Article Figures & Data
Figures
- Figure 1.
Illustration of artificial datasets creation for the simulation. The artificial signal was initialized using WN drawn from a Gaussian distribution with μ = 0 and σ = 10 arbitrary units. These random data were then bandpass filtered at the frequency of interest plus or minus 1 Hz, and a Hilbert transform was applied to extract the phase at 40 ms after time 0, the time of target presentation. The phase angle at this time was then used to separate the trials between outcome A and B, with a given probability following a cosine function. Finally, an outcome independent ERP wave form (with slight random variations between trials) was added to each trial’s signal to create the final artificial dataset.
- Figure 2.
Simulation results. A, Median latency (blue line) at which phase modulation effects can be measured depending on the frequency of the phase modulation introduced at 40 ms (true latency represented by the green dashed line) when the ERP is included in the artificial datasets. The color bar indexes the percentage of significant datasets at each time point after Bonferroni correction. The purple dashed line represents the outer edge of the window function (Morlet wavelets). B, Representation of the evoked response (lower panel) included in the artificial data in A, and its time-frequency content (upper panel). The color bar represents the oscillatory amplitude (arbitrary unit) at each frequency and time point. C. Same analysis as in A but without ERP included in the artificial datasets. Note the absence of temporal latency distortion in this case. Note also that the temporal smearing created by the window function is slightly shorter than the window duration (i.e., it does not reach the purple line) as Morlet wavelets have Gaussian tapers on either end.
- Figure 3.
White-noise (WN) paradigm. The IRF to WN sequences can be extracted by cross-correlating the stimuli sequence with the recorded EEG: here, an example IRF is shown from one subject on electrode POz. This is what we did in the first session of our experiment. This IRF can, in turn, be used to reconstruct the brain activity (reconstructed EEG) to any new WN sequence by convolution. This was done in the second session of the experiment. The stimulus fluctuations around the target were removed in the second session to avoid any target masking by the luminance (Fig. 5).
- Figure 4.
Correlation between reconstructed and recorded EEG. Two models of brain EEG activity were tested: one based on the ERP to targets (left) and the other based on the IRF to WN sequences (right). In both cases, we systematically correlated the prediction of the model with the single trial recorded EEG. For the ERP models, the “isolated targets” coefficients are based on the dataset presented in Busch et al. (2009), while the “targets in WN” are from our own dataset, with the ERP extracted relative to the targets embedded in WN (Fig. 5C). For the IRF models, we used the IRF to model the brain response to the WN sequences, and then correlated reconstructed and recorded EEG data using the raw signals, as well as signals filtered in different frequency bands (δ: 2–4 Hz; θ: 4–8 Hz; α: 7–14 Hz; β: 14–28 Hz; γ: 30–60 Hz). The gray dots represent the mean coefficient for each subject (1 dot per subject) across cross-validation runs at the maximum electrode (red dot on the topographies); the error bars represent the 95% CI of the mean (black dot) coefficient across subjects. The topographies represent the mean correlation coefficients across subjects and cross-validation runs. Shaded areas represent channels not significant after FDR correction. Note the difference of color scales for the β and γ correlation coefficients relative to the other topographies.
- Figure 5.
Classification image and evoked response. A, Example stimuli sequences for session 1, when the EEG was recorded, and session 2, when only the behavioral response was recorded. The sequences are centered on the embedded target presentation (at t = 0 ms). B, Classification image of the luminance values around the target (at t = 0 ms) for detected (orange) and missed (blue) targets. Darker colors represent the mean across subjects and trials. The lighter shades represent the standard error of the mean across subjects. The dotted line represents target presentation. Note that the same sequences were shown to all subjects in session 2. C, ERP evoked by the detected (orange) and missed (blue) targets embedded within the WN sequences. These are computed separately for the recorded EEG for session 1 (top) and the reconstructed EEG for session 1 (middle) and 2 (bottom). Note the absence of visible target-evoked ERP in reconstructed signals for session 2 (bottom). While there seems to be a very strong phase opposition between hits and misses in the reconstructed EEG to session 1, this is likely to be an artifact of the strong relationship between luminance value and behavioral outcome illustrated in B, rather than a direct relationship between phase and perception. Darker colors represent the mean across trials and the lighter shades represent the standard error of the mean across subjects. The dotted line represents target presentation.
- Figure 6.
Z-score of POS for the reconstructed EEG and phase dependent performance. A, Z-scores of the POS values aggregated across electrodes and subjects for the reconstructed EEG for each time and frequency point. Significant time points after a FDR correction (α = 0.05) are outlined in green. B, Z-scores of the POS values aggregated across subjects for the reconstructed EEG at the frequencies and time points inside the largest significant cluster (from 3–8.17 Hz and from −12.5 ms to 250 ms, A). C, Normalized hits ratio depending on the phase bin for the reconstructed EEG at the peak significance (left hand side: 6.1 Hz, +75 ms). For comparison, previously published results linking detection performance and prestimulus phase of actually recorded (rather than reconstructed) EEG signals are replotted on the right hand side (7 Hz, −120 ms; taken from Busch et al., 2009). Both phase dependence curves are measured on the same fronto-central channel. Note that in our experiment the phase bins were not realigned between subjects, but that they were realigned (to produce peak performance at 0 for all subjects) for the experiment by Busch et al. (2009). The black line represents the mean across subjects (left hand side: 20 subjects, right hand side: 14 subjects), the bars represents the SEM, the red line represents a cosine fit, whose modulation amplitude (in %) is noted above the curve.
In this issue
