Atypical paroxysmal slow cortical activity in healthy adults: Relationship to age and cognitive performance

Paroxysmal patterns of slow cortical activity have been detected in EEG recordings from individuals with age-related neuropathology and have been shown to be correlated with cognitive dysfunction and blood-brain barrier disruption in these participants. The prevalence of these events in healthy participants, however, has not been studied. In this work, we inspect MEG recordings from 623 healthy participants from the Cam-CAN dataset for the presence of paroxysmal slow wave events (PSWEs). PSWEs were detected in approximately 20% of healthy participants in the dataset

Recently, Milikovsky and colleagues (2019) observed that this pathological slowing is associated with transient (5-10 second) patterns of slow-wave activity, rather than a persistent slowing of the sustained signal.The authors termed these transient shifts in network activity "Paroxysmal Slow-Wave Events" (PSWEs) and showed that the duration and number of PSWEs were underlying the slowing effects observed in the average spectral power analysis (Milikovsky et al., 2019).They also detected PSWEs at a higher rate in both epilepsy and AD populations compared to healthy controls and replicated these findings in rodent models of ageing and epilepsy (Senatorov et al., 2019;Milikovsky et al., 2019;Zelig et al., 2022).
In addition to being related to a diagnosis of epilepsy or AD, PSWEs were found to be related to cognitive impairment and blood-brain barrier dysfunction (BBBd) in animals and patient populations (Milikovsky et al., 2019).A negative relationship was observed between cognitive performance and PSWE occurrence in patients with AD suggesting that PSWEs could be a marker for cognitive impairment in patient populations.This observation supported previous findings that low frequency (e.g., delta and theta) oscillations are related to cognitive impairment in older adult populations (Adler et al., 1999;Benwell et al., 2020).In addition, it was found that patients with a high rate of PSWEs were more likely to have a higher percentage of blood-brain barrier leakage than patients with fewer PSWEs and that PSWEs and BBBd tended to be spatially colocalized in patients with epilepsy.
The established relationship between PSWEs, cognitive impairment, and BBBd suggests that PSWEs may be a signal of interest in other neurological processes that are associated with cognitive impairment and BBBd.One such process for which the presence of PSWEs has not yet been investigated, is the process of normal (i.e., non-pathological) ageing.Non-pathological ageing is associated with a decline in cognitive functions including memory, executive function, processing speed and reasoning that occurs from middle age onwards (Deary et al., 2009a;Hedden and Gabrieli, 2004;Park and Reuter-Lorenz, 2009).This cognitive decline is thought to be related to subtle and widespread changes that lead to atrophy of neural tissue and overall reduction in brain volume (Deary et al., 2009a).The extent of cognitive ageing varies significantly between individuals and depends on a variety of factors including genetic (Deary et al., 2004;Deary et al., 2009b), cardiovascular (Hochstenbach et al., 1998;Rafnsson et al., 2009Rafnsson et al., , 2007)), dietary (e.g., lack of B-vitamins, antioxidants, omega-3 fatty acids, etc.) (Deary et al., 2009a), and lifestyle (e.g., smoking, alcohol consumption, physical activity levels, etc.) (Fratiglioni et al., 2004;Ganguli et al., 2005;Nooyens et al., 2008).In addition, non-pathological ageing is associated with BBBd which may further contribute to cognitive decline (Senatorov et al., 2019).A combination of physiological factors including the accumulation of iron in astrocytes (Connor et al., 1990) and decreased activity of transporters involved in the extrusion of toxins from the brain (Toornvliet et al., 2006) can participate in alterations to the blood-brain barrier that are seen with age (Popescu et al., 2009).Other conditions that increase in prevalence with age such as hypertension and type-2 diabetes are risk factors for BBBd and vascular dementia which can further contribute to age-related cognitive decline (Popescu et al., 2009).The evidence for cognitive decline and BBBd with normal ageing and age-related pathology (Montagne et al., 2015;Sweeney et al., 2018;van de Haar et al., 2016) raise the hypothesis that PSWEs may also be found in non-pathological ageing and be associated with cognitive decline.
Electrophysiological studies of non-pathological ageing, however, have provided mixed results.In animal models of ageing, older mice were found to have increased low-frequency activity in EEG (Senatorov et al., 2019) and a higher number of PSWEs compared to younger mice (Milikovsky et al., 2019;Senatorov et al., 2019).Human electrophysiological studies, however, have either found no significant changes in low-frequency oscillatory activity with age (Caplan et al., 2015;Cesnaite et al., 2023), or have found reduced low-frequency activity in healthy older adults (Emek-Savaş et al., 2016;Leirer et al., 2011;Meghdadi et al., 2021;Vlahou et al., 2015).To our knowledge, the presence of PSWEs in non-pathological ageing has not been investigated in humans, and all other studies of low-frequency activity in healthy ageing have used average spectral power approaches and relatively small datasets.The limitation of average spectral power analysis, however, is that changes in spectral power can be caused by various rhythmic signal characteristics including power, frequency, duration, or number of highpower events in the signal.This limitation can be overcome by detecting and characterising individual transient bursts in the raw data.A large body of literature has explored the presence and characteristics of high-power transient bursts in the mu, beta, and gamma frequency bands (Brady et al., 2020;Milikovsky et al., 2019;Sherman et al., 2016;Shin et al., 2017;van Ede et al., 2018) and has demonstrated the diverse relationships between burst characteristics and spectral power.The success of previous literature suggests that the investigation of the characteristics of individual PSWEs in a large dataset of healthy individuals will allow us to extract more specific information about event characteristics and relate these to age-related effects previously observed using spectral power analyses.The limitations of previous spectral power analyses along with the discrepancy between electrophysiological studies of healthy J o u r n a l P r e -p r o o f ageing and studies of pathological populations and animal models, suggests a need for further detailed investigation of slow-wave activity in a large cohort of healthy participants.
The objectives of the current work are to determine the prevalence of PSWEs in healthy populations and to identify the relationship between age, cognitive performance, and PSWE characteristics.We used a large open-access dataset of healthy human MEG data collected by the Cambridge Centre for Ageing and Neuroscience (CamCAN;Shafto et al., 2014;Taylor et al., 2017) for the detection of PSWEs.We hypothesised that PSWEs increase in prevalence with participant age and are associated with lower cognitive performance.The current work provides complementary insight into the functional impact of slow-wave activity in healthy ageing, improves our understanding of age-related functional brain changes, and may promote the use of EEG/MEG for the early diagnosis of age-related cognitive decline.

Participants and Experimental Paradigm
MEG data was collected by the Cambridge Centre for Ageing and Neuroscience (Cam-CAN) in Phase 2 of the Cam-CAN examination of healthy cognitive ageing.A key characteristic of this large open access dataset is the uniform age distribution between 18 and 88 years.The current work uses MEG data acquired during the "resting state" component of the study.In this component, approximately 9.5 minutes of data were obtained from 623 healthy participants while they rested with their eyes closed.
Participants also completed a series of 14 domain-specific cognitive tasks as part of the Cam-CAN examination, spanning several cognitive domains including emotion, language, memory, and executive function (Shafto et al., 2014;Taylor et al., 2017).Six of these tasks were excluded from the current analysis due to subjectivity of their scoring metrics.The remaining 8 tasks included 2 language tasks, 3 memory tasks, 2 executive function tasks, and 1 emotion recognition task.Descriptions of all included tasks are provided in Table 1.Of the 623 participants included in the MEG data analysis, 500 had available data for all 8 cognitive tasks.To assess differences in cognition related to PSWE characteristics, an average Z-score was calculated to represent the overall cognitive ability for each participant.Z-scores were calculated relative to all participants (i.e., 500 participants with available cognitive data) for each of the 8 cognitive tasks individually and were then averaged across all tasks to generate an overall cognitive score.In the case of the tip-of-tongue and hotel tasks, lower scores were indicative of better cognitive performance and therefore the signs of the raw scores were flipped prior to calculating the Z-scores.For all other tasks, the raw scores were not modified.
In addition, participants completed Addenbrooke's Cognitive Examination-Revised (ACE-R) which is a brief cognitive assessment that is useful for detecting dementia and mild cognitive impairment.This assessment was used as part of the Cam-CAN screening procedure to exclude participants with neuropathology, but it also provides an overall standardized cognitive score J o u r n a l P r e -p r o o f that is a useful addition to our analysis.The ACE-R score is therefore included in our cognitive analysis.However, ACE-R scores are strongly skewed towards high values in this relatively cognitively intact population.Thus, the ACE-R scores are analysed separately from the aggregate Z-scores described above.

MEG Data Acquisition
Data were obtained from the Cam-CAN repository (available at http://www.mrccbu.cam.ac.uk/datasets/camcan/ (Shafto et al., 2014;Taylor et al., 2017).MEG data were acquired at 1000 Hz with inline band-pass filtering between 0.03 and 330 Hz using a 306channel Vectorview system with continuous head position monitoring (Elekta Neuromag, Helsinki, Finland).Digitization of anatomical landmarks (i.e., fiducial points; nasion and left/right preauricular point) as well as additional points on the scalp was also performed for registration of MEG and MRI coordinate systems.Electrooculogram (EOG) and electrocardiogram (ECG) were recorded concurrently.

MEG Data Processing
Data were pre-processed by the Cam-CAN group using temporal signal space separation to perform environmental noise reduction, reconstruction of missing or corrupted MEG channels, continuous head motion correction, and a transform of each dataset to a common head position (Taulu and Simola, 2006).All subsequent MEG processing was completed in the Python programming environment (v.2.7.13), using the MNE-python library (v.0.18.1) (Gramfort et al., 2014).Data analysis scripts used in this work are available at https://github.com/lindseypower/CamCAN_detectPSWEs.Raw MEG data were bandpass filtered between 1 and 50 Hz.The data was then cropped into a single 9-minute epoch beginning 15 seconds after the onset of the recording.Independent component analysis was performed on the cropped data using the FASTICA algorithm (Delorme et al., 2007;Hutchinson, 2002) to remove artifacts using a fully automated process.Epochs with signals that exceeded 5 pT (magnetometers) or 400 pT/cm (gradiometers) were not included when calculating the deconstruction.Following this, components were excluded if the amplitude and phase of the component was similar to that of the EOG or ECG (Bardouille et al., 2019).An average of 4 +/-1 (mean +/-standard deviation) independent components were removed from each subject's data.This process resulted in cleaned MEG data (i.e., channels x time) which was used for subsequent analysis.

Paroxysmal Slow Wave Event Detection
J o u r n a l P r e -p r o o f Paroxysmal slow wave events (PSWEs) were detected at each MEG channel (102 magnetometers and 204 gradiometers) in each participant.Briefly, MEG data was windowed into 2 second windows with a 1 second overlap.The power spectrum was then calculated for each time window and the median power frequency (MPF) was selected for that window.The algorithm then found all windows that had an MPF between 1-6 Hz, and if more than 4 consecutive windows (i.e., at least 5 seconds of data) met this low-frequency criteria, then this segment of data was considered a PSWE.The low frequency threshold of 6 Hz was defined based on ROC (receiver operating characteristic) analysis to determine the frequency that best separated patients with AD from healthy age-matched controls in prior work.Thresholds of 2,4,6, and 8 Hz were compared, and it was found that 6 Hz resulted in the greatest area under the curve in the ROC analysis and was therefore selected as the frequency threshold that would characterize PSWEs (Milikovsky et al., 2019).Following PSWE detection, characteristics of each PSWE, including the onset and offset times (s), duration (s), and mean frequency (Hz) were calculated (see Figure 1).For each participant, events were collapsed across channels such that any events that occurred at different channels but had overlapping onset and offset times were averaged together to create a single event.Specifically, events were ordered based on their onset times, and for each event, if the offset time was later than the onset time of an event from another channel, then the events were pooled.The characteristics of the overlapping individual channel events were averaged to obtain an average event duration and frequency that would characterise the multichannel event.The number of channels involved in the multi-channel event was also recorded as an event characteristic.After pooling events across channels, events with only 1 active channel were excluded because activity generated within brain tissue is generally captured by J o u r n a l P r e -p r o o f multiple sensors and single-channel effects are often artefactual.In addition, due to the lowfrequency nature of the signal of interest, events were excluded from the analysis if they occurred during significant head movements.Specifically, 1-sample t-tests were used to determine whether the head position velocities during an event were significantly different from the mean head position velocity during the entire scan.Those events that had velocities that were significantly different (p<0.05/number of events) from the mean were excluded from further analysis.
PSWEs were further characterised based on their spatial distribution in the MEG sensor array.MEG channels were divided into 8 regions based on their lobe (frontal, parietal, temporal, occipital) and hemisphere (left, right) placement (see supplemental Figure S1 for illustration of channels by region).For each PSWE, the number of contributing channels from each of the 8 regions was counted and recorded as an indicator of spatial distribution.

Demographic Analysis
Following the detection of PSWEs in all datasets, participants were split into groups based on PSWE presence (i.e., "PSWEs" and "No PSWEs").Chi-squared tests were performed for characteristics of age, sex, ACE-R scores, and aggregate cognitive scores using a significance threshold of alpha=0.05.Effect size was measured using Cramer's V.

Statistical Analysis
Spectral power in canonical frequency bands including delta (0-4 Hz), theta (4-8 Hz), alpha (8-12 Hz) and beta (12-30 Hz) bands were compared between participants with and without PSWEs using t-tests.Spectral power was calculated for gradiometers and magnetometers separately and a Bonferroni-corrected significance threshold of 0.00625 (0.05/8 tests) was used for all ttests.Effect size was measured using Cohen's D.
For those participants who had detectable PSWEs (N=127), statistical analyses were conducted assessing the relationship between PSWE characteristics and participant age/cognitive performance using the R statistical software packages (R Core Team, 2021).Characteristics including event duration, event frequency, number of channels, and number of regions were calculated for each PSWE.In addition, an overall indicator of event prevalence (i.e., time in events) was calculated for each participant.Time in events was calculated as the sum of the durations of all detected PSWEs (in seconds) for a given participant.For each characteristic, the distribution of values was plotted and Kolmogorov-Smirnov tests with alpha=0.05 were used to assess the normality of the distribution as well as the normality of the residuals for a linear model.In cases where the relevant distribution was significantly different from normal (e.g., right skewed), a logarithmic transform was applied to the data to improve normality.
For each characteristic, simple linear regression models were used to assess the relationship between the characteristic and participant age, and between the characteristic and cognitive score.A forward stepwise approach was then used to add cognitive score as an additional J o u r n a l P r e -p r o o f predictor to the linear model with age, and to add age as an additional predictor to the linear model with cognitive score.Akaike information criterion (AIC) was used to compare models with and without the additional predictors to assess the contribution of the additional predictors to the model.
For time in events, the relationship with age and cognitive score were relatively weak (see Figure 6).However, when inspecting the distribution of log-transformed time in events values across all participants, it was found to be approximately bimodal with an excess mass of 0.0871 (p=0.016).Based on this, a bimodal curve was fit to the distribution and the data were split into "less" and "more" time in events groups based on a splitting point that was defined as the local minimum between the two modes.Differences in participant age between groups with more, less, and no PSWEs were then assessed using a Kruskal-Wallis test (i.e., non-parametric ANOVA), followed by pairwise Wilcoxon rank sum tests in cases where the Kruskal-Wallis was significant with alpha=0.05.Similar statistical analysis was used to assess differences in cognitive score between groups.To evaluate the association between PSWE prevalence and particular cognitive domains, Kruskal-Wallis tests were also conducted comparing Z-scores for individual cognitive tasks between participants with more time in events, less time in events, and no events.A Bonferroni-corrected alpha of 0.00625 (0.05/8 cognitive tests) was used to assess significance.Effect sizes for Kruskal-Wallis tests were measured using eta 2 .
To contextualize the PSWE findings in relation to traditional metrics, model comparisons between linear models of 1-6 Hz average power, 1-6 Hz root mean square (RMS), and time in events with age were conducted.Linear models were constructed using the entire Cam-CAN cohort (N=623).Average power and RMS were calculated and analysed separately for magnetometers and gradiometers.In addition, previous literature investigating the relationship between low-frequency spectral power and age has reported decreases in low-frequency power with age in healthy populations and increases in low-frequency power in age-related neuropathology.To contextualize the differences between PSWE prevalence groups in relation to these previous findings, the interaction effect of age and PSWE group on power in the 1-6 Hz band was tested using a multiple linear regression model.Similarly, the interaction effect of cognitive score and PSWE group was also investigated.Average power was calculated and analysed separately for magnetometers and gradiometers.

Spatial Analysis
To assess the spatial distribution of PSWEs, the number of active channels in each of 8 spatial regions were compared.Statistical analyses were conducted for gradiometers and magnetometers separately.For each set of events, a Kruskal-Wallis test with alpha = 0.05 was used to determine whether there were significant differences in the number of active channels in the different spatial regions during PSWEs.When significant differences were found, multiple pairwise comparisons were subsequently used to determine which regions were significantly different from one another.A Bonferroni-corrected alpha of 0.00179 (0.05/28 pairs) was used to assess significance of the pairwise comparisons.

J o u r n a l P r e -p r o o f
Spatial maps were also created to provide a visual representation of PSWE distribution across channels.The total number of PSWEs detected at each channel (across all participants) was calculated and plotted on a spatial topography.Separate plots were created for magnetometer and gradiometer arrays.

PSWEs and Power Spectral Density
Of the 623 participants included in the PSWE analysis, 127 (~20%) had any detectable PSWEs.

Demographic Differences
Significant differences in demographic characteristics were also found between participants with and without PSWEs (see Figure 5).Chi-squared tests comparing the distributions of each group revealed that participants with PSWEs have a significantly different age distribution than those without PSWEs (X 2 (df=9) = 87.2,p=5.96e-15, effect size=0.265) such that there is a higher J o u r n a l P r e -p r o o f proportion of older participants with PSWEs compared to those without.In addition, there was a higher proportion of females with PSWEs compared to those without PSWEs (X 2 (df=1) = 4.62, p=0.0316, effect size=0.0609).Cognitive performance was also skewed to lower values for participants with PSWEs compared to those without PSWEs for both aggregate cognitive scores (X 2 (df=8) = 16.5, p=0.0361, effect size=0.115) and ACE-R scores (X 2 (df=9) = 36.5,p=3.19e-5, effect size=0.265).

J o u r n a l P r e -p r o o f
For participants with PSWEs, the PSWE prevalence (i.e., time in events) was further investigated for trends related to age and aggregate (mean Z) cognitive score.The aggregate cognitive scores were chosen over ACE-R scores for further investigation because the values were more normally distributed.
Within the group of participants with PSWEs (N=127), log-transformed time in events values were regressed against age and cognitive score using simple linear regression.A significant increasing trend was found with age (F(df=1,125) = 4.46, p=0.0366,R=0.164), but no significant effect was found with cognitive score (see Figure 6).Stepwise addition of age to the linear model of cognitive score resulted in a significant improvement in the model AIC, but stepwise addition of cognitive score to the model of age did not improve AIC.Based on a comparison of AIC between all models, the model containing age as the only predictor was the best fit for the data.
In addition to the regression analysis, it was determined based on an excess mass test, that the distribution of the log-transformed time in events values were approximately bimodal (excess mass = 0.0871, p=0.016).Participants who had detectable events were therefore split into two groups based on the log-transform of their time in events.Fitting a bimodal curve to the distribution revealed modes at 0.8 and 2.1 and a splitting point at 1.6."Less PSWE" participants were thus defined as participants with a log-transformed time in events value less than 1.6 (approximately 40 seconds; N=76), and "more PSWE" participants were defined as those with a log-transformed time in events value greater than or equal to 1.6 (N=51).Note that the total recording time for each participant was 540 seconds, therefore, participants with more time in PSWEs spent at least ~7% of the recording in PSWEs.Kruskal-Wallis tests comparing the ages of participants with more PSWEs, less PSWEs, and no PSWEs found significant differences between groups (H(df=2) = 26.6,p=1.65e-6, effect size = 0.0397).Pairwise Wilcoxon rank sum tests revealed that those with more PSWEs were significantly older than those with less PSWEs (p=0.019) and those with no PSWEs (p=7.7e-6), and that those with less PSWEs were significantly older than those with no PSWEs (p=0.015).The mean (+/-standard error) participant age was 52 +/-1 for participants with no PSWEs, 58 +/-2 for participants with less PSWEs, and 65 +/-2 for participants with more PSWEs.Kruskal-Wallis tests comparing the cognitive scores of participants with more, less, and no PSWEs revealed significant differences between groups (H(df=2) = 15.1, p=5.25e-4, effect size=0.0211).Pairwise Wilcoxon rank sum tests revealed that participants with more PSWEs had significantly lower cognitive scores than those with less PSWEs (p=0.014) and those with no PSWEs (p=3.2e-4).No significant differences in cognitive scores were found between those with less and no PSWEs.The mean (+/-standard error) cognitive score was 0.030 +/ 0.029 for participants with no PSWEs, -0.060 +/-0.084 for participants with less PSWEs, and -0.29 +/-0.079 for participants with more PSWEs.
J o u r n a l P r e -p r o o f Analysis of individual cognitive task data revealed that participants with more time in PSWEs had lower mean z-scores than those with less or no time in PSWEs for all cognitive tasks (see Figure 7).However, the only cognitive tasks for which this difference was statistically significant with Bonferroni-corrected alpha=0.00625were the fluid intelligence task (H(df=2) = 14.5, p=7.20e-4, effect size=0.0201), and the hotel task (H(df=2) = 11.2, p=3.74e-3, effect size=0.0148).In both cases, the more PSWEs group had significantly lower Z scores than the no PSWEs group (fluid intelligence p=0.0015; hotel p=0.004).
J o u r n a l P r e -p r o o f

Comparison to Common Metrics
Linear model comparisons between 1-6 Hz average power, 1-6 Hz RMS, and time in PSWEs with age revealed that time in PSWEs explained the most variance in age (see Figure S4, Table S1).This finding suggests that the time in PSWEs metric is more sensitive to age-related effects than other traditional metrics, and thus provides novel and valuable information about age-related brain activity.A complete breakdown of this analysis, including figures and statistical results is included in the supplemental material (Figure S4, Table S1).
Multiple linear regression assessing the effects of age and PSWE prevalence on low-frequency band power revealed significant effects (gradiometers: F(df=5,617) = 12.7, p=9.40e-12,R=0.293; magnetometers: F(df=5,617) = 18.4,p<2.2e-16,R=0.351; see Figure 8).Particularly, an interaction effect was found whereby participants in the less and no PSWE groups showed a decreasing trend in 1-6 Hz band power with age, in line with previous findings of healthy ageing (Emek-Savaş et al., 2016;Leirer et al., 2011;Meghdadi et al., 2021;Vlahou et al., 2015), while participants in the more PSWEs group showed an increasing trend in 1-6 Hz band power with age.This interaction effect was significant for magnetometer power values (age by more-less: p=0.0496; age by more-none: p=0.0429) but did not reach significance for gradiometers (age by more-less: p=0.0563, age by more-none: p=0.0517).Multiple linear regression assessing the effects of cognitive score and PSWE prevalence on low-frequency band power also revealed significant effects (gradiometers: F(df=5,617) = 11.3,p=2.09e-10,R=0.276; magnetometers: F(df=5,617) = 16.6, p=2.38e-15,R=0.333).In this case however, there was a significant effect of J o u r n a l P r e -p r o o f group (high-none) for gradiometers (p=9.60e-5) and magnetometers (p=1.03e-7),but no significant interaction effect of cognitive score and PSWE prevalence.
Figure 8. Interaction effect of age and PSWE prevalence on 1-6Hz band power for gradiometers (left) and magnetometers (right).Significance is indicated by asterisks.

Burst Characteristics of PSWEs
In addition to PSWE prevalence, PSWE characteristics including event duration, event frequency, number of channels, and number of regions were regressed against age and cognitive score (see Figure 9).For event duration, there was a significant increase in logtransformed event duration with age (F(df=1,916) = 23.6,p=1.40e-6,R=0.155) and a decrease with cognitive score (F(df=1,916) = 7.63, p=5.84e-3,R=0.0847).The stepwise addition of age to the linear model of cognitive score improved the model AIC, but the addition of cognitive score to the model of age did not.The model of age alone was the best model to explain the effects related to event duration.Log-transformed mean frequency was found to decrease with age (F(1,916)=25.1,p=6.66e-7,R=0.160), but no significant effect of cognitive score was found.However, both the addition of age to the cognitive score model and cognitive score to the age model resulted in an improved model AIC, and the model containing both predictors was found to be the best fit for the mean frequency data.Log-transformed number of channels and logtransformed number of regions both showed decreases with cognitive score (channels: F(df=1,916) = 32.4,p=1.68e-8,R=0.182; regions: F(df=1,916) = 23.2,p=1.68e-6,R=0.154), but no significant changes with age.In both cases, the addition of age to the cognitive score model and the addition of cognitive score to the age model resulted in an improved model AIC, and the model containing both predictors was the best fit for the data.The relationship between event spread and age and cognitive performance is further explored in supplemental Figure S5.
J o u r n a l P r e -p r o o f

Spatial Distribution of PSWEs
The spatial distribution of PSWEs is shown in Figure 10.A Kruskal-Wallis test comparing the number of channels affected by PSWEs in each spatial region revealed that, for both gradiometers and magnetometers, there were significant differences between regions (gradiometers: H(df=7) = 949, p<2.2e-16, effect size=0.128; magnetometers: H(df=7) = 727, p<2.2e-16, effect size=0.0981).A diagrammatic representation of the significant differences between brain regions is shown in Figure 10c.In particular, for gradiometer events, the number of affected channels in the temporal regions (left and right) were significantly higher than all other regions, and the number of affected channels in the parietal regions (left and right) were significantly lower than all other regions.The number of affected channels in the frontal regions were significantly greater than in the occipital regions.There were no significant differences between the left and right hemispheres with the exception of the occipital region where the right hemisphere had a greater number of affected channels than the left hemisphere.For magnetometer events, the results were quite similar with a few exceptions.Firstly, for magnetometers, there were no significant differences between the number of channels affected in the frontal and occipital regions.And secondly, while there was no significant difference between left and right occipital regions, there was a significant difference between left and right frontal regions.
J o u r n a l P r e -p r o o f

Discussion
This work provides the first characterisation of the presence and prevalence of PSWEs in a healthy human population.In this work, PSWEs were identified in MEG recordings from a subset (~20%) of healthy participants.This finding demonstrates for the first time, that PSWEs can be found in some individuals in the absence of a known neuropathological diagnosis and that the prevalence of this purportedly pathological activity in persons without a diagnosis is as high as 1 in 5.The presence and prevalence of PSWEs in healthy participants was found to be related to the age and cognitive performance of participants such that older participants and those with lower cognitive scores tend to have more time spent in a PSWE state.In addition, it was found that PSWE characteristics including event duration and event frequency changed linearly with age, resulting in longer, and slower events in older adults.On the other hand, lower cognitive performance was found to be related to more widespread events (i.e., higher number of channels and regions affected).This work provides unique insight into the transient nature of low-frequency brain activity.While the debate regarding whether a transient or continuous analysis of neurophysiological data is most appropriate is still unresolved, this work provides a novel, complementary perspective on slow wave activity in a healthy cohort by J o u r n a l P r e -p r o o f highlighting the characteristics of transient events that are most relevant to the age and cognitive score.Finally, this work provides the first sensor-level evidence for potential sources of PSWEs in healthy ageing.
The findings of this work have several implications for our understanding of changes in cortical network activity during ageing.Importantly, the findings suggest that there may be a meaningful difference in the neurophysiology of a subset of participants leading to the emergence of atypical brain activity.Participants with PSWEs tend to be older and have lower cognitive performance than those without PSWEs, but it remains unclear whether the presence of these events can be attributed to normal variability in non-pathological ageing, or if the presence of PSWEs suggests underlying undiagnosed pathology in this subset.Previous studies suggest that increased low-frequency brain activity is more indicative of pathology than healthy ageing.While increased low-frequency activity is commonly observed in age-related pathology such as dementia and AD (Brenner et al., 1986;Hier et al., 1991;Jeong, 2004;Musaeus et al., 2018;Penttilä et al., 1985;Weiner and Schuster, 1956), studies of healthy older adult populations have often observed an opposite effect whereby low-frequency activity actually decreases with age (Emek-Savaş et al., 2016;Leirer et al., 2011;Meghdadi et al., 2021;Vlahou et al., 2015).Interestingly, when relating our PSWE prevalence findings to average band power, we found an interaction effect between PSWE prevalence and age whereby participants with more PSWEs showed an increasing trend in band power with age, while participants with low or no PSWEs showed a decreasing trend in band power with age.In the context of previous agerelated band power findings, this suggests that prevalent PSWEs are likely not characteristic of a healthy ageing brain and rather may be indicative of underlying cortical dysfunction.This observation suggests that PSWEs may have utility as a biomarker for atypical brain ageing.However, further work is required to elucidate the clinical applicability of this marker.
This work provides the first evidence for an association between PSWEs and cognitive processes in healthy adults.Particularly, cognitive performance was found to be related to the spatial spread of PSWEs such that those with lower cognitive scores had events that were captured on a large number of sensors and spatial regions.The relationship between overall cognitive score and PSWE prevalence was found to be primarily driven by age-related effects, however, we made an interesting observation whereby some cognitive tasks (e.g., fluid intelligence and hotel tasks) showed more prominent changes with PSWE prevalence than others.While the origin of this effect cannot be fully disentangled from the effect of age, this suggests that some cognitive processes, particularly higher order fluid intelligence and executive function processes have a stronger relationship to PSWE prevalence compared to other cognitive processes such as memory, language, and emotion recognition.Fluid Intelligence tasks are typically cognitively demanding and require the integration of sensory, language, working memory, and higher-order reasoning and decision-making processes that require activation of broadly distributed brain networks (Barbey et al., 2014).Therefore, the relationship between PSWEs and cognitive performance seems to depend on the cognitive demands of the task.This, together with the high occurrence of PSWEs in patients with AD (Milikovsky et al., 2019), suggests the need for future studies to test whether PSWE analysis can J o u r n a l P r e -p r o o f be used as a sensitive biomarker for early diagnosis of neurodegenerative brain disorders associated with cognitive decline.
The current work also provides insight into the mechanisms underlying PSWE generation using sensor-level regional analyses.Spatial topographies suggested widespread activation of large cortical networks during PSWEs, with particularly high activation of temporal sensors compared to other regions.These regional findings are in line with cognitive performance findings that showed a strong association between fluid intelligence and executive function and PSWE occurrence, since widespread sources are involved in these functions (Barbey et al., 2014).This evidence, taken together, provides some clues as to the functional relevance of PSWEs.In particular, it suggests temporal regions (i.e., temporal cortex or subcortical temporal structures such as the hippocampus), or widely distributed sources as potential sources of PSWE generation.However, it should be noted that there are limitations associated with the sensorbased spatial analysis conducted in this work.In particular, a region-based approach was used to index spatial spread, which provides only a rough estimate of underlying source activity.The sensor/region-based approach is limited due to variability in the position of the sensors relative to different participants' heads, and the complexity of the inverse problem.While we generally expect that sensors strongly activated by an event are close to the source of the event, it is impossible to confirm this without source estimation.Therefore, further studies using source localization approaches and/ or intraoperative recordings are required to test the spatial hypotheses formed in this work.
One logical approach to investigating PSWEs at the source level is by applying a transient burst localization method as previously proposed (Power and Bardouille, 2021).This method was attempted by our group and the results are included in the supplemental material (see Figure S6).However, the source estimation resulted in high activity primarily in ventral regions of the cortex.Because MEG source estimation generally has low accuracy for ventral and/or deep sources, we were not confident in the accuracy of these activation maps.In addition, MEG source localization is limited for this application because the PSWEs are widely distributed across cortical regions, as is evident from Figure 10.The MEG imaging modality is unreliable in its detection of deep and diffuse cortical sources because most source localization methods model a source as a focal (i.e., <10mm 2 ) cortical patch (Hansen et al., 2010).Due to these limitations, future work is required before definitive conclusions about PSWE sources can be made.To mitigate these limitations, future work could consider resting-state fMRI or multimodal imaging (e.g., simultaneous EEG-fMRI) results from participants known to have PSWEs.The differences between resting-state brain activity of participants with and without PSWEs could be investigated to identify the presence of any neurophysiological correlates of the PSWE signal.
While this work has provided promising results towards the identification of a potentially relevant marker of atypical brain function, there are some potential confounds of the work that should be addressed.Firstly, the significant correlation between age and cognitive score (p<2.2e-16,R=0.524) is a confound to our interpretation of the age and cognitive performancerelated effects.We attempted to investigate the relative contributions of age and cognitive J o u r n a l P r e -p r o o f score to the model using stepwise regression.This suggested that age was the primary predictor of most of the PSWE effects, with the exception of spatial spread metrics which showed primary effects with cognitive performance.Despite our attempts, it is difficult to fully disentangle these highly related predictors, and this should be acknowledged when interpreting the results of this work.
In addition, given the high power, transient nature of the signal of interest, it is possible that some relevant signal was excluded during pre-processing steps.In particular, the exclusion of independent components and the removal of data segments corresponding to head movements may have inadvertently resulted in the loss of PSWE segments which could have confounded our results.Several approaches were applied to investigate this potential confound, revealing evidence both for and against number of excluded components as a confounding factor in our analyses.On one hand, there were significant negative relationships between age and the number of excluded components (p=0.00119) and time in PSWEs and number of excluded components (p=0.00512)suggesting a potential confounding effect; and the inclusion of number of excluded components in the model of age by time in PSWEs resulted in a reduction in the effect of age (p=0.102).These relationships suggest that number of excluded components is a potential confounding variable in the relationship between age and time in PSWEs.However, these analyses do not provide insight into the more relevant question of whether the number of excluded components is responsible for the observed relationship between age and time in PSWEs.In order to provide insight into this, we conducted two additional analyses.Firstly, we implemented the procedure described by Judd & Kenny (1981) to test for the indirect effect of the number of excluded components as a mediator variable to the relationship between age and time in events.Using this procedure, a coefficient of indirect effect was calculated for the number of excluded components and a bootstrap confidence interval with 1000 repetitions was created for the coefficient of indirect effect.This analysis revealed that the number of excluded independent components mediator variable had a coefficient of indirect effect of 0.00134 with a confidence interval of (-0.0041, 0.0069), suggesting that the number of excluded components was not a significant mediator of the relationship between age and time in events.In addition, the data contained within the excluded components were analysed to determine whether the age and time in events relationship differed systematically compared to the included components, which would indicate a relevant bias in the exclusion procedure.The age-related effects were found to be consistent across included and excluded components, suggesting that dropping components did not alter the relationship between age and time in PSWEs.Taken as a whole, these tests suggest that the role of excluded components as a potential confound is indeterminate; our analyses suggest that the number of excluded components is not solely responsible for driving the interesting effects described in this work.Our findings highlight an important consideration for future large data projects and suggest that artefact rejection with methods such as ICA requires cautious investigation of excluded data.As we have done, excluded components should be interrogated carefully in order to obtain fulsome and balanced insight into their relationship to identified trends.
J o u r n a l P r e -p r o o f In addition, there are some avenues for future research that could further improve our understanding of the presence and prevalence of PSWEs.(Milikovsky et al., 2019).The PSWEs were detected in this work using identical methods to those described in the original characterisation of PSWEs by Milikovsky and colleagues (2019) to allow for comparison to previous results.However, it is possible that the constraints applied to the PSWE signal are not capturing the entire nature of the relevant signal in this cohort of healthy adults.For this reason, future studies should consider alternative methods of transient signal detection that make fewer a-priori assumptions about the temporal features of the signal of interest.For example, a method such as convolutional dictionary learning (Dupré la Tour et al., 2018;Power et al., 2023) could be used on this dataset to naively detect transient electrophysiological markers of ageing.

Conclusion
This work provides unique insight into electrophysiological characteristics of the ageing brain by utilizing a large open-access dataset and a transient event framework.We suggest that PSWEs could become a biomarker for the detection of atypical brain activity in ageing.The identification of such a biomarker that is easily detectable through fast, non-invasive measurements is particularly desirable for rapid assessment of cognitive dysfunction which would normally require significant time and human resources to complete.These findings have important implications for the future of the field of cognitive ageing research as they provide new information on pathophysiological mechanisms underlying the ageing brain and a promising target for therapy.

None
J o u r n a l P r e -p r o o f

Figure 1 .
Figure 1.Methods for detecting PSWEs in MEG data.PSWEs are characterised by their frequency, duration and channel spread.

Figure 2 .
Figure 2. Median power frequency plots, spectrograms, and raw data traces for a representative participant with PSWEs (left) and an age-matched participant with no PSWEs (right).The top panels show the MPF for all channels over time, the second-row panels show the MPF for a single channel (MEG1421) over time, the third-row panels show the spectrograms for channel MEG1421 where colour represents the magnitude of activity in dB, and the bottom panels show the raw data for channel MEG1421.The horizontal dashed lines show the 6 Hz frequency cut-off, and the vertical dashed lines indicate the onset of each detected PSWE.Note that not all indicated PSWEs were captured by MEG1421.

Figure 4 .
Figure 4. Power spectral density plots in dB for all gradiometers (left) and magnetometers (right) averaged across groups of participants with PSWEs (blue), and no PSWES (orange).The solid lines represent the group mean and the shaded region represents the 95% confidence interval.

Figure 5 .
Figure 5. Distributions comparing participants with PSWEs (light grey) and those with no PSWEs (dark grey) Groups are scaled to the same total density for comparison.Ridge plots show the distributions of participant age (A), aggregate Z scores of cognitive performance (B), and ACE-R scores of cognitive performance (C).A stacked bar plot is used to compare the distributions of males and females in the two groups (D).

Figure 6 .
Figure 6.The relationship between PSWE prevalence and participant age/aggregate cognitive score.(A) Linear regression relating age to log-transformed time in events values.The solid black line is the best fit regression for this data.(B) Violin plots showing the age distribution of participants with no PSWEs (light grey), less PSWEs (medium grey), and more PSWEs (dark grey).(C) Linear regression relating cognitive score to log-transformed time in events values.(D) Cognitive scores for participants in each PSWE prevalence group.Asterisks indicate a significant effect between groups.

Figure 7 .
Figure 7. Relationship between individual cognitive task scores and PSWE prevalence.Distributions of mean Z-scores for each of 8 cognitive tasks are shown for participants with no PSWEs (light grey), less PSWEs (medium grey) and more PSWEs (dark grey).Significant differences are indicated by asterisks.

Figure 9 .
Figure 9. Results of regression between PSWE characteristics and age/cognitive score.Each point represents the characteristic associated with a single PSWE and black lines show the best fit result of the regression.Significance is indicated by asterisks.

Figure 10 .
Figure 10.Spatial distribution of gradiometer and magnetometer PSWEs.(A) Number of channels in each spatial region for magnetometer (light grey) and gradiometer (dark grey) PSWEs.Error bars represent standard error.The 8 spatial regions indicated are left occipital (LO), right occipital (RO), left temporal (LT), right temporal (RT), left parietal (LP), right parietal (RP), left frontal (LF), and right frontal (RF).The diagram of the sensors included in each region is provided in the supplemental material.(B) Spatial topographies showing the total number of PSWEs across all participants that were detected at each channel.Plots are shown for magnetometers and gradiometers separately.(C) Visual representation of the difference in the number of active channels between spatial regions, averaged across participants.Significant differences between brain regions are indicated by arrows.Arrows point in the direction of the brain region with a greater number of active channels.Line width represents the magnitude of the difference.
Firstly, while previous research examining PSWE prevalence used EEG data, our work detected PSWEs in MEG data.Due to differing sensitivity and channel numbers between technologies it is difficult to directly compare our findings to previous work.Therefore, future research should consider comparing PSWE detection in EEG and MEG from the same cohort of participants to provide insight into the sensitivity of each to PSWE activity.Another potential limitation is that the data analysed in this work was exclusively collected as part of the Cam-CAN initiative, and therefore consisted of a limited population of participants from a limited geographic region.While the Cam-CAN dataset has ideal demographics for studying cross-sectional ageing trends, it is unclear how generalizable the results are to a wider population.For that reason, it would be beneficial to consider analysing additional open-access datasets for the presence of PSWEs to determine whether the 20% subset of participants with PSWEs is consistent across other populations.In addition, because the data used in this work is cross-sectional, we can only infer age-related trends but cannot confirm that the prevalence and characteristics of PSWEs change across the lifespan of an individual.For that reason, additional longitudinal studies of PSWEs examining the same participants several years apart would also be a useful direction for future research.Another limitation of this work is that the temporal characteristics of the PSWE signal studied in this work were pre-defined based on previous observations of low frequency transient signals in pathological populations

Table 1
Cognitive task descriptions.

Table 2 .
Summary demographics for participants with and without PSWEs and for the overall group.Means and standard deviations are given for continuous metrics.3.3PSWE Prevalence with Age and Cognitive Performance