhythms arise from sensorimotor circuits (e.g. Halgren et al. 2019). In addition, motoric mu suppression was associated with activity over frontal sites, beyond the central or “Rolandic” sensor locations typically associated with the mu rhythm (e.g. Hari 2006). Although this frontal distribution likely reflects non-motoric differences due to task demands, as well as other factors such as volume conduction, it is worth noting that the amplitude of the mu rhythm is correlated with activity across multiple brain networks (Yin et al. 2016). More importantly, however, these results did not differ substantially as a function of whether the analysis included the aperiodic component, suggesting that aperiodic activity has relatively limited effects on the estimation of the motoric mu rhythm.

Comparing mu suppression for observation of coherent versus scrambled PLDs

What about AO? As a first pass, we looked at activity within the frequency window identified from the AE analysis for Coherent PLDs versus Scrambled versions of the same stimuli. Because the low-level motion information in both types of stimuli is the same, this comparison should provide a more stringent measurement of biological motion processing distinct from other visual motion cues. In addition, participants were tasked with monitoring for infrequent stimulus repetitions of both types of stimuli, ensuring similar levels of attention to both conditions. Although in theory the subtraction of these two conditions should be sufficient to find AO mu suppression, we also corrected for potential baseline differences in spectral power using subtraction of the pre-stimulus log-transformed signal, as commonly implemented in previous studies (e.g. Hobson and Bishop 2016). Comparing the analyses with and without baseline correction, we found no major differences in the scalp topography or statistical significance of our mu suppression results.

However, using the standard FFT analysis in which aperiodic activity is retained, we found no strong reduction in power for Coherent—Scrambled conditions (Fig. 4a) and no statistically significant negative effect in this range (Fig. 2b). To plot the average AO power spectrum, we used the conjunction SOI from the AE task (Fig. 4b, red dots), showing a much noisier and smaller difference between Coherent and Scrambled in the 8–14 Hz range (Fig. 4c). This result suggests that when sample sizes are large enough to account for false positives and visual and attentional factors are appropriately controlled, mu suppression may be less robust than previously thought.

Figure 4.

Estimation of mu suppression in the AO task for Coherent—Scrambled human movements from (a–c) standard Fourier analysis (FFT) and (d–f) following removal of the aperiodic component (FOOOF). In all cases, the frequency range of interest was determined based on the window of significant effects in the corresponding AE analysis. (a) Spectral power and (b) statistical maps computed using FFT, 8–14 Hz. Since no significant negative effects were observed in this analysis, SOIs from the AE task (dots) were used to plot the average power spectrum in (c). (c) Baseline-corrected average power spectrum across SOIs during observation of Coherent versus Scrambled PLDs. (d) Spectral power and (e) statistical maps computed using FOOOF, 10–14 Hz. Dots in (e) indicate sensors at which cluster permutation-corrected P < .05 for negative (white) direction of effect. (f) Aperiodic-corrected average power spectrum across SOIs during observation of Coherent versus Scrambled PLDs.

Further examination of the same data following removal of the aperiodic activity showed a striking difference. The scalp topography shows a clear negative effect, which appears to largely reflect the same central sources seen during the finger-tapping task (Fig. 4d), with a sizable number of sensors now reaching significance (Fig. 4e). Although the distribution of significant activity is slightly different from that for the AE task, with more activity over midline electrodes, the overall pattern is highly consistent with engagement of motor circuits during viewing of others’ actions. Likewise, the average power spectrum now displays a clear and prominent peak in the range associated with the mu rhythm, similar to the results for AE, with reduced power for Coherent PLDs (Fig. 4f). Confirming these results, an exploratory frequency-by-­frequency analysis revealed no significant mu suppression in the standard FFT analysis (Fig. 5a), but significant effects in the 11–12 Hz range with aperiodic activity removed via spectral parameterization (Fig. 5b).

Figure 5.

Exploratory frequency-by-frequency analysis of AO Coherent—Scrambled analysis for (a) standard (FFT) analysis, 8–14 Hz and (b) aperiodic-corrected (FOOOF) analysis, 10–14 Hz. Frequency ranges were derived from the corresponding AE results for each analysis. Dots indicate sensors at which cluster permutation-corrected P < .05 for positive (black) and negative (white) direction of effect.

Why might removing aperiodic activity have such a dramatic effect on our ability to recover AO mu suppression effects? To explore this question, we obtained aperiodic parameters for each participant at sensors associated with mu suppression, defined as showing a significantly greater response to Action versus Rest in the AE task (Fig. 3b, red dots). Specifically, we obtained two parameters: offset, conceptualized in terms of the overall levels of measured activity; and slope, the rate at which power falls off at higher frequencies. We first computed these parameters for the AE task, where we would expect to see large differences in offset and slope given the many differences between the finger-tapping condition and resting baseline (task demands, arousal, etc.). As seen in Fig. 6a and b, Action was associated with highly significant increases in offset (average offset ± SD: Action = −0.79 ± 0.25, Rest = −0.82 ± 0.26; t(108) = 5.35, p = 4.89 × 10⁻⁷) and slope (average slope ± SD: Action = 1.14 ± 0.12, Rest = 1.08 ± 0.13; t(108) = 10.5, p = 4.19 × 10⁻¹⁸). These results suggest that aperiodic parameters change systematically as a function of differing task demands, consistent with prior reports (Waschke et al. 2021, Gyurkovics et al. 2022).

Figure 6.

Aperiodic parameters by condition for (a–b) AE and (c–d) AO tasks. Aperiodic parameters (a and c) offset and (b and d) slope were computed separately for each participant and condition, averaged across sensors showing a significant effect in the AE task (Action—Rest). Points indicate individual participants, jittered based on the distribution of the data. Notches indicate median values; non-overlapping notches signify that medians differ at the 5% significance level. * P < .05.

For the AO task, we would expect little to no difference in offset or slope of the aperiodic components (Fig. 6c and d), as the two conditions are more closely matched aside from the PLD stimuli themselves. Consistent with this idea, the difference in the ­aperiodic slope term was not significant between conditions (­Coherent = 1.121 ± 0.13, Scrambled = 1.122 ± 0.13; t(108) = −0.23, P = .82). However, analysis of the aperiodic parameters revealed a small but significant increase of the offset in the Coherent condition relative to Scrambled (Coherent = 0.09 ± 0.21, Scrambled = 0.08 ± 0.21; t(108) = 2.45, P = .02). These results indicate that, at least for this dataset, there is a significant increase in the broadband (aperiodic) offset for the Coherent condition (Coherent > Scrambled) in the opposite direction of the periodic mu suppression effect (Coherent < Scrambled). Thus, when aperiodic activity was included in the data, it appears that our estimation of the periodic mu rhythm in the AO task was inflated for the Coherent condition relative to Scrambled, effectively masking sensorimotor mu suppression.

Although our findings following removal of the aperiodic component are consistent with previous claims that AO elicits mu suppression, the current results suggest that this effect is much smaller and less robust than previously theorized. To assess this idea, we computed estimates of Cohen’s d for AE and AO using standard statistical tools in the FieldTrip software package. For this analysis, we focused on the data in which the aperiodic component had been removed, averaging across the significant SOIs for AE in the 10–14 Hz frequency range associated with significant motoric mu suppression. As might be expected, the AE contrast of Action versus Rest was associated with a Cohen’s d of 0.72, indicating a medium to large effect size. In contrast, comparing Coherent versus ­Scrambled AO yielded Cohen’s d = 0.25, representing a small effect size. To translate these values into concrete terms, we entered the calculated effect size into G*Power (Faul et al. 2007) to compute the required sample size to detect an effect using a one-tailed dependent-sample t-test with alpha level of 0.05 and power of 0.8. Whereas a sample size of 14 individuals would be sufficient to detect mu suppression during AE, given the effect size of d = 0.72, the much lower effect size of d = 0.25 obtained from the AO analysis would require a sample size of at least 101 participants. To reliably detect effects with a two-tailed test, this number climbs to 128 participants (cf N = 18 for AE). Notably, these estimates are substantially larger than the sample sizes used in most prior studies, and well above the recommendation of N ≥ 40 in a previous critique (Hobson and Bishop 2017). It is therefore likely that many previous estimates of mu suppression during AO reflect false positives, potentially stemming in part from inadequate visual and attentional controls. These power issues are further compounded by potential masking of the oscillatory mu rhythm by aperiodic activity. Notably, rerunning our original FFT analysis with a one-tailed t-test still failed to recover any significant negative activity in the 8–14 Hz range, implying that a larger sample size alone would not necessarily be sufficient to recover a significant mu suppression effect.

Finally, we examined whether estimation of the AO mu suppression effect would be improved by the use of an additive removal of aperiodic activity (Gyurkovics et al. 2021). Building off our exploratory analysis, we compared estimation of the AO mu suppression effect to Coherent versus Scrambled PLD conditions, averaging over 11–12 Hz. The resulting scalp topographies were largely similar, with significant negative effects over central and temporal electrode sites, though both the observed significance and estimated effect sizes were lower for the additive analysis (estimated Cohen’s d = 0.18, cluster-­corrected p’s < 0.015) compared to the multiplicative aperiodic removal (estimated Cohen’s d = 0.25, cluster-­corrected p’s < 0.0005). Therefore, the additive model does not appear to provide better estimation of the aperiodic component for this dataset.

Discussion

Oscillatory brain rhythms obtained from EEG signals have been associated with specific aspects of human behaviour. One such oscillation, the sensorimotor mu rhythm, has been the object of intense study in the domain of social perception, as it is thought to index common processing underlying the execution and understanding of intentional actions (Springer et al. 2013). In support of this interpretation, researchers have reported reductions in the mu rhythm during both the performance of one’s own and viewing of others’ actions (Babiloni et al. 2002), similar to suppressive effects negatively correlated with sensorimotor activation during motor execution in fMRI (Ritter et al. 2009). However, recent critiques of mu suppression in the literature have highlighted numerous methodological shortcomings in the measurement of this effect, including small sample sizes, limited recording sites, and inadequate controls for visual and attentional processing (Hobson and Bishop 2017). At the same time, there has been growing awareness within the field that estimates of periodic oscillations such as the mu rhythm may be conflated with aperiodic, or non-oscillatory, signals in the EEG (Donoghue et al. 2020).

In this study, we addressed these questions by combining a relatively large sample size (N = 109) with high-density 128-channel EEG. To establish the canonical scalp distribution and frequency range of mu suppression, we collected data from participants performing an AE task. Within the same people, we also recorded EEG during an AO task with careful controls for attention and visual motion cues, allowing us to compare the frequency range and scalp distribution associated with executing one’s own versus observing others’ movements. Consistent with the literature, we found significant and highly localized mu suppression over sensorimotor cortex during motor execution (i.e. finger tapping) compared to rest. Notably, this effect was not substantially altered by removal of the aperiodic component via FOOOF, suggesting that the motor mu rhythm is robust to differences in aperiodic activity.

In contrast, our initial attempts to find mu suppression during AO failed to identify significant reductions over sensorimotor cortex for biologically plausible Coherent PLDs relative to biologically implausible Scrambled controls. This null result gives weight to recent criticisms of the literature regarding AO mu suppression, suggesting that small sample sizes and insufficiently matched control conditions may have inflated previous measurements of mu suppression during AO. When these factors are properly controlled, the difference between mu oscillations to human movements and low-level motion becomes small enough that other factors may contaminate estimates of mu. One such factor is the presence of aperiodic “1/f-like” activity in the EEG signal, which has previously been hypothesized to affect estimation of periodic EEG oscillations (Donoghue et al. 2022, Gyurkovics et al. 2022, Cunningham et al. 2023). In line with this idea, we found that the removal of aperiodic activity enabled recovery of characteristic sensorimotor mu suppression signals during AO. This effect appeared to be driven by differences in the offset, or overall activity level, between Coherent and Scrambled PLD conditions, with the Coherent condition showing a small but significant increase relative to Scrambled. This result would appear to explain our initial failure to find a significant mu suppression effect during AO, as the heightened offset may have masked reductions in the sensorimotor mu rhythm elicited by meaningful human body movements.

One caveat is that the physiological basis of the aperiodic offset remains poorly understood, making it difficult to predict exactly when and how this signal may influence estimation of periodic oscillations. In an earlier study, for example, we found significant spectral power changes consistent with sensorimotor mu suppression despite not removing aperiodic activity (Siqi-Liu et al. 2018). One possibility is that previous estimates of the necessary sample size to detect mu suppression were overly optimistic, such that even samples of ∼40 participants are vulnerable to an increased likelihood of false positives (Fox et al. 2016, Hobson and Bishop 2017). Supporting this idea, a power analysis of our results suggested that AO mu suppression has a Cohen’s d = 0.25, consistent with a small effect size. Moreover, differences in the offset between conditions may vary as a function of other, as yet unidentified factors. Indeed, recent work suggests that aperiodic aspects of the EEG signal vary consistently across participants (Demuru and Fraschini 2020). In this case, estimation of the oscillatory activity evoked by the same paradigm may be pushed in different directions depending on the offset, resulting in inconsistent results between samples. Such effects would presumably be heightened in the presence of rigorous experimental controls, since the effect of the offset would be more pronounced when differences between conditions are small. Future research may benefit from more systematic quantification of the aperiodic exponent and offset terms, so that researchers can build a better understanding of how these signals affect estimation of periodic EEG oscillations.

Although our data suggest that aperiodic activity distorts estimation of the oscillatory sensorimotor mu rhythm, the exact contribution of aperiodic versus periodic activity to the EEG power spectrum in this and other settings remains under debate. For example, a recent study used biophysical modelling to argue that aperiodic activity has relatively little effect on the amplitudes of oscillatory spectral peaks (Brake et al. 2024). Conversely, a study using electrocorticography (ECoG) during simple movement tasks found that finger movements were associated with changes to both broadband power (i.e. offset) and rhythmic oscillations in the beta band (12–20 Hz), but no entrainment in the 8–12 Hz range typically associated with mu (Miller et al. 2012). Clearly, further research will be needed to explore the interplay of aperiodic and periodic components of the EEG spectrum, including the modelling of aperiodic activity and the roles of different frequency bands during motor execution and observation.

Finally, it is worth noting that spectral power is not the only oscillatory information carried by the EEG signal. In particular, the peak frequencies of periodic brain rhythms vary across individuals, potentially differentiating between arousal states, cognitive demands, and psychological disorders (e.g. Mierau et al. 2017). Using a subset of the data reported here, we recently found that individually determined peak frequency of the mu rhythm is significantly correlated with an independent measure of autism-spectrum traits, which have been linked to impairments in body movement perception (Strang et al. 2022). Further exploration of this relationship with larger datasets could shed light on the extent to which peak frequency and spectral power provide overlapping or distinct information about the processing of human body movements.

In conclusion, despite the theorized importance of the sensorimotor mu rhythm for understanding others, recent reports have raised questions about the reproducibility of mu suppression during AO. Here we addressed these concerns using high-density EEG in a relatively large sample, while controlling for low-level motion cues and attentional demand. Using standard analysis methods, we found that these adjustments rendered mu suppression for AO non-significant, suggesting that mu suppression is less robust than previously thought. Nonetheless, we were able to recover significant sensorimotor mu suppression during AO by removing aperiodic activity from the signal, with these effects being driven by increased offsets during Coherent PLD stimulus blocks in our sample. Consistent with previous reports, these results suggest that mu suppression for AO may be confounded by underlying aperiodic activity in the EEG signal, supporting the use of methods to correct for aperiodic activity in the estimation of mu.

Author contributions

Alison Harris (Conceptualization [lead], Formal analysis [lead], Funding acquisition [equal], Investigation [lead], Methodology [lead], Project administration [lead], Resources [lead], Software [lead], Supervision [lead], Validation [lead], Visualization [lead], Writing—original draft [lead], Writing—review & editing [lead]), Chandlyr Denaro (Formal analysis [supporting], Methodology [supporting], Project administration [supporting], Supervision [supporting], Visualization [supporting], Writing—original draft [supporting], Writing—review & editing [supporting]), Catherine Reed (Conceptualization [equal], Formal analysis [supporting], Funding acquisition [equal], Methodology [supporting], Project administration [supporting], Resources [supporting], Supervision [supporting], Writing—­original draft [supporting], Writing—review & editing [supporting])

Conflict of interest: None declared.

Funding

This work was supported by the National Science Foundation [BCS-1923178 to A.M.H. and C.L.R.].

Data availability

The data underlying this article will be shared on reasonable request to the corresponding author.