Fatigue interferes with task performance in multiple ways—slowing down, improper execution, or failure to accomplish. Furthermore, unpleasant sensations such as pain, discomfort, and increased effort can accompany the muscle fatigue. In healthy population, the consequences due to fatigue may be reflected as reduced physical performance, difficulty to carry out strenuous or prolonged activity, and demand for extra effort to perform everyday tasks, which will subside after a period of rest or decreased activity. Nevertheless, in individuals suffering from muscular, neurological, cardiovascular, and respiratory diseases and those who are aging or frail, the effects of fatigue are amplified and posing restrictions on daily life (Taylor et al., 2016).

Physical fatigue can be viewed as the decline in the voluntary-force-generating capacity of the neuromuscular system induced by a physical activity (Fry et al., 2017). It stems from active skeletal muscles that are involved in the peripheral processes as well as supraspinal mechanisms within the brain, thus comprising physiological and psychological aspects (Gandevia, 2001; Berchicci et al., 2013). The fatigue processes at or distal to the neuromuscular junction is referred to as peripheral fatigue and those attributed to the central nervous system affecting the neural drive to the muscle as central fatigue (Bigland-Ritchie et al., 1978).

The force generated by muscle contractions depends on the motor unit (MU) firing in the neuromuscular pathway: the cessation or slowing down of MU firing marks fatigue, whereas the increased recruitment of MUs serves as the compensatory mechanism. The central fatigue changes the MU firing by altering the intrinsic properties of motoneurons, feedback from the sensory input, and descending drive, which in turn modifies the strength and timing of muscle contractions. Besides, a sensation of discomfort and fatigue is facilitated by the firing of group III/IV muscle afferents. Furthermore, the modulation of brain neurotransmitters impacts the task endurance performance (Taylor et al., 2016). Therefore, the neural contribution to muscle fatigue is multi-pronged as depicted in Figure 1 and gaining better insights into this mechanism holds paramount importance. Even though the physical fatigue is influenced by the supraspinal mechanisms within the brain, our understanding on how fatigue modulates the brain activity is far from complete.

Aside from electrophysiological markers, fatigue can also be subjectively quantified by the perception of effort during physical activities, e.g., psychophysiological scale developed in Borg (1982). To understand the central and peripheral mechanisms underlying neuromuscular fatigue while performing isometric muscle contractions, several studies (Falvo et al., 2011; Pincivero, 2011; De Morree et al., 2012) have measured the perception of effort with the Borg category ratio scale.

The functional magnetic resonance imaging (fMRI) studies were dedicated to observe the changes in the cortical activation due to single-limb contractions until self-perceived exhaustion. The sustained maximal (Post et al., 2009; Steens et al., 2012) and submaximal contractions (Benwell et al., 2007; Van Duinen et al., 2007) with hand muscles progressively enhanced the activation in the motor cortex, supplementary motor area (SMA), and sensorimotor cortex. On the contrary, the repetitive fatigue task (Liu et al., 2005b) distinctly differs from the continuous one (Yang et al., 2010) in the sense that the former does not alter the blood-oxygen-level-dependent (BOLD) signal from the motor function cortices. The study findings would thus resolve an important research dilemma: do the motor-related cortices handle the fatigue induced by the repetitive and continuous task differently? Our fatigue study involving repetitive submaximal elbow flexion attempts to clarify whether the band-specific power of the electroencephalographic (EEG) signals recorded over specific cortical regions could be modulated by fatigue. To this end, we analyze the event-related spectral perturbation (ERSP), which is the estimation of post-onset and post-offset changes in the EEG spectrum with respect to the baseline (rest). The onset and offset refer to the time instant marking the application of a task-related stimulus and the cessation of the task, respectively. A more formal definition of ERSP is deferred to Section 4.

A few magnetoencephalographic (MEG) studies reported the effect of fatigue on the spectral power of signals recorded from the sensorimotor cortex. The post-movement beta rebound (PMBR) was reported to increase because of submaximal contractions until fatigue (Fry et al., 2017). Likewise, the muscle fatigue enhanced the beta- and gamma-band power of MEG signals while a motor task was carried out (Tecchio et al., 2006). Nevertheless, a repetitive maximal contraction fatigue task reduced the MEG alpha-band power from the sensorimotor and prefrontal areas (Tanaka et al., 2015). These contradictory results led us to speculate that even among the repetitive tasks, the influence of fatigue on the neural mechanisms may differ between the submaximal and maximal contractions. Our study is therefore intended to verify if the EEG spectral power changes due to fatigue caused by repetitive submaximal contractions agree with the MEG study findings.

There are ample evidences to believe that the fatigue task involving submaximal isometric contractions causes a significant rise in movement-related cortical potentials (MRCP) in brain regions that include primary motor cortex (M1), SMA, and premotor cortex, e.g., studies related to upper limb (Johnston et al., 2001; Schillings et al., 2006; De Morree et al., 2012; Guo et al., 2014) and lower limb (Berchicci et al., 2013). Similarly, a significant increase in the alpha- or beta-band EEG power in SMA, frontal, parietal lobe, and Brodmann area 11 was attributed to fatigue caused by cycling (Schneider et al., 2009; Hilty et al., 2011; Enders et al., 2016). By contrast, intermittent maximal and sustained submaximal contractions significantly reduced the alpha- and beta-band power measured at the EEG electrodes over the central and parietal lobes (Nishihira et al., 1995; Liu et al., 2005a). Thus, the findings related to band-specific EEG power changes due to a fatigue task remain inconclusive, which warrant further investigation.

Fatigue gradually sets in soon after the commencement of contractions, regardless of the fact that the individual continues the task execution (Barry and Enoka, 2007). It explains the reason why fatigue, task failure, and exhaustion are distinguishable (Berchicci et al., 2013). On this premise, we hypothesize that as the execution of repetitive submaximal contractions proceeds in time, the individual would experience increased levels of fatigue, and hence the modulation in the activation of motor-related cortices marking the fatigue level would also scale proportionally. Our study outcome would add knowledge of central mechanism of muscle fatigue as the experiment has been designed bearing the following objectives in mind. (i) The primary goal is to investigate how the EEG spectral power is altered with regard to graded fatigue levels, which has not been explored by past studies. (ii) The secondary goal is to help reach a consensus with the contradictory findings from fMRI, MEG, and EEG fatigue studies involving voluntary muscle contractions. (iii) The future goals are to make use of this knowledge to understand how the spectral power changes with fatigue are modified by a pathology and to design effective rehabilitation techniques targeting patients who have impaired movement functions.

The current analysis was carried out on a previously collected dataset for which only the analysis on the muscle and force data was published (Cai et al., 2014).

The 124-channel EEG data were preprocessed as outlined in Supplementary Figure S1 with the standard tools available in EEGLAB 2021.0 software (Delorme and Makeig, 2004) prior to carrying out the ERSP analyses.

Filtering and Noise Removal. The EEG signals were first band-pass filtered using a Hamming windowed sinc finite impulse response filter with the lower and higher cut-off frequencies of 1 and 100 Hz, respectively. Next, they were filtered using a notch filter to remove the 60 Hz power-line noise and its higher harmonics. The noisy bursts in the filtered EEG data were corrected with Artifact Subspace Reconstruction (ASR) approach, which is an EEGLAB plug-in (Mullen et al., 2015). The threshold for burst detection criterion (specified as k in the ASR algorithm) was set as 20 as recommended in Chang et al. (2019) based on a rigorous evaluation. Afterwards, the EEG signals were re-referenced to the common average, i.e., the arithmetic average of all electrodes (Nunez and Srinivasan, 2006), which is advocated for EEG source estimation.

Epoching and Data Cleaning. The continuous EEG data were then epoched as illustrated in Figure 2 (left) with respect to the event-markers denoting the onset and offset of the elbow flexion task as follows. For analyzing the EEG pertaining to the steady contraction, the epoch limits were selected as 1.5 s before and 5 s after the task onset. To study the post-contraction EEG, the epoch limits were 4 s before and 2 s after the task offset. In the case of post-contraction, the baseline EEG (from 1.5 s before the task initiation up to the onset) from every epoch was used to replace the data within the time range, 4–2.5 s, before the task offset in the respective epoch.

The epoched data was visually inspected for the presence of noisy data. The epochs contaminated with noise, which would possibly affect the outcome of ERSP analysis, and those containing stim pulses (generated by supra-maximal stimulation of the muscle belly or motor nerve connecting the muscle to assess the peripheral fatigue, which is not investigated in this work) were removed.

ICA Decomposition and Dipole Estimation. The multichannel EEG data free of noisy epochs and stim pulses was decomposed by an Independent Component Analysis (ICA) algorithm, namely Extended Infomax (Lee et al., 2000) (“runica” in the EEGLAB toolbox), to estimate the underlying temporally independent EEG sources. The projection of EEG source signals to the scalp surface gave rise to scalp maps. We employed an EEGLAB plug-in, namely DIPFIT (Oostendorp and Van Oosterom, 1989), to identify the location of equivalent dipoles that could best describe the scalp maps. For this purpose, the boundary element spherical head model from the standard Brain Electrical Source Analysis [BESA (BESA GmbH, Gräfelfing, GE)] was adopted. Since the subject-specific digitized electrode positions were not available, the standard Montreal Neurological Institute (MNI) space that contains the dipole locations was applied to the head model.

Artifactual IC Removal. The degree of resemblance between an independent component (IC) scalp map and the projection of the equivalent dipole is quantified using the residual scalp map variance (RV); for instance, the best-fitting model has the smallest RV. The ICs were classified with an artificial neural network framework into two categories: (i) “non-brain-related ICs” representing electrical artifacts stemming from muscle, eye, heart, line noise, and channel noise and (ii) “brain ICs” accounting predominantly for activities originating within the brain. The classifier was implemented with an EEGLAB plug-in (Pion-Tonachini et al., 2019), known as ICLabel. The ICs designated as non-brain-related and those having RV values exceeding 20% were discarded. This process called IC pruning will help remove artifactual IC components that are spatially stereotypical. With the retained neural ICs (RV ≤ 20 %), the EEG signals were reconstructed to produce ICA-cleaned EEG for further analysis.

We conducted the following ERSP analyses. (i) Channel-Level ERSP Analysis. It was performed directly with the EEG data recorded using electrodes mounted on the scalp over brain regions of interest. (ii) Source-Level ERSP Analysis. First, the ICs were estimated by decomposing the scalp EEG data with the Extended Infomax ICA algorithm (Bell and Sejnowski, 1995). Next, the ICs were clustered with the k-means algorithm from the MATLAB Statistics Toolbox based on the location and orientation of the equivalent dipoles to obtain physiologically and anatomically separate brain sources. The parameters of k-means were selected as the default values, i.e., the number of clusters and the standard deviation to decide the outliers were eight and three, respectively. The rule of thumb to determine the number of k-means clusters is suggested as V/2 in p. 365, Mardia et al. (1980), where V is the number of data points to be clustered. In our case, this value coincides with the default choice in EEGLAB: on average there will be one IC component per subject per cluster. Finally, the analysis was carried out with the plausible EEG sources (IC clusters) located within motor-related cortical regions. In both cases, we performed the ERSP analysis with the EEG data corresponding to the steady elbow flexion and post-contraction (hatched blue region in Figures 2A,B, respectively). By visually inspecting the force data in Figure 2 (right) that was simultaneously measured with the force transducer, the steady-contraction period was determined to be between 1 and 3 s after the task initiation. The EEG data within the interval 0–1 s (unhatched blue region in Figure 2A) was excluded from the analysis since the subjects were ramping up to reach certain target force that was set for the IMT trials. Likewise, while performing the ERSP analysis for post-contraction, the data recorded from 0 to 0.25 s after the task offset (unhatched blue region in Figure 2B) was disregarded because it would contain remnant task-related EEG as the subjects were ramping down to rest after cessation of the task. The EEG acquired during the time interval immediately before the task onset, i.e., –0.5–0 s (unhatched pink region in Figure 2A), was not considered as the baseline data, because this data segment would be associated with movement anticipation and hence does not represent the ideal state of rest. For the same rationale, the baseline data from –3 to –2.5 s with respect to the task offset (unhatched pink region in Figure 2B) was not included in the ERSP calculation.

We computed the ERSP based on the formal definition provided in Fuentemilla et al. (2006), Delorme et al. (2007), and Meltzer et al. (2008) for the EEG data in a frequency band of interest during the steady contraction or post-contraction to compare the power spectral density (PSD) between the fatigue conditions at the channel or source level. Furthermore, we implemented the framework illustrated in Supplementary Figure S2 to make statistical inferences. Recall from Section 2.2.2 that the healthy volunteers repeated the elbow flexion until they experienced severe fatigue (SevFatg), which eventually led to the cessation of intermittent contractions. We suppose that there would have been a steady progression of fatigue as the subjects prolonged the IMT execution and ultimately they reached SevFatg just before discontinuing the motor task. On this premise, we divided the entire EEG data into three sets with equal number of trials by preserving the order to maintain the correspondence to minimum fatigue (MinFatg), moderate fatigue (ModFatg), and SevFatg. For the ERSP analysis, we selected only the first, middle, and the last 10% of trials from the first, second, and third set of trials, respectively, to ensure that the data germane to the three fatigue conditions are distinctly disparate. In our EEG dataset, 10% of trials from every fatigue condition varies between the subjects within the range [5, 9] and has a median value of ≈ 8.

Statistical Framework. For notational convenience in the sequel, the subject (or component) index j and the fatigue condition are omitted in Equations (1)–(4). The mean event-related spectrum (ERS¹) is the average data power spectrum across M trials computed for the sliding time windows centered at t in each trial, given by

with S_k(f, t) being the spectral estimate for trial k at frequency f and time instant t. In the present setting, M denotes the total number of trials for a given subject (channel-level) or component in an IC cluster (source-level) corresponding to any one of the three fatigue conditions. The term |Sk(f,t)2| represents the power spectrum for the k-th trial of subject (or component) j. We apply the gain model, which is the default choice in EEGLAB software that divides the contraction-induced spectral activity by the pre-contraction spectral activity. For a given frequency band, the absolute ERSP is computed by dividing the ERS power in Equation (1) computed for the steady contraction or post-contraction at each time-frequency point by the average spectral power in the pre-contraction baseline period at the same frequency.

As pointed out in Grandchamp and Delorme (2011), the log-transformed measure of ERSP is endowed with the following merits: (i) The EEG signal distribution, in general, is skewed. By taking its logarithm, the distribution is made more normal, and hence it becomes suitable for parametric inference testing. (ii) The log-transformed power values help visualize a wider range of power variations, whereas its absolute value leads to the masking of power changes at high frequencies by those at low frequencies. Therefore, the ERSP analyses and the results reported here are based on the log-transformed ERSP in Equation (2) and its time-averaged quantity in Equation (3), both expressed in decibels (dBs):

where

⊘ denotes Hadamard division, and r is the cardinality of the set R of steady- or post-contraction time points. In Equation (4), B_AVG(f) denotes the pre-contraction spectral power as a function of f averaged over a total of q time points belonging to the set Q in the baseline period across p trials. We derived a common baseline for every subject (or component), i.e., the power spectra of the pre-contraction data for the j-th subject (or component) were averaged across the trials, time points, and fatigue conditions. The rationale for using a common baseline across fatigue conditions is to counteract the effect due to fatigue-induced changes in the baseline to the best possible extent. Consequently, p in Equation (4) represents the sum of selected number of trials for a subject (or component) under MinFatg, ModFatg, and SevFatg. We vertically concatenated the row vectors ERSP~jFatgCond for j = 1, 2, …, N to construct a matrix of size N × f as given by

To draw statistical inferences, we conducted the surrogate permutation test with the following: (i) log-transformed ERSP matrices ERSPjFatgCond defined in Equation (2) for frequencies in the interval 3 Hz to 90 Hz and time period shown in Figure 2 (left); (ii) log-transformed time-averaged ERSP vectors that are vertically concatenated as in Equation (5) to build the matrix ERSP~FatgCond. Subsequently, we applied a false discovery rate (FDR) correction test in both scenarios to assess how the FDR was controlled for multiple comparisons (Lage-Castellanos et al., 2010). Finally, we performed a one-way repeated measures analysis of variance (ANOVA) test after averaging the PSD values in ERSP~FatgCond over the EEG frequency bands² of interest, which is henceforth referred to as ERSP≈FatgCond.

Permutation Test and FDR Correction. We studied the log-transformed matrices ERSPjFatgCond and the vertical concatenation of log-transformed time-averaged row vectors ERSP~FatgCond with the surrogate permutation test from the EEGLAB toolbox. We adopted the permutation test (Pesarin, 1992) that draws samples without replacement (Rimbert et al., 2019), because the EEG data may not conform to the normality assumption required by many statistical approaches. Moreover, the test is deemed suitable for our small sample size, i.e., 14, and in scenarios where the data contain outliers, which is more likely our case. We selected the following values for the permutation test: p < 0.05 and 2000 permutations. The test results validated the differences among the three fatigue conditions in terms of ERSP values at each time-frequency point. We applied the FDR correction to control the type I errors in null hypothesis testing when conducting multiple comparisons with a p-value threshold of 0.05. The reason for employing FDR is that it is less conservative compared to family-wise error rate (FWER) controlling procedures and have greater power (Benjamini and Hochberg, 1995).

One-Way Repeated Measures ANOVA and Bonferroni Correction. A one-way repeated measures ANOVA (Schober and Vetter, 2018) was performed with ERSP≈FatgCond obtained by averaging the log-transformed ERSP over the steady-/post-contraction time interval and the frequency band of interest to verify whether their means are different. While ANOVA is considered to be robust when the normality assumption is violated (Kirk, 2012), we rely on the log transformation that enables the data distribution to be made normal, thus rendering it suitable for parametric inference testing. We conducted the multiple comparison test (post-hoc analysis) with Bonferroni correction (Abdi, 2007), when the p-value for the ANOVA was statistically significant to explore the difference between the three group means while controlling for the FWER.

The ERSP analysis was carried out across the three fatigue conditions and the results were reported for the following four schemes: (i) channel-level ERSP for steady contraction; (ii) source-level ERSP for steady contraction; (iii) channel-level ERSP for post-contraction; and (iv) source-level ERSP for post-contraction.

Cortical Regions/Channels of Interest. For the ERSP analysis, we have selected the EEG signals recorded from the electrodes, namely FC3, C1, C2, C3, CP1, and CP3, that overlay the following brain regions: (i) motor cortex comprising M1 (Brodmann area 4), premotor cortex, and SMA (Brodmann area 6); (ii) primary somatosensory cortex (Brodmann area 1, 2, 3); and (iii) superior parietal lobule (Brodmann area 5) (Donoghue and Sanes, 1994; Rosenbaum, 2009). The motor cortex is responsible for executing motor activities, especially contralateral upper-extremity muscle contractions as well as learned motor sequences. The premotor and SMA offer sensory guidance of movement, control proximal and trunk muscles, and plan complex and coordinated motor actions. The primary somatosensory cortex is associated with sensory perception involving arms and hands and helps with motor learning. The superior parietal lobule plays a crucial role in planned movements, receives sensory inputs from peripheral sources, and coordinates fine motor skills (Donoghue and Sanes, 1994; Rosenbaum, 2009). Since all the 14 individuals are right-handed, the EEG electrodes on the motor areas of the left (contralateral) hemisphere were selected.

Relevance of IC Cluster Locations to Motor Task/Fatigue. The task-relevant IC clusters obtained from the steady and post-contraction EEG data point to neural sources residing in the following four regions: left cingulate gyrus, left superior parietal gyrus/left precuneus, right paracentral lobule (SMA), and left transverse temporal gyrus. These brain regions were found to have an association with the upper-extremity motor task performance and/or fatigue by the previous studies. The dorsal anterior cingulate cortex is responsible for the mediation of task-related motor control and has connections to premotor and sensorimotor areas (Asemi et al., 2015). In an fMRI study on the submaximal fatigue task, a steady increase of the BOLD signal in the cingulate gyrus has been observed during both the sustained and intermittent muscle contractions (Liu et al., 2003). The superior parietal cortex has been demonstrated to play a crucial role in the reach-to-grasp action that involves the elbow flexion movement (Cavina-Pratesi et al., 2010). An MEG study with the bimanual load-lifting task showed the involvement of precuneus, which is contralateral to the load-lifting forearm, in the elbow-joint flexion movement (Di Rienzo et al., 2019). The SMA is an important source of upstream drive to M1 and sends projections to M1 (Muakkassa and Strick, 1979). It has an extensive transcallosal connectivity and plays a critical role in uni/bimanual muscle activation (Grefkes et al., 2008). Reportedly, the fatiguing tasks reduce the SMA activity, which in turn decrease the excitatory drive to the corticospinal neuron pool (Sharples et al., 2016).

Frequency Bands of Interest. We investigated the ERSP measure computed in the theta, alpha, and beta band across three fatigue conditions for the following rationale. Theta Frequency Band. An increase in the region-specific theta-band activity has been observed with sensorimotor processing (Cruikshank et al., 2012). The theta activity on the contralateral motor cortex is modulated by acceleration and associated with adaptive adjustments during upper limb movements (Ofori et al., 2015). The theta-band power in the contralateral motor cortex serves as a marker for enhanced control engagement across cortical regions. Moreover, theta synchrony observed at the start of movement in the premotor cortex and anterior cingulate cortex is critical for motor execution (Paus, 2001; Ofori et al., 2015). Therefore, the theta-band power is thought to be linked to the upper limb movement and the acceleration of the movement (Ofori et al., 2015). Alpha Frequency Band. A decrease in the alpha-band activity is commonly encountered during movement [referred to as event-related desynchronization (ERD)], which reflects the excitability of neurons due to activation in a specific cortical region (Toro et al., 1994; Babiloni et al., 1999; Neuper and Pfurtscheller, 2001). A suppression in the alpha-band activity is interpreted as a release from the inhibition caused by alpha ERD (Klimesch, 2012). This phenomenon has been observed in the ipsilateral motor and somatosensory areas near the movement cessation in proportion to the task demands (Ofori et al., 2015). Beta Frequency Band. The beta-band spectral power tends to reduce over bilateral motor cortex at the beginning of movement (Cruikshank et al., 2012; Gwin and Ferris, 2012; Pastötter et al., 2012; Kilavik et al., 2013). However, the beta-band activity in the contralateral motor area is not altered at the acceleration phase of movement but it increased during the deceleration phase (Ofori et al., 2015).

fMRI Studies with Corroborating Findings. The study findings corroborate with the inferences drawn from many fatigue studies. Imaging studies conducted during fatigue tasks involving single-limb contractions induced changes in the activity of motor cortex and other brain regions. Engaging in prolonged submaximal contractions with dominant hand muscles progressively increased the fMRI BOLD signal from the contralateral and ipsilateral sensorimotor cortex, SMA, and cingulate motor cortex (Liu et al., 2003; Benwell et al., 2007; Van Duinen et al., 2007). The activation in the motor cortex (Post et al., 2009) and sensorimotor cortex (Liu et al., 2002; Steens et al., 2012) increased during sustained maximal efforts as inferred from the BOLD signal. The increase in the BOLD signal due to submaximal and maximal tasks is attributed to simultaneous excitatory and inhibitory activity, recruitment of neurons to engage muscles noncritical to the task, and effects induced by the modified sensory feedback (Post et al., 2009). However, in the case of submaximal efforts, an augmented BOLD signal implies an increase in motor cortex output due to excitatory input (Taylor et al., 2016). In a fatigue task study involving repetitive MVCs of finger flexor muscles (Liu et al., 2005b), the fMRI signal level in the primary, secondary, and association motor-function cortices remained unaltered throughout the task as opposed to the findings from a continuous muscle contraction (Yang et al., 2010). The fMRI studies suggest that disparate cell populations might be responsible for controlling sustained and intermittent voluntary motor activities. Thus, our study offered additional evidence to support that the motor cortical centers control the repetitive and continuous fatigue task differently.

MEG Studies. An MEG-based study concerning the influence of physical fatigue due to submaximal contractions on the PMBR and movement-related beta decrease (MRBD) within the contralateral sensorimotor cortex concluded the following: the physical fatigue gives rise to (i) an enhanced PMBR and (ii) cancellation of attenuation in MRBD related to task habituation (Fry et al., 2017). The MEG beta- and gamma-band power reduction with respect to relaxation period was less pronounced during the execution of a non-fatiguing motor task performed after producing muscle fatigue (Tecchio et al., 2006). In contrast to our findings, a fatigue-inducing task session with repetitive handgrips at MVC led to a decrease in the MEG alpha-band power in the ipsilateral sensorimotor and prefrontal areas (Tanaka et al., 2015).

EEG Studies with Corroborating Findings. Submaximal handgrip isometric contractions inducing progressive fatigue resulted in a significant increase in MRCP at the precentral (Cz and FCz) and central contralateral (C3) electrode sites (Johnston et al., 2001). In like manner, the central fatigue induced by submaximal (Guo et al., 2014) and lower limb (Berchicci et al., 2013) isometric contractions increased the amplitude of MRCP in M1, prefrontal cortex, SMA, and premotor cortex. The muscle fatigue was reported to increase the MRCP amplitude during movement execution (De Morree et al., 2012). A significant increase in EEG power within the alpha and beta band was noticed as fatigue developed from a high-intensity cycling exercise in the following cortical regions: parietal and limbic regions (Schneider et al., 2009); Brodmann area 11 (both alpha and beta bands) and parietal lobe (only beta band) (Hilty et al., 2011); SMA, frontal, and parietal lobes (Enders et al., 2016). During repetitive force grip submaximal contractions, the area under the curve of readiness potential⁴ doubled at electrode Cz and increased fourfold at electrodes C3' and C4' (1 cm anterior to C3 and C4, respectively) that cover the motor cortex (Schillings et al., 2006).

EEG Studies with Contradictory Findings. Contrary to what we have observed in the submaximal contractions, the intermittent handgrip MVCs were reported to have significantly decreased the MRCP with fatigue during the sustained phase of muscle contractions. A significant fatigue-related EEG power decline was observed at alpha 2 (C3), beta 1 (C3 and Fz), and beta 2 (C3, Cz, and C4) bands (Liu et al., 2005a). Similar contradictory outcome was reported from a sustained (submaximal) voluntary handgrip contraction to fatigue (Nishihira et al., 1995); a decrease in alpha 1 and alpha 2 power was observed at Cz, C4, P3, and P4 during the second half of the contraction compared to the first half, whereas the power increased at the frontal position. Moreover, the mean EEG power measured during the rest, pre-fatigue, and post-fatigue state from a sustained Adductor pinch task until fatigue performed at two different target values did not produce a significant difference across the three states in alpha, beta, or gamma band (Sh et al., 2020).

Influence of Age and Gender on the PSD-Fatigue Relationship. Even though the participants' age falls within the range 26 − 72 years, only two of them exceeded 65 years. This means that the rest of the subjects were middle-aged (26 − 59 years). Therefore, we refrain from comparing the fatigue-related PSD modulation across different groups based on the age criterion, e.g., two groups consisting of young (< 65 years) and old (≥ 65 years) participants. Nevertheless, we repeated the analyses by excluding the EEG signals from older (≥ 65 years) subjects to verify whether the results could have been impacted by the data from the two outliers—subjects of age 70 and 72 years—in the former analyses. Interestingly, the outcome of both the channel- and source-level-analysis with the data from the remaining 12 subjects is in total agreement with what we have already inferred from the original dataset. Recall that the dataset consists of EEG from six male and eight female volunteers. To make sure that the conclusions arrived at are gender-independent, we carried out the channel-level analysis with the male and female participants separately and examined the outcome. Notwithstanding the gender difference, the spectral power varies relative to the fatigue level for the data from the channels of interest, provided theta-, alpha-, and beta-band frequencies are considered for statistical inferences (refer to Supplementary Figure S4).

Future Perspectives. Past studies have reported the association of fatigue with the following frequency bands in pathological conditions: theta, alpha, or beta band in chronic fatigue syndrome (Siemionow et al., 2004; Flor-Henry et al., 2010) and multiple sclerosis (Leocani et al., 2001; Vecchio et al., 2017); alpha and/or beta band in burnout syndrome (Luijtelaar et al., 2010); theta, alpha, beta, and gamma band in cancer (Allexandre et al., 2020). In this respect, our study outcome will provide a reference and help investigate how fatigue would alter the spectral power in pathologies in comparison to healthy individuals. In particular, our future research direction would be to investigate how the EEG-based spectral power changes due to fatigue compare between cancer or multiple sclerosis patients and healthy volunteers, thus enhancing our understanding of the effect of a brain pathology on the central fatigue mechanism.

Furthermore, in lieu of the common baseline considered in this work that takes into account the rest period corresponding to all the trials, the common baseline derived from only the MinFatg trials may be used. We tested both the baseline options and the findings were consistent.

Study Limitations. Although the EEG data represent brain activities related to fatigue conditions derived from the scalp electrodes and underlying source regions, the data are influenced by connectivity with other cortical and subcortical areas (Jiang et al., 2012, 2016), which could not be determined by this study. Aside from a small sample size, a major limitation of this study is the inter-subject variability in terms of the ERSP measure. For instance, 1-D ERSP plots from two subjects are portrayed in Supplementary Figure S5, where the PSD changes with respect to fatigue levels within a frequency band of interest are quite dissimilar. Admittedly, in a few subjects as exemplified in these plots, we failed to notice a monotonous increase of PSD in a specific motor-related cortical area (or EEG channel) within the theta, alpha, or beta frequency band as the perceived fatigue increases. One plausible explanation could be that the availability of multiple cortical centers that control one motor task or muscle group makes it possible to “shift the activation center” to compensate for fatigue as demonstrated in Liu et al. (2007). However, this phenomenon was not widely observed and the avoidance of central fatigue by shifting neuron populations linked to a fatiguing motor task requires further investigation.

In the channel-level ERSP analysis, the EEG recordings do not entirely represent the electrical potential generated by the cortical tissues right beneath the respective scalp electrodes due to the volume-conduction effect (Nunez et al., 1997). The scalp EEG signals will further be linearly combined when they are reconstructed with the neural ICs extracted via IC pruning. On the other hand, the source-level analysis is carried out with IC clusters produced by k-means algorithm, which does not guarantee that the ICs in each cluster are in close proximity to each other and located in one functional area of cerebral cortex. Furthermore, the source localization is performed with the DIPFIT toolbox in EEGLAB, which introduces inaccuracies—deviation between the location of estimated dipoles and true sources—in the order of around 10 mm (Knyazev, 2013). Also, the lack of subject-specific MRIs and 3-D-digitized EEG electrode locations in the present study would interfere with the precise estimation of dipole positions. The aforementioned factors would have led to the overlapping of information borne by different scalp electrodes or IC clusters in the channel- or source-level analysis, respectively. Consequently, the traces of fatigue-induced alterations in ERSP are likely to be present even in channels and sources that are not directly linked to motor tasks.

Pointers for Improvement. The inherent shortcomings of EEG can be overcome by performing the analysis with other neuroimaging techniques having a better spatial resolution, e.g., electrocorticography and fMRI. To enhance the source localization accuracy, the future studies may investigate distributed source modeling methods recommended in Handiru et al. (2018) as well as incorporate individual MRIs and 3-D-digitized EEG positions (Shirazi and Huang, 2019).

We remark that the ERD—decrease of cortical spectral power in the alpha and beta frequency bands during contractions in relation to the baseline period—was not conspicuous, which is contrary to the observation reported in Cremoux et al. (2013) with the EEG data from an isometric elbow flexion task at different force levels. We speculate that it could be because of our experimental design dedicated to study the effect of fatigue on the cortical spectral power during the steady⁵- and post-contraction period. Finally, we opine that the source-level analysis can be further improved by employing more sophisticated source-separation approaches instead of traditional ones that assume that the underlying sources have a large kurtosis, whereas the distributions of EEG/MEG sources tend to be multimodal (Knuth, 1998).

In the present study, by dividing the EEG data from intermittent submaximal elbow flexions until subjective exhaustion into three equal segments, each one representing progressive levels of fatigue—minimum, moderate, and severe—the changes in various EEG band power with regard to fatigue levels were studied. Key Findings in a Nutshell. Our channel- and source-level analyses offer insights into the effect of progressive fatigue during a prolonged intermittent contraction task on the EEG cortical rhythms: (i) The theta, alpha, and beta PSDs scale with fatigue in cortical motor areas. These signal changes most likely are associated with progressively increasing the effort to continue the motor task to compensate for progressive fatigue. The fMRI (Dai et al., 2001) and EEG-measured MRCP (Siemionow et al., 2000) studies have reported a linear relationship between fMRI/MRCP signals and level of effort (level of voluntary force). (ii) The pairwise PSD differences between the fatigue conditions maintain a statistical significance—especially in minimum vs. severe fatigue. (iii) The relationship described in (i) and (ii) between the band-specific spectral power and the fatigue level is exhibited by the steady- and post-contraction EEG data. (iv) The source-level analysis reveals that the fatigue effect on PSD is prevalent in SMA, superior parietal cortex, and cingulate gyrus during the steady and post-contraction period. (v) The age-related and gender-based differences do not seem to have an influence on the fatigue-PSD behavior.

To the best of our knowledge, the spectral power changes in relation to graded fatigue levels were not investigated by past studies. Furthermore, as mentioned earlier, the inferences drawn would help clarify the inconsistencies and contradictions among the findings from fMRI, MEG, and EEG fatigue studies involving intermittent/sustained maximal/submaximal contractions. Importantly, since the results point to fatigue-induced changes in the cortical activation in healthy individuals, they would serve as a reference to evaluate the alterations in spectral power caused by fatigue in pathological conditions such as cancer and multiple sclerosis. We envisage that the neural underpinnings of graded fatigue would provide pointers to design effective rehabilitation therapies for patients suffering from movement disorders and related fatigue symptoms.

The datasets presented in this article are not readily available because of IRB restrictions. However, the MATLAB scripts and the de-identified preprocessed EEG data associated with this study can be shared, if requests are made to the corresponding author. Requests to access the datasets should be directed to eselvan@kesslerfoundation.org.

The studies involving human participants were reviewed and approved by Local Institutional Review Board. The patients/participants provided their written informed consent to participate in this study.

ES led the research study in terms of conceptualization, scripting, data analysis, interpretation of results, and manuscript writing. VS was involved in conceptualization, scripting, data analysis, interpretation of results, and writing. DA assisted in data analysis, interpreting the results, and writing. AH preprocessed the data. SS contributed to the research idea through discussion. GY oversaw the data collection and helped with conceptualization, interpretation of results, and writing. All authors contributed to the article and approved the submitted version.

This study was supported in part by research grants from the National Institutes of Health (R01CA189665), Cleveland Clinic (RPC6700), Department of Defense (DAMD17-01-1-0665), New Jersey Commission on Brain Injury Research (CBIR19PIL014), and National Institute on Disability, Independent Living, and Rehabilitation Research (ARHF17000003).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.