Behavioral Measures of Word Visibility
Unmasked words were consciously reportable, and were categorized better than chance level in a forced-choice categorization task on the emotional valence of words (mean discriminability index d′ = 2.24 (+1.14 to +3.04), all individual χ2 p-values and group analysis Student t-test p-value < 0.0001). In sharp contrast, masked words were not consciously visible, and forced-choice performance was at chance level for each of the implanted patients (mean d′ = 0.02 (−0.18 to +0.27), all p-values >0.2). Response times (RTs) were similar across the two masking conditions (p > 0.38 in Student t-test performed on mean RTs; masked mean RT = 1,640 ms, unmasked mean RT = 1,300 ms).
We defined masked effects by subtracting the voltages measured on masked blank trials from those associated with masked word trials. This subtraction allowed us to isolate, on a sample-by-sample basis, activations associated with masked word processing (see and Materials and Methods
for our detailed three-step statistical procedure). Unmasked effects were defined similarly by subtraction of the unmasked word and unmasked blank conditions.
shows the anatomical distribution of the 176 reconstructed bipolar montages (“electrodes”) from which we obtained valid data across the ten patients. The bipolar subtraction of nearby recording sites reduced distant influences, including those from the reference electrode, and resulted in a signal tightly localized to the implanted structure. Although measures were obtained for all four lobes, it should be kept in mind that major sectors of dorsolateral and polar prefrontal cortex as well as parietal cortex were not sampled.
Among the 176 electrodes, 24.4% (43 electrodes) showed a significant effect for masked words. These effects were observed across all implanted structures but with a predominance of effects on occipital electrodes: 22/55 (40%) within the occipital lobe, 11/78 (14.1%) within the temporal lobe, 4/24 (16.7%) within the parietal lobe and 6/19 (31.6%) within the frontal lobe (χ2 p-value = 0.004).
Concerning unmasked words, 68.8% of all electrodes (121 electrodes) showed a significant effect of word presence—a remarkably high percentage, given that electrodes had been placed at clinically relevant sites without consideration of their relevance to our visual stimuli. Unmasked effects were observed across all implanted structures but with a particular emphasis on the frontal lobe: 42/55 (76.4%) within the occipital lobe, 49/78 (62.8%) within the temporal lobe, 12/24 (50%) with the parietal lobe, and 18/19 (94.7%) within the frontal lobe (χ2 p-value = 0.005). The frontal lobe showed a major difference between trials containing masked and unmasked words: almost all contacts were systematically activated during conscious processing of unmasked words (~95%), whereas this was not the case during unconscious processing of masked words (~32%).
In order to better assess the specificity of this last result, we ran an ANOVA to directly compare the impact of masking on the proportion of activated electrodes between occipital and frontal lobes. A main effect of masking was observed (86% versus 36%; p < 10−4), confirming the larger spatial extension of unmasked activations as compared to masked activations. No main effect was observed between frontal and occipital electrodes (58% versus 63%, p > 0.5). Crucially, we observed a significant interaction between frontal and occipital cortices and masking condition (p = 0.05), assessing the larger differential activation of frontal lobe between masked and unmasked conditions, as compared to the pattern observed in posterior visual cortex.
Note that this spatial analysis is affected by the nonhomogenous sampling of brain regions, minimizing the contribution of cortical structures that were less frequently implanted. Nevertheless, masked effects were more frequent on posterior than on anterior electrodes, whereas unmasked effects were homogeneously distributed. To demonstrate this point, we examined the distribution of the anterior-posterior (y
) coordinate, in Talairach space, of the electrodes showing a significant effect, and compared it to the spatial distribution of all 176 recorded electrodes (see Figure S1
). For masked words, the spatial distribution of significant electrodes was strongly shifted towards posterior sites (p
, Kolmogorov-Smirnov test, relative to the distribution of either the whole set of 176 electrodes or to those showing an unmasked effect). The same analysis conducted on the cumulative distribution of unmasked effects showed a spatial distribution statistically indistinguishable from that of the whole set of electrodes.
Masked and unmasked words were also distinguished by the temporal extension of their activation. A crude analysis, averaging across all electrodes, revealed that masked effects had a mean duration of 60 ms, much shorter than the mean of 378 ms for unmasked effects (p
). Masked effects also showed an earlier onset (mean = 366 ms; median = 301 ms) than unmasked effects (mean = 522 ms; median = 497 ms; t
). A more relevant analysis focusing on the first significant effect within the subset of electrodes with both masked and unmasked effects showed similar latencies between these two conditions (299 ms and 348 ms, respectively, for masked and unmasked conditions; t
= 0.30). Indeed, up to approximately 200 ms after word onset, glass brain visualization of the spatiotemporal dynamics of masked and unmasked effects showed a strikingly similar pattern of activations within posterior occipitotemporal cortical regions (see ). This dynamic pattern is very comparable to the “feedforward sweep” described by Lamme in the nonhuman primate visual cortex using multiunit recordings for latencies up to 100 ms after visual stimulus onset [5
]. Clear differences between masked and unmasked effects appeared after 150 ms, with a progressive increase in the intensity and spatial extension of unmasked effects, while masked effects decayed and did not show a similar spatial extension (see Videos S1
Spatiotemporal Dynamics of iERP Effects
This general pattern was observed on individual electrodes (see ). The initial response was often indistinguishable between masked and unmasked effects. This initial common response was usually followed by later effects specifically for the unmasked condition. Out of 14 electrodes showing this pattern with our statistical criteria, 11 of them also showed a polarity inversion of the late sustained effects relative to the polarity of the initial effect.
iERP Effects on Three Representative Electrodes
A cortical lobe analysis focusing on the proportion of electrodes showing a significant effect over time () showed a similar proportion of electrodes activated by masked and unmasked words at short latencies, whereas at later latencies, the effects were increasingly specific to the unmasked condition. An analysis of the mean voltage power, averaged across electrodes within one lobe, showed a similar temporal dynamics, and additionally allowed us to detect a progressive time delay in the peak of the initial activation common to masked and unmasked words. The time point at which the first significant divergence between masked and unmasked effects occurred, as estimated by a t-test (p < 0.05), progressively increased from 215 ms to 275 ms and 347 ms, respectively, for the occipital, temporal, and frontal lobes (see ). The divergence did not reach significance for the small set of 14 parietal lobe electrodes tested (those showing at least one significant effect, masked or unmasked).
Event-Related Spectral Perturbations
We then turned to frequency-domain analyses of the intracranial signals. A shows a typical single-electrode example of the time-frequency transform applied to our data. The masks alone evoked a very strong sequence of event-related increase in the beta and gamma bands accompanied by alpha decrease, followed by a reversal of this pattern. Subtraction of each mask-only condition from the corresponding word-present condition, however, isolated the ERSP induced by the word alone, as a function of whether it was masked or unmasked. As can be seen in this example, masked words induced a slight increase in gamma power 100–200 ms after the stimulus, whereas unmasked words induced a much bigger effect that lasted throughout the epoch and was accompanied by alpha suppression.
To evaluate the generality and significance of such effects, we averaged the time-frequency diagrams across all electrodes (B). Statistical comparisons over time-frequency regions of interest, with Bonferroni correction (see Methods and Materials
), identified several significant effects. In the 100–200-ms time window, masked words evoked highly significant power changes (beta suppression: p
= 0.0004; high-gamma increase: p
= 0.0005). In this time period, there was no significant difference with unmasked words, confirming that a volley of activation, reflected primarily in a gamma increase, can propagate nonconsciously while being largely unaffected by masking [7
In the next time window (200–300 ms), whereas unmasked words created an even larger power increase in the high-gamma band (p < 10−11) and decreases in alpha (p < 10−8) and beta bands (p = 10−5), masked words induced only small effects of alpha suppression (p = 0.0014) and high-gamma increase (p = 0.038). Beta and high-gamma bands showed significantly stronger changes for unmasked compared to masked words (all p < 0.0007). In the subsequent time window (300–500), alpha suppression, beta suppression, and high-gamma increase were very strong in the unmasked condition (all p < 0.0003), but altogether absent in the masked condition, creating a significant difference (all p < 0.0002).
In summary, masked words induced significant changes in the power spectrum, particularly increases in the high-gamma band, but these induced oscillations quickly dissipated with time, whereas the ERSPs evoked by unmasked words exhibited a greater power and lasted significantly longer. Note that the above analysis was based on the pooling of ERSP results from all electrodes, regardless of their location. We also replicated this ERSP analysis while separating the electrodes as a function of their lobe of origin. Both the high-gamma increase and late alpha and beta suppression specific to the unmasked condition were replicated within each of the four lobes (see Figure S2
). Interestingly, the high-gamma increase peaked earlier in occipital cortex than in temporal, parietal, or frontal, following an approximate posterior to anterior progression (see Videos S3
). Furthermore, the lobar analysis showed that the early nonconscious effects were confined to the occipital and temporal lobes: the only significant effects were a high-gamma power increase in occipital cortex in the 100–200-ms and 200–300-ms windows (respectively, p
= 0.013 and p
= 0.016), and decreases in alpha (200–300 ms, p
= 0.007) and beta (100–200 ms, p
) in temporal cortex. In brief, the early ERPS evoked by nonconscious stimuli originated only from occipitotemporal regions, whereas conscious perception was associated with stronger and longer-lasting power changes spreading towards anterior cortical regions.
In this respect, analyses of induced high-gamma power yielded conclusions very similar to those derived from iERP analyses. To better evaluate the relation between induced gamma activity and iERPs, we calculated for each segment for which a significant ERP difference was seen, the absolute value of the iERP effect as well as the mean power evoked in the high-gamma range, averaged over the same time period. A significant positive correlation between these two measures was found, for unmasked words (r2
= 14.9%, 241 time segments, p
) and, crucially, for masked words (r2
= 9.3%, 59 time segments, p
= 0.028). This means that periods in which a high-gamma–band activity is seen are also periods in which a high-voltage difference between word-present and word-absent conditions exists. In brief, high-gamma activity and iERP are correlated measures that both jointly reflect conscious as well as nonconscious processing stages. Indeed, videos of ERP and of high-gamma activity, provided as supplementary online material, present high similar profiles (see Videos S1
, and S4
Spectral changes are complex phenomena that can be sensitive to local as well as global neuronal synchronization of thalamocortical networks [56
]. To evaluate the global workspace model's prediction that access to consciousness is associated with long-distance synchronization, we measured the phase synchrony between all electrode pairs. Phase synchrony can occur independently of changes in induced power: it solely evaluates whether oscillations are reproducibly synchronized across two distant sites in the sense that across trials, they exhibit a systematic phase relationship.
shows a time-frequency diagram of the intertrial phase coherence (ITC) changes induced by the masked and unmasked words, both in an example electrode and in the mean overall electrodes. All statistics were Bonferroni corrected. Statistical analyses revealed no significant coherence changes induced by the masked words. For unmasked words, an increase in beta synchrony in the 300–500-ms time window was highly significant (t(1,282 df) = 7.12, p < 10−10; difference with masked condition, t(1,282 df) = 5.43, p < 10−6). It is particularly interesting to note that (1) this phase synchrony increase was concomitant with a decrease in induced spectral power (ERSP) within the same frequency band (see A and B); (2) no phase synchrony increase was detected in the high-gamma band, although in this band, a highly significant increase in induced power had been detected by ERSP analysis. Thus, ERSP and phase synchrony appear to double dissociate, and beta synchrony appears as a highly selective marker of the late phase of conscious access.
ERSP and Phase Synchrony across Three Time Windows
shows in graphic form the value of the beta coherence increase due to word presence in the critical time window 300–500 ms, separately for unmasked and masked words. Clearly, unmasked words create a more globally synchronous brain state than masked words. The figure makes apparent that this phase coherence analysis is importantly limited by the available electrodes: we can only analyze coherences between electrodes within a given patient, and these tend to be regrouped within a cortical area, thus preventing a thorough analysis of how coherence evolves across distant anatomical sites. For instance, it was not possible to evaluate the prediction that frontal electrodes should cohere more with posterior sites during conscious processing, because in our sample, these two regions were very rarely recorded simultaneously. Still, to probe long-distance connections, we could analyze a subset consisting of electrode pairs in which the two electrodes lie in different hemispheres, thus imposing a long-distance transfer across the corpus callosum. As predicted by global workspace theory, we observed an increase in long-distance interhemispheric beta coherence selective to unmasked words (t(71) = 2.50, uncorrected p = 0.015). In fact, interhemispheric beta coherence actually decreased when masked words were presented (t(71) = 3.14, uncorrected p = 0.003), thus creating a strong difference between visible and invisible conditions (t(71) = 3.83, p = 0.0003).
Phase Synchrony and Granger Causal Gain between 300 and 500 ms after Word Onset
Conversely, suggests that in the masked condition, there might have been a small local increase in beta coherence within posterior occipitotemporal cortices, which might have been missed in our analysis pooling across all electrode pairs. Indeed, when we restricted only to intrahemispheric electrodes lying within occipital cortex or within temporal cortex posterior to y = −20, we detected a significant increase in beta coherence for masked words during the 200–300-ms time window (t(734) = 2.34, uncorrected p = 0.02), which ceased to be significant in the 300–500-ms time window (t = 0.41, not significant [n.s.]). No such increase was seen in other frequency bands, or in other regions (e.g., within frontal electrodes). Thus, nonconscious word processing resulted in only small and barely detectable transient increases in phase coherence within visual cortex, whereas conscious words yielded a massive increase in long-distance beta coherence
A final measure of conscious processing that we evaluated is Granger causality [60
], a mathematical tool that can estimate the causal influence that one electrode site exerts on another. Global neuronal workspace theory predicted that access to consciousness for unmasked words would be accompanied by a massive web of causal relations among distant cortical sites, not seen in the masked condition. Granger causality and phase coherence are similar in that both estimate the correlations among pairs of electrodes, but Granger causality looks for temporal contingencies inaccessible to coherence analyses. In a nutshell, the method estimates whether past samples of electrode j
account for a significant amount of variance in electrode i
, over and above a simpler “autoregressive” model using only past samples of electrode i
] for details). It is possible for two time series to be strongly phase coherent, yet not causally related (for instance, two sine waves with constant phase lag and independent noise). Thus, Granger causality analysis is not redundant with phase coherence analysis: finding that Granger causality increases during conscious perception, perhaps simultaneously with the beta coherence increase, would provide additional evidence in favor of a large-scale reverberating neuronal assembly linking distant sites. Furthermore, unlike phase coherence, Granger causality is a directional measure: it is possible for electrode j
to causally influence i
causally influencing j
(although it is also possible for two signals to exert mutual causal influences on each other). This analysis therefore provided an opportunity to examine the top-down versus bottom-up propagation of activation during conscious and nonconscious processing.
As a concrete example, illustrates the causality analysis of a sample electrode pair consisting of one frontal and one occipital electrode. At the time of stimulus presentation, a massive increase in Granger causality is seen in the feedforward, occipitofrontal direction (A, left panel) and, to a smaller extent in the top-down, fronto-occipital direction (A, right panel). Importantly, the curves showing the evolution of the F-test for causality as a function of time exhibit two successive peaks: one early peak is evoked by the masks alone (146 ms after mask onset), whereas a second peak (325 ms after word onset) is seen only when a word is present and unmasked. As illustrated in B, a strong “causal gain” is therefore observed, approximately 200–450 ms after word onset, when the word-present condition is contrasted to the word-absent condition. This effect is seen mostly in the feedforward direction, thus engendering a “causal imbalance” (higher causal gain in one direction than in the other).
Similar increases in causal gain and causal imbalance in the unmasked condition were seen in a large set of electrode pairs. To evaluate their statistical significance, we first averaged the causal gains across all electrode pairs and both causal directions, separately for masked and unmasked conditions, and used t-tests to evaluate the significance of changes within three temporal windows (100–200, 200–300, and 300–500 ms, similar to the ERSP and phase coherence analyses, with Bonferroni correction over the number of windows tested) (see C). A massive increase in mean causal gain was observed during the 300–500-ms window in the unmasked condition (t(1805) = 7.60, p < 10−13), but not in the masked condition (t = 1.47, n.s.), resulting in a significantly larger causal gain during conscious than during nonconscious processing (t(1805) = 5.46, p < 10−8). The effect was already perceptible in the 200–300-ms window, though it was much smaller (unmasked: t(1805) = 3.03, p = 0.0075; masked, t = −0.86, n.s.; difference: t(1805) = 2.89, p = 0.012). No effect reached significance in the 100–200-ms window.
B illustrates the anatomical distribution of the mean causal gains during the 300–500-ms window. In the unmasked condition, causal relations increased massively among many distant sites, both within the occipitotemporal cortex, between occipitotemporal cortex and distant frontal and insular sites, and across the corpus callosum. By contrast, increases were very scarce in the masked condition and never reached significance even when restricted to posterior electrodes only.
With similar methods, we evaluated the statistical significance of changes in the variable of “causal imbalance,” which is the subtraction of forward causal gain minus backward causal gain in the same electrode pair. This variable evaluated the dominant directionality of causality (posterior to anterior = feedforward, or anterior to posterior = feedback). During the 300–500-ms time window, in the unmasked condition, there was a small imbalance with a higher causality gain in the feedforward compared to the feedback direction (t(1,850) = 2.07, p = 0.039 before Bonferroni correction). Although marginally significant, this finding occurred in the predicted late time window and fits with our prior hypothesis that during this time period, and only in the unmasked condition, perceptual information gains access to consciousness and is therefore able to invade anterior areas in a feedforward manner. Indeed, in this time window, the imbalance was not significant for masked targets (t = −0.90, n.s.), creating a difference for unmasked as opposed to masked targets (t(1,805) = 2.09, p = 0.037 before Bonferroni correction).
Quite surprisingly, however, in the preceding time window (200–300 ms, see D and S3
), there was a significant imbalance in the converse direction (higher causality in the top-down or feedback). This was true only for the masked condition (t
(1,805) = −2.66, p
= 0.024, Bonferroni corrected), not the unmasked condition (t
= 1.11, n.s.), a significant difference (t
(1,805) = 2.70, p
= 0.021 Bonferroni corrected). This unexpected finding, further discussed below, may indicate that in the masked condition, there is a top-down component of attentional amplification, perhaps relating to an unsuccessful effort to identify the masked word.