Slow fluctuations in ongoing brain activity decrease in amplitude with ageing yet their impact on task-related evoked responses is dissociable from behaviour

In humans, ageing is characterized by decreased brain signal variability and increased behavioural variability. To understand how reduced brain variability segregates with increased behavioural variability, we investigated the association between reaction time variability, evoked brain responses and ongoing brain signal dynamics, in young (N = 36) and older adults (N = 39). We studied the electroencephalogram (EEG) and pupil size fluctuations to characterize the cortical and arousal responses elicited by a cued go/no-go task. Evoked responses were strongly modulated by slow (< 2 Hz) fluctuations of the ongoing signals, which presented reduced power in the older participants. Although variability of the evoked responses was lower in the older participants, once we adjusted for the effect of the ongoing signal fluctuations, evoked responses were equally variable in both groups. Moreover, the modulation of the evoked responses caused by the ongoing signal fluctuations had no impact on reaction time, thereby explaining why although ongoing brain signal variability is decreased in older individuals, behavioural variability is not. Finally, we showed that adjusting for the effect of the ongoing signal was critical to unmask the link between neural responses and behaviour.


Introduction 1
Brain signal variability holds valuable information about brain dynamics that should not be ignored 2 when studying the link between neuronal activity and cognition. Brain signal variability can take many 3 forms. Moment-to-moment variability of the ongoing signal reflects how much brain activity changes 4 from one moment to the next, i.e., the brain activity range. Trial-by-trial variability in task-related 5 evoked responses reflects differences in the brain responses to repeated task conditions, while brain 6 signal entropy quantifies the irregularity of the time series. 7 Within-subject brain signal variability has been shown to change with ageing and is a robust marker 8 of age (Grady & Garrett, 2014). The effect of ageing on brain signal variability has been studied using 9 several measures, including temporal standard deviation (SD) and mean square successive not coincide suggesting that BOLD and EEG variability capture different aspects of brain function. In 20 fact, age differences in BOLD SD might be explained by changes in cardiovascular and 21 cerebrovascular factors that affect the link between neuronal activity and changes in blood flow 22 (Tsvetanov et al., 2020). Moreover, BOLD MSSD, a measure of how much the BOLD signal changes 23 across successive observations, is more often increased rather than decreased in older adults 24 related changes in signal dynamics can result in more or less "signal variability" depending on the 26 used metric. In fact, the relationship between the non-invasive measures of brain signal variability 27 used in human neuroscience and neuronal noise is not trivial. Large signal fluctuations will increase 28 signal standard deviation but might originate in a more stable and predictable signal associated with 29 reduced noise. Higher signal variability is, therefore, not equal to a noisier signal. Consistent with this 30 idea, EEG signal entropy increases with ageing and is associated with the EEG spectra slope 31 (Waschke et al., 2017). Older adults present EEG signals with flatter spectra, i.e., more similar to 32 white noise, potentially reflecting increased neuronal noise (Voytek et al., 2015). Higher brain noise 33 is also consistent with the fact that older people present increased behavioural variability measured However, the link between brain signal variability, neuronal noise and behavioural variability remains 1 to be clarified. 2 The ongoing brain signal (i.e. spontaneous brain activity) measured with fMRI or EEG during "task-3 free", "resting" periods reflects activity in large-scale networks each identified as a set of brain regions 4 that show activity co-fluctuations, the resting-state networks (Beckmann et al., 2005;Liu et al., 2017). 5 Notably, these networks are functionally active also during task performance (Abreu et al., 2020;6 Smith et al., 2009). The ongoing activation and de-activation of these networks presents a backdrop 7 of activity on top of which task-related activity occurs (Fox et al., 2006). Not surprisingly, ongoing 8 activity modulates task-related evoked responses that in turn affect task performance (Becker et al., knowledge, no previous study has investigated the effect of ageing on the link between the dynamics 17 of the ongoing brain activity, trial-by-trial variability of evoked responses and behavioural variability. 18 In this study, we aimed at clarifying the effect of ageing on brain signal dynamics and its impact on 19 behaviour by investigating the relationship between variability of ongoing brain signals, evoked 20 responses, and behaviour. We analysed EEG and pupil data from a previously published study 21 acquired while a group of young and a group of older adults were engaged in a cued auditory go/no-22 go task (Ribeiro & Castelo-Branco, 2019a, 2019b. The EEG allows the non-invasive 23 measurement of cortical electrical activity with high temporal resolution, while the pupillogram, Podvalny et al., 2021), these two signals capture distinct aspects of brain function, both important and 29 complementary for our understanding of the ageing brain. In our previous study, we showed that, in 30 both signals, the cue stimulus evoked a preparatory response. In the EEG, the cue evoked a 31 frontocentral negative potential [the contingent negative variation (CNV)] that increased in magnitude 32 throughout the preparatory period. The CNV has been previously shown to correlate with reaction     were associated with more negative ERPs in both groups. Data are represented as mean ± standard error of 4 the mean across participants.

6
Older individuals presented reduced trial-by-trial variability in cue-locked pupil dilation (PD) responses 7 (Fig. 4). The warning cue elicited a pupil dilation response that peaked around 3 s after cue-onset (as   In summary, trial-by-trial variability in the amplitude of the evoked responses was markedly reduced 1 in older adults, while reaction time variability was increased. These observations suggest that the 2 additional variability observed in the young group originates in a signal component that is not 3 behaviourally relevant. In the following sections, we will delve into the origin of the variability in these 4 evoked responses.  The ongoing EEG signal of older adults presented power spectral density (PSD) data that are less 14 steep, with reduced offset (reduced amplitude at low frequencies), reduced alpha power and 15 increased beta power (Fig. 5). Ongoing EEG activity was estimated from the pre-stimulus period (3.5 16 s) just before cue onset, taking advantage of the long inter-trial intervals used. We analysed the PSDs 17 using the FOOOF toolbox that separates the aperiod from the periodic component of the spectra 18 (Donoghue et al., 2020). The aperiodic component can be described by its two parameters: the 19 spectral exponent (a measure of how steep the spectrum is) and the offset (which reflects the uniform 20 shift of power across frequencies). Both were significantly reduced in the older group (Fig. 5B). 21 Occipital alpha power was reduced in older individuals mainly in occipital channels. In contrast, beta 22 power in frontal and central EEG channels was increased in the older group.  offset, and alpha and beta power, extracted from the pre-stimulus PSD of each group of participants. Group 7 differences for each of the parameters studied were estimated using independent t-tests. Channels that showed 8 significant differences after controlling for multiple comparisons are highlighted in black. The graphs on the right 9 show the exponent, offset, alpha power and beta power of the PSDs of young and older groups measured in 10 the channel FCz. Graphs depict individual data points (circles), mean (black horizontal line) and ± standard 11 error of the mean across participants (grey box). Participants where alpha or beta peaks were not detected 12 were excluded from these graphs.

1
The PSDs of ongoing pupil signals of older individuals presented reduced exponent and offset (Fig.  2   6). Ongoing pupil signal fluctuations happen on a slower time scale than the EEG fluctuations, 3 therefore, it is important to analyse longer time periods to study their dynamics. We took advantage 4 of the pupillary recordings obtained at the beginning of the acquisition protocol while participants were 5 fixating and passively listening to the cue stimulus being presented with the same frequency as in the 6 task (Ribeiro & Castelo-Branco, 2019a). Due to the absence of task or luminance changes, the 7 dynamics of pupil fluctuations in these recordings can be mostly assigned to ongoing pupillary activity. 8 We calculated the pupil PSDs on 20 s epochs. The spectra of the ongoing pupillary signal followed a 9 1/f distribution, with power decreasing with increasing frequency (Fig. 6). The aperiodic parameters 10 were significantly decreased in the older group [exponent: t(70) = 7.48, p < .001; offset: t(70) = 4.17, p < 11 .001], i.e., the older group presented a flatter pupil spectrum with lower amplitude at the low 12 frequencies. No obvious peaks were detected in the pupil spectra.  Spectral properties of ongoing signals predict trial-by-trial variability in evoked responses (Table 1). 24 We used correlational analyses to investigate if the dynamics of the ongoing EEG and pupil signals 25 captured by the spectral parameters was associated with the variability in task-related evoked 26 responses across participants. For the EEG, we focused this analysis on the FCz electrode where 27 the CNV showed its highest amplitude. Participants with higher spectral offset and exponent 1 presented higher CNV and PD variability (Table 1; Supplementary Fig. 2). Alpha and beta power were 2 not associated with CNV variability. These findings suggest that the aperiodic signal fluctuations are 3 intrinsically linked to the variability observed in the evoked responses. We then sought to determine 4 which frequencies in the aperiodic spectra were more predictive of the variability in the amplitude of 5 the evoked responses. We found that the correlation between variability in the amplitude of the evoked 6 responses and the amplitude of the power spectra was particularly strong in the lower frequencies 7 ( Supplementary Fig. 3). Importantly, the correlations were significant in both the aperiodic component 8 of the fitted (FOOOF) spectra and in the raw (total) power spectra. In the total spectral power, the 9 correlation between PD variability and spectral power was highest around .44 Hz and decreased for 10 lower frequencies. In the EEG analyses, correlation was highest at 1 Hz, the lowest frequency 11 analysed. These findings suggest that slow aperiodic fluctuations observed in the ongoing signals

20
Adjusting for group differences in the power of the slow ongoing fluctuations eliminated the group 21 differences in the variability of the evoked responses (Fig. 7). Once the effect of the spectral exponent, 22 offset or slow spectral power were regressed out, the CNV variability was no longer significantly  modulating its shape and amplitude -a simple additive mechanism. In this latter case, the phase of 20 the background fluctuations at the cue-onset rather than the instantaneous amplitude would better 1 capture the relationship between the measured evoked responses and the ongoing signal (at the 2 same amplitude values, the phase of the oscillation will define if the signal is increasing or 3 decreasing). We explored this relationship by estimating the phase of the slow fluctuations in the EEG 4 and pupil signals at cue-onset. For illustrative purposes, we plotted the relationship between the pre-5 stimulus phase and the amplitude of the respective evoked responses (Fig. 8) the evoked responses were affected solely by the amplitude of the baseline signal. This is further 10 explored quantitatively in the next paragraph. to the phase of the ongoing signals at cue-onset (mean ± standard error of the mean across participants).

4
Ongoing signal fluctuations strongly affected the amplitude of the evoked responses (Fig. 9). We 5 quantified the association between EEG/pupil phase of slow fluctuations estimated at cue-onset (θ) 6 and the amplitude of the respective evoked responses using linear regression. This relationship can 7 be captured in a linear regression by representing the phase as a pair of variables: the cosine and 8 the sine (Nurhab et al., 2017). We multiplied the cosine and sine functions by the amplitude envelop 9 (r) of the slow signal fluctuations because the amplitude envelop varies throughout the recordings and 10 these fluctuations will have differential effects on the evoked responses depending on its amplitude. 11 The two predictors associated with the slow ongoing signal fluctuations were r • sin θ and r • cos θ. 12 Note that r • cos θ is equal to the signal amplitude and therefore to the amplitude at baseline, while r 13 • sin θ will capture the additional effect of the signal phase. To study how these factors affected the 14 trial-by-trial amplitude of the task-related responses, we run generalized linear regression models 15 (GLMs) for each participant and separately for each signal type -EEG and pupil. We included as 16 dependent variable the amplitude of the evoked response and as predictors r • sin θ, r • cos θ and, 17 additionally, time-on-task (that can lead to differential responses due to learning, adaptation, or 18      responses explained by the models tested increased substantially by including the ongoing signal 31 phase [r • sin(θ)] at cue-onset in comparison with including only the signal amplitude (as captured by 32 the cosine function). Importantly, when we adjusted for the effect of the ongoing fluctuations on the 33 evoked response's amplitude, the correlation between the evoked response and reaction time was at 34 least as strong as without this adjustment, suggesting that the effect of the ongoing signal fluctuations 1 on the evoked responses has no impact on behaviour. Interestingly, in both the pupil and EEG signals 2 adjusting for ongoing signal fluctuations unmasked a relationship between the evoked responses 3 amplitude and reaction time that, without the adjustment, was not detectable in the pupil and was 4 evident only in a restricted number of channels in the EEG. This observation suggests that the 5 additional variability imposed on the evoked responses by the ongoing signal can mask brain-6 behaviour associations and therefore should be attended to and adjusted for. 7 Previous fMRI studies suggest that task evoked responses are affected by ongoing signal fluctuations 8 occurring in the same brain areas but that these ongoing fluctuations are shared across brain regions 9  ., 2004). Likely, these areas will display ongoing activity that will sum to the evoked 16 responses. Nevertheless, the ongoing EEG fluctuations studied probably originate in a large set of 17 brain areas that will include, but not be limited to, the brain areas involved in the preparatory 18 mechanisms linked to the CNV potential. observe changes in pupil size) remains to be solved (Joshi & Gold, 2020). It is possible that, in warned 27 reaction time tasks like the one used in the current study, ongoing fluctuations in pupil size reflect 28 activity in a different brain area from the brain area eliciting the task-related responses. The idea that 29 the ongoing pupil signal and the task-related pupil evoked response can be two largely independent 30 signals, challenges previous beliefs. The prevailing model suggests that high baseline (tonic) activity 31 in the LC is associated with weaker task-related (phasic) responses, and that pupil data parallels this Our analyses showed that ageing affects the dynamics of the ongoing EEG signal (reflecting cortical 3 activity) and the ongoing pupillary signal (reflecting activity in brainstem neuromodulatory systems). 4 In both signals, older people presented flatter spectra with reduced power in the slower aperiodic 5 fluctuations. This flattening of brain signal spectra with ageing has been described before in EEG and 6 ECoG data (Voytek et al., 2015;Waschke et al., 2017), is also evident in MEG data (Vlahou et al., 7 2015), and is here reported in the pupil signal suggesting an age-related decrease in the amplitude fluctuations might therefore be related to impaired recruitment of large-scale brain systems. 22

Conclusions 23
The current study shows that the amplitude of the preparatory evoked potential, the CNV, and the 24 preparatory pupil dilation responses show comparable trial-by-trial variability in young and older adults 25 once the effect of the ongoing signal fluctuations is considered. This finding suggests a dissociation 26 between ongoing brain activity, which presents reduced variability in older adults, and task-related 27 responses, which does not. This dissociation might explain why although ongoing brain activity is less 28 variable in older adults, behaviour variability increases with ageing. Thirty-six young adults (mean age ± SD = 23 ± 3 years; 29 women; 3 left-handed) and thirty-nine 3 older adults (mean age ± SD = 60 ± 5 years; 31 women; 3 left-handed) were included in this study. 4 Participants' characteristics were reported elsewhere (Ribeiro & Castelo-Branco, 2019a). EEG data 5 from one older participant and pupil data from one young adult and one older adult with light-coloured 6 eyes were not included due to poor data quality. 7 The study was conducted in accordance with the tenets of the Declaration of Helsinki and was 8 approved by the Ethics Committee of the Faculty of Medicine of the University of Coimbra. Written 9 informed consent was obtained from the participants, after explanation of the nature and possible 10 consequences of the study. 11 12

Task design 13
The task was designed and run with the Psychophysics Toolbox, version 3 (Brainard, 1997), for 14 Figure 1 and has been 15 described in detail elsewhere (Ribeiro & Castelo-Branco, 2019a, 2019b. Briefly, we applied two cued 16 auditory tasks: a cued simple RT task and a cued go/no-go task. In this study, for the sake of 17 conciseness, we report only the analyses of the go/no-go task data. Equivalent results were obtained 18 with the data from the simple RT task with the exception that PD amplitude, without the adjustment 19

Matlab (The MathWorks Company Ltd). Task design is schematized in
for the slow ongoing fluctuations, correlated with reaction time. Nevertheless, after the adjustment the 20 correlation was significantly stronger. The auditory stimuli used were three different pure tones, 21 suprathreshold, with duration of 250 ms, with the following frequencies: cue -1500 Hz; go stimulus -22 1700 Hz; no-go stimulus -1300 Hz. Participants performed the two tasks, cued simple RT and cued 23 go/no-go, sequentially. The order of the tasks was counterbalanced across participants, i.e., half of 24 the participants started with the simple RT and the other half with the go/no-go. In the simple RT task, 25 the cue was followed by the go stimulus (100 trials) to which participants were instructed to respond 26 by pressing a keyboard key as fast as possible with their right index finger. In the go/no-go task, the 27 cue was followed or by the go stimulus (80 trials) or by the no-go stimulus (20 trials). Participants 28 were instructed to respond as fast as possible to the go stimulus with their right index finger, while 29 refraining from responding to the no-go stimulus. The intertrial interval was variable with a median of 30 7.6 s (min 6.7 and max 19.6 s). The interval between the cue and the target stimuli and between 31 target and the beginning of the next trial were drawn from a nonaging distribution, -W*ln(P), where 32 W is the mean value of the interval distribution and P is a random number between 0 and 1 (Jennings 33 et al., 1998). In our task design, the cue-target interval was 1.5-0.25*ln(P) and the interval between 1 the target and the beginning of the next trial (cue) was 5.2-1*ln(P) in seconds. 2 In the analysis of task performance, we assessed reaction time and task accuracy. We considered as 3 error trials all trials where the participants responded after cue presentation, failed to respond to the 4 go stimulus (misses), responded to the go stimulus too slowly (slower than 700 ms), or responded to 5 the no-go stimulus in the go/no-go condition. These trials were signalled with a feedback tone warning 6 the participants that an error was committed. 7 8 EEG data acquisition and pre-processing 9 As previously described (Ribeiro & Castelo-Branco, 2019a, 2019b, the EEG signal was recorded 10 using a 64-channel Neuroscan system (Compumedics EUROPE GmbH) with scalp electrodes placed 11 according to the International 10-20 electrode placement standard, with reference between the 12 electrodes CPz and Cz and ground between FPz and Fz. Acquisition rate was 500 Hz. Vertical and 13 horizontal electrooculograms were recorded to monitor eye movements and blinks. The participants' 14 head was stabilized with a chin and forehead rest to record pupillographic data simultaneously. 15 Consequently, the electrodes on the forehead, FP1, FPz, and FP2, displayed signal fluctuation 16 artefacts due to the pressure on the forehead rest. These were excluded from the analyses. A trigger 17 pulse was generated at the onset of each stimulus and at every button press. EEG data analysis was 18 performed with the EEGLAB toolbox versions 14.1.1 and 19.1 (Delorme & Makeig, 2004) and Matlab 19 custom scripts. 20 We used independent component analysis (ICA) to eliminate nonbrain artefacts from the data, as 21 described previously. The data, re-referenced to linked earlobes, was then band pass filtered with cut 22 off frequencies of 0.1 and 35 Hz, and periods containing further artefacts were manually removed. 23

24
Pupil data acquisition and pre-processing 25 As described previously (Ribeiro & Castelo-Branco, 2019a), the pupil diameter of the right eye was 26 measured by an infrared eye-tracker (iView X Hi-Speed 1250 system from SMI) with a sampling rate 27 of 240 Hz. Analysis of pupil data was performed using Matlab custom scripts and the EEGLAB toolbox 28 in Matlab. Artefacts and blinks were corrected using the blink correction Matlab script by Greg Siegle 29 (stublinks.m) available in http://www.pitt.edu/~gsiegle/ (Siegle et al., 2003). Briefly, artefacts, including 30 blinks, were identified as large changes in pupil dilation occurring too rapidly to signify actual dilation 31 or contraction. Linear interpolations replaced artefacts throughout the data sets. Data were smoothed 32 using a 3-point unweighted average filter applied twice. 33 Pupil epochs were visually inspected for artefacts not adequately corrected by the linear interpolation 1 procedure. Epochs with remaining artefacts were manually rejected. In addition, we discarded all error 2 trials and correct trials that immediately followed error trials. The numbers of included trials did not 3 differ significantly across groups for each condition. 4 For each run, pupil data was expressed as the percentage of the mean across the whole run by 5 dividing the run data by the mean pupil size within that run.  In the EEG analyses, PSDs were estimated in single trials in 3.5 s long pre-cue epochs. PSDs were 4 then averaged across trials for each electrode and fit using the FOOOF algorithm with the following 5 settings: peak width limits = [1 8]; peak threshold = 1; aperiodic mode = 'fixed'; minimum peak height 6 = 0.1; max number of peaks = 4; frequency range = [1 35]. Peaks with peak frequency between 7 and 7 14 Hz were considered within the alpha range and peaks with peak frequency between 14 and 30 Hz 8 were considered within the beta range. Peaks outside these frequencies were rarely detected and 9 were not analysed. 10 In the pupil analyses, PSDs were estimated in 20 s long epochs from a 4 min recording acquired at 11 the beginning of the experiment during which the participants were passively fixating and listening to 12 the cue stimulus being presented with the same frequency as in the cued auditory tasks but without FOOOF model fitting quality was estimated from the model R 2 . For the EEG data, average R 2 across 20 electrodes and across participants was .99 for the young group and .97 for the older group. These R 2 21 values reflected good model fittings, however, it is important to note that they showed significant 22 differences across groups in most EEG electrodes (Supplementary Figure 5A and B). As the R 2 values 23 were lower in the older group, it is possible that the parameter estimation was not as good in this 24 group. Surprisingly, the model R 2 correlated with the estimated values for exponent and offsetlower 25 R 2 values were associated with lower exponents and offsets (Supplementary Figure 5C). This 26 relationship was evident in both groups of participants. To ensure that the study conclusions were not 27 driven by differences in fitting quality, we compared the correlation between the variability in the 28 evoked responses and the fitted spectra with the correlation between the variability in the evoked 29 responses and the raw PSD (supplementary Fig. 3). We obtained similar results: the power of low 30 frequencies in the aperiodic fitted spectra and in the raw spectra correlated with the evoked responses 31 amplitude variability. The R 2 for the fitting of pupil PSDs were not significantly different across groups 32 (R 2 young = .99; R 2 older = .98; Supplementary Figure 5D). Pupil R 2 correlated with the estimated 33 PSD offset values but not with the PSD exponent. 34 For the correlation analyses between the spectral parameters of the ongoing signal measured at FCz 1 and CNV variability, participants with spectral goodness-of-fit more than 2.5 standard deviations away 2 from the mean, calculated as the z-score of R transformed into Fisher's Z, were excluded from the 3 analyses (one young and one older adults). These were also the participants with lowest spectral 4 exponent. The instantaneous phase and amplitude envelope of the slow fluctuations was estimated using the 8 Hilbert transform. Before applying the Hilbert transform, the pre-processed EEG signal was bandpass 9 filtered between .1 Hz and 2 Hz and the pre-processed pupil signal was bandpass filtered between .1 10 and .9 Hz. Phase (θ) and amplitude envelope (r) at the time of cue-onset were extracted and used to 11 calculate two variables that represent the circular phase of the signal in Cartesian coordinates: r • sin 12 θ and r • cos θ (Nurhab et al., 2017). These variables were included as predictors in the linear 13 regressions studying the effect of pre-stimulus phase angle on the amplitude of the evoked responses. 14 Note that r • cos θ is equal to the amplitude of the filtered signal at each time point. 15 16

The effect of pre-stimulus variables on single trial amplitude of evoked responses 17
We run within-subject generalized linear regression models, with the single-trial amplitude of the cue-18 locked evoked responses (average within the time interval from 1 to 1.5 s after cue-onset) as response 19 variable and including run as categorical predictor (two sequential runs were acquired per participant), 20 and, as continuous predictors, time-on-task (defined as the trial number within each run), alpha power 21 (estimated from the FOOOF model fitting of 3.5 s pre-stimulus epochs in electrode POz where it 22 showed the highest amplitude; when no alpha peak was detected, alpha power was set at zero), r • 23 sin θ and r • cos θ (estimated as described above for each signal). 24 The models explained a large amount of variance of the evoked responses amplitude, however, the 25 explained variance was higher in the young than in the older group [independent samples t-test 26 comparing the models'R 2 of both groups: EEG t(72) = 2.11 p = .039; pupil t(70) = 2.88, p = .005]. To adjust the evoked responses amplitude for the effect of ongoing signal fluctuations, we regressed 31 out the effects of the phase of the signal at cue-onset by taking, for each participant, the residuals 32 from the multiple linear regression with response variable the evoked responses amplitude and r • sin 1 θ and r • cos θ as predictors. The regression was run separately for each signal (EEG and pupil).