Fluctuations in Neural Complexity During Wakefulness Relate To Conscious Level and Cognition

There has been considerable recent progress in measuring conscious level using neural complexity measures. For instance, such measures can reliably distinguish healthy awake from asleep subjects and vegetative state patients. However, this line of research has never explored the dynamics of conscious level during normal wakefulness. Being able to capture meaningful differences in conscious level during wakefulness may provide a vital new insight into the nature of consciousness, by demonstrating what biological, behavioural and cognitive factors relate to such differences. Here we take advantage of a large MEG and fMRI dataset of healthy adults, to examine within-subject conscious level fluctuations during resting state and tasks, by using a range of complexity measures. We first establish the validity of this approach in both neuroimaging domains by relating neural complexity measures to pre-existing techniques for capturing transitions of consciousness from full wakefulness into drowsiness and the earliest stages of sleep, finding decreased complexity as participants become increasingly drowsy. We further demonstrate that neural complexity measures in both MEG and fMRI change both within and between tasks, and relate to performance on an executive task, with higher complexity associated with better performance and faster reaction times. This approach provides a powerful new route to further explore the cognitive and neural underpinnings of consciousness.


Introduction
In the MEG data, we quantified the effect of drowsiness with a linear mixed-effects (LME) model predicting LZ using 48 AMA as predictor and subject identity as random effect. The results showed a very reliable effect of drowsiness on LZ, with 49 greater drowsiness associated with less complex brain activity (β = −0.0128 ± 0.0003 bit, t = −35.6). It should be noted that 50 AMA is a novel measure, not used in MEG before. Therefore we validated our results with the more traditional alpha-theta 51 ratio measure, 22 with qualitatively similar results (β = −0.0091 ± 0.0003 bit, t = −26.6). Furthermore, to demonstrate that 52 these results are not dependent on any single measure of neural complexity, we also carried out these comparisons with other 53 measures (see Supplementary Figure S1), with again broadly similar results. 54 Next, we performed a similar analysis on the fMRI data, with two main differences: First, alertness level was quantified via 55 the functional connectivity clustering method of Haimovici et al. 21 instead of AMA. Second, given the much smaller number of 56 time points in the fMRI data (5 windows of 50 TRs each), instead of using an LME we split the subjects into alert and drowsy 57 groups, labelling as alert those who had all 5 time windows classed as awake by the Haimovici algorithm (N = 352), and as 58 drowsy those who had at least 3/5 time windows classed as drowsy or early sleep (N = 74). As a sanity check, we verified 59 that the proportion of awake subjects in each window decreased with time, replicating previously reported results of subjects 60 becoming more drowsy during the scan 23 (see Supplementary Materials). 61 To relate alertness and complexity, we averaged LZc across all 5 temporal windows for each subject and performed a 62 two-sample t-test comparing the two groups. This analysis showed strong agreement with the MEG results, with drowsy 63 subjects having significantly lower complexity than alert subjects (t = −12.5, p < 0.001). We repeated this analysis for each of 64 Yeo's seven brain networks 19 in turn, and all networks showed a significant LZc reduction in drowsiness (all p < 0.001) to varying extents. In particular, the visual and somatomotor networks showed the lowest levels of resting LZc but the largest 66 reductions with drowsiness, while the frontoparietal and salience network showed the smallest changes.  Having established that LZ meaningfully fluctuates with alertness levels between full alertness and drowsiness, we next explored 69 how LZ changes by task. The CamCAN database offers an excellent opportunity to explore this, thanks to its high sample 70 numbers and wide repertoire of tasks available (see Methods). To this end, we computed whole-brain average complexity for 71 all subjects and all tasks in both MEG and fMRI sessions (averaged across channels in MEG, concatenated in fMRI), revealing 72 striking changes both within and between tasks (Fig. 2).

73
In the MEG session, participants showed a wide spectrum of LZ values, which varied markedly between tasks (see 74 Supplementary Figure S5). Of all 12 pairwise comparisons between tasks (including sets A and B), 11 of them had significant 75 LZ differences at the p < 0.001 level using one-sample t-tests, with the only non-significant comparison being between 76 multi-mismatch negativity and incidental memory in task set A. Interestingly, all active tasks (i.e. tasks that required subjects' 77 input; including SNG, picture naming, and SC) had lower complexity than wakeful rest, suggesting that more cognitively providing further evidence that contextual factors can affect subjects' spontaneous neural complexity. Note that, however, in 86 fMRI LZc is higher during task than in wakeful rest, unlike the results for the MEG data. The cause of this discrepancy is 87 unclear, but it is unlikely to be caused by the concatenation of channels in LZc since it also occurs with average channel-wise 88 LZ (see Supplementary Figure S4).

89
Finally, it is worth mentioning a few important differences in LZc between brain networks as they relate to the tasks in 90 the fMRI session. For both task and movie-watching, the visual and somatosensory networks showed the largest increases 91 with respect to rest, while the frontoparietal network showed slightly weaker task distinctions compared to the other networks.

92
Additionally, the dorsal attention network was the only one to show a positive interaction effect between task and movie-93 watching, with higher complexity during task (t = 6.77, p < 0.001; see Supplementary Figure S6). (right) Whole-brain LZc in each window of the fMRI session, showing both variation across tasks and negative drifts within tasks, suggesting subjects became progressively drowsy during each task. Note that MEG and fMRI represent different neural processes and LZc is different from LZ, so values are not directly comparable between the two. 95 The fact that these complexity measures track alertness and are modulated by task suggests they can be used to explore the 96 connection between consciousness and cognition. For this we focused on the one attentionally demanding, executive task 97 present in both fMRI and MEG task sets: the SNG task.

98
In each trial of the SNG task, subjects are presented with one of four types of visual stimuli. For two of them (go-left and 99 go-right signals) the subject must respond via button press as quickly as possible (with their left or right hand, respectively).

100
For the other two stimuli (stop and no-go signals), the subject must withhold their response and wait for the next trial. For this 101 analysis we focused on the reaction times in the go trials and the total number of omission and commission errors, regardless of 102 which stimulus type they were triggered by (see Supplementary Figure S7).

103
In the MEG session, we analysed the trial-by-trial relationship between LZ and reaction time using an LME model to 104 predict reaction time in a given trial, using the LZ of the preceding 4 s as predictor and subject identity as random effect.

105
This revealed a small but consistent negative effect (β = −533± 50 ms bit −1 , t = −9.13), showing that subjects responded 106 faster in trials with a higher level of pre-stimulus LZ. To investigate this further, we analysed whether subjects with higher 107 inter-trial LZ variability also had higher RT variability. In agreement with the previous result, we found a significant positive In this paper we have demonstrated, using both MEG and fMRI, that neural complexity measures can be used to track fluctuating 121 conscious level during wakefulness. Furthermore, these same complexity measures change for different tasks, reflecting changes 122 in the subject's condition and cognitive state. Finally, we have shown that these complexity measures correlate with accuracy 123 and reaction time in an executive task. These results suggest that conscious levels not only change from awake to sleep states, 124 but fluctuate, in line with alertness and drowsiness, on a moment by moment basis -and these spontaneous fluctuations offer us 125 an opportunity to investigate the inner workings of consciousness during wakefulness.

126
In addition to the temporal dynamical aspects captured by LZ, this type of analysis allows us to explore the spatial 127 heterogeneity of complexity throughout the brain. Although there were no differences between local regions in the MEG data, 128 in fMRI there were intriguing differences between the seven Yeo brain networks. For example, the somatomotor and visual 129 networks had lower complexity overall, both during alertness and drowsiness. In contrast, the frontoparietal network had one 130 of the highest LZc overall, demonstrated the least decline from alert to drowsy states, and was less modulated by task than 131 other networks. One explanation for this is that the frontoparietal network, by having the highest LZc generally, even sustained 132 during drowsiness, is more centrally involved in supporting conscious contents than lower-level networks involved in sensory 133 processing or motor output (which would be closely in accord with the existing literature 11, 13 ).

134
Connecting neural complexity and cognitive processing 135 On the cognitive level, our results demonstrate that complexity measures are predictive of behaviour and cognitive performance. 136 We demonstrated this in fMRI using broad temporal averages, as well as in MEG with a trial-by-trial analysis, showing in 137 both cases that higher neural complexity was associated with better performance (in terms of fewer errors and faster reaction 138 times). Although the posited association between executive processing and consciousness is not new, 24 here we've provided, to 139 our knowledge, the first neurally-driven evidence supporting its connection, albeit in a provisional form we hope that we, and 140 others, will build on in future studies.

141
Interestingly, however, there is one important aspect for which the MEG and fMRI results disagree: in the MEG session LZ 142 was lower for all tasks than during resting state, while the opposite was true in fMRI. To the best of our knowledge this is the 143 first reported discrepancy between LZ changes across multiple imaging modalities, with previous works showing consistent 144 results in fMRI, MEG, and EEG. 2, 4, 5, 25 However, it is worth noting that these reported consistent changes all concern drastic 145 changes in conscious state, such as between wakeful rest and deep sleep or anaesthesia. Thus, the fine-grained structure of the  Nonetheless, there are also a few consistent patterns across fMRI and both MEG sessions that are worth mentioning. Most 149 notably, both fMRI and MEG passive tasks (i.e. multi-mismatch, word recognition, and movie watching) induce consistently 150 higher complexity than active tasks. This could be caused by these tasks calling on different underlying cognitive mechanisms, 151 although without further experiments we cannot discard an effect of salience driving the LZ results: for example, the word 152 recognition task in MEG had unexpected non-words, and the movie shown in the fMRI session was engaging and fast-paced, 153 while the active tasks were generally repetitive -which could explain the differences in complexity between the two.

154
Overall, although future work is needed to explain these results, we can attempt to interpret them within the framework 155 of the Entropic Brain Hypothesis (EBH). 26 Put briefly, the EBH states that the richness of phenomenal experience should be 156 accompanied by a similarly rich repertoire of neural dynamics, quantifiable via complexity or entropy measures. 27 If, as the 157 EBH postulates, LZ essentially captures the variety of phenomenological events, and given the cognitive tasks were relatively 158 prescribed and uniform compared to resting state, it is natural that entropy is higher in resting state. At the same time, this 159 variety of phenomenology is more likely to be captured by the higher temporal resolution of MEG, but not fMRI, which only 160 captures slower neural processes due to the timescale of the haemodynamic response 28 . Overall, this discrepancy highlights 161 the need for future work exploring the profile of neural complexity across scales, from aggregated cellular activity to BOLD 162 signals, and paints a more nuanced picture regarding the interpretation and physiological relevance of LZ deserving further 163 investigation.

164
Limitations and future work 165 Our approach in this study has been openly empirically-driven, applying an experimentally validated set of measures to a 166 large neuroimaging dataset and analysing the results. Nonetheless, like previous studies linking complexity measures to 167 consciousness, 1-3 a theoretical foundation behind these analyses is largely underspecified. For instance, it is not clear how 168 LZ (and its variants) map onto elements of candidate theories of consciousness, such as the differentiation or integration 169 components of IIT 1 -or, for that matter, whether the different measures proposed historically by IIT can in fact meaningfully 170 1 Although there is some preliminary evidence in toy logic-gate models, 29 it is unclear how or to what extent these results apply to whole-brain dynamics.

5/10
assess integration and differentiation in brain dynamics, 30, 31 and how these relate to cognition. 32 Therefore, we need better theories to fully interpret these results and incorporate them into a broader picture of consciousness as a neurobiological phenomenon. In this sense, we see these results as a more challenging testbed for upcoming theories of consciousness, that are 173 willing to go beyond gross changes in conscious level (such as between wakefulness and sleep or coma) and able to explain 174 these subtle fluctuations of consciousness during wakefulness.

175
One concern with the analyses presented here is whether there may be some specific details of LZ that generate these results. For these analyses we leveraged the power of big data via the CamCAN 18, 41 database, which allowed us to resolve subtle 194 changes in the measures of interest thanks to its large size, and make first steps towards elucidating the components of 195 consciousness. We believe that this will be a powerful, flexible new approach to accelerate progress in consciousness science 196 generally.

197
Although the results presented here provide intriguing putative clues as to the components of consciousness, we believe that 198 their main purpose is to demonstrate the utility of this approach for future, more directed, studies than was possible with the 199 CamCAN dataset. We envisage these future studies to focus on fluctuations in complexity measures during wakefulness and 200 these changes to be linked to specific cognitive components. In this way, a taxonomy of the potential cognitive machinery of 201 conscious could be found.

203
Participants 204 We examined a pre-existing dataset, collected by the Cambridge Centre for Ageing and Neuroscience (CamCAN) project, 205 involving adults who had undergone both fMRI and MEG at two stages, approximately 2 years apart. In the first stage, here Ratio; ATR), a robust and empirically well-established marker of drowsiness in eyes-closed EEG. 22, 43 . We took theta power in 245 the 3-5 Hz range, alpha power in the 8-12 Hz range, and calculated the ratio between these two frequency bands as the mean of 246 all MEG gradiometer sensors per epoch. More drowsy epochs were those where alpha power was reduced and theta power was 247 increased.

248
In terms of neural complexity measures, in line with recent literature [1][2][3][4] we focused on Lempel-Ziv complexity (LZ) 44 as a 249 measure of neural signal diversity. In short, the procedure to estimate LZ from a time series of neural activity of length T is as 250 follows: first, the signal is binarised around its median. Then it is scanned sequentially using the algorithm by Kaspar and 251 Schuster, 45 counting the number of different "patterns" in the signal. Finally, following Ziv 46 this number is normalised by 252 log 2 (T )/T to yield an estimate of the signal's entropy rate. 47 This process is repeated for all channels and the results averaged 253 into a single whole-brain average LZ.

254
Although our focus here is on using LZ, other suitable complexity measures exist, some of which have already been 255 successfully applied to distinguish between conscious levels. 3, 48, 49 Therefore, to demonstrate the robustness of our findings k-means algorithm to cluster subjects' functional connectivity into "awake" and "sleep" clusters. For our analysis, we took the 264 cluster centroids from the Haimovici study in order to classify the CamCAN data into high-and low-alertness segments (of 265 100 s each), roughly equating to the "awake" and "drowsy" micromeasures classification in the MEG dataset. 4

266
To measure complexity we again focus on LZ complexity, for consistency with the MEG analysis. It is worth noting that 267 although fMRI has less temporal resolution than MEG, recent studies 25 have shown that fMRI LZ is as robust of an index of 268 conscious level as it is in other modalities. Nevertheless, given the short length of fMRI BOLD time series (compared to MEG), 269 4 Note that the Haimovici study labelled the non-alert state as "sleep" -however, given that in the CamCAN650 dataset subjects still responded regularly on the sensorimotor task when classed in this state, we believe that "drowsy" is a more appropriate label than sleep. See Supplementary Material for details. instead of computing LZ on each region separately, we concatenated all time series into a single one-dimensional signal, and then computed and normalised LZ as described in the MEG section, resulting in a "concatenated LZ" (LZc). This procedure 271 was performed for all parcels in the Schaefer 300-region atlas, as well as in those subsets of parcels that belong to each of Yeo's 272 seven networks. 19  Components Analysis (ICA) was used to reduce the effects of eye blinks. The data was split into gradiometer and magnetometer 278 channels, and only the 204 gradiometer channels were selected for further preprocessing, due to superior signal-to-noise ratios.

279
Following this, the data was downsampled to 250 Hz, and filtered with a 0.5-30 Hz bandpass filter. Data were epoched to 4 s 280 segments in both RS and task recordings. For a further analysis relating task performance with complexity at the trial level, we 281 focused on the only CamCAN executive task (SNG), and extracted 1 s epochs immediately prior to each trial. Epochs with 282 muscle artefacts (widespread frequency above 120 Hz across channels) or unusually reduced signal (less than 25% of average 283 signal compared with the rest of the session) were excluded from the analysis.