Abstract
Electroencephalogram oscillations recorded both within and over the medial frontal cortex have been linked to a range of cognitive functions, including positive and negative feedback processing. Medial frontal oscillatory characteristics during decision making remain largely unknown. Here, we examined oscillatory activity of the human medial frontal cortex recorded while subjects played a competitive decision-making game. Distinct patterns of power and cross-trial phase coherence in multiple frequency bands were observed during different decision-related processes (e.g., feedback anticipation vs. feedback processing). Decision and feedback processing were accompanied by a broadband increase in cross-trial phase coherence at around 220 msec, and dynamic fluctuations in power. Feedback anticipation was accompanied by a shift in the power spectrum from relatively lower (delta and theta) to higher (alpha and beta) power. Power and cross-trial phase coherence were greater following losses compared to wins in theta, alpha, and beta frequency bands, but were greater following wins compared to losses in the delta band. Finally, we found that oscillation power in alpha and beta frequency bands were synchronized with the phase of delta and theta oscillations (“phase–amplitude coupling”). This synchronization differed between losses and wins, suggesting that phase–amplitude coupling might reflect a mechanism of feedback valence coding in the medial frontal cortex. Our findings link medial frontal oscillations to decision making, with relations among activity in different frequency bands suggesting a phase-utilizing coding of feedback valence information.
INTRODUCTION
Decision making is a critical function of the brain and is supported by a network of neural structures that relies on functioning in the medial frontal cortex (Kennerley, Walton, Behrens, Buckley, & Rushworth, 2006; Ridderinkhof, Ullsperger, Crone, & Nieuwenhuis, 2004; Bush et al., 2002). Indeed, lesions to the rat or nonhuman primate medial frontal cortex impair animals' ability to adapt reward-seeking behavior flexibly according to changes in the environment (Rushworth, Buckley, Behrens, Walton, & Bannerman, 2007; Kennerley et al., 2006; Walton, Bannerman, Alterescu, & Rushworth, 2003; Walton, Bannerman, & Rushworth, 2002). Although the neurocognitive mechanisms of decision making are receiving increasing attention in cognitive neuroscience (Rushworth et al., 2007; Volz, Schubotz, & von Cramon, 2006; Ridderinkhof et al., 2004; Montague & Berns, 2002), much remains unknown about even the basic neural processes involved in simple decision-making situations. Electroencephalography (EEG), which measures the sum of dendritic activity and has a high temporal resolution, offers a way to examine the time course of neural events, including oscillation power and cross-trial phase coherence. Researchers who study EEG correlates of human decision making typically focus on the feedback-related negativity (FRN), a relatively negative potential beginning around 200 msec following negative compared to positive feedback, that is maximal over fronto-central scalp electrodes (Cohen, Elger, & Ranganath, 2007; Cohen & Ranganath, 2007; Frank, Woroch, & Curran, 2005; Holroyd, Nieuwenhuis, Yeung, & Cohen, 2003; Holroyd & Coles, 2002; Miltner, Braun, & Coles, 1997). The FRN is thought to reflect the engagement of a medial frontal network that uses positive and negative feedback to adjust future behavior toward more optimal or reward-maximizing levels (Nieuwenhuis, Holroyd, Mol, & Coles, 2004; Holroyd & Coles, 2002).
EEG data may comprise both brief bursts of neural activity as well as ongoing oscillations. Oscillations in EEG are driven by rhythmic, synchronized dendritic activity in populations of neurons. When creating stimulus-locked event-related potentials (ERPs), oscillations survive averaging only when those oscillations are phase-locked to the stimulus. Although the traditional view has been that non-phase-locked oscillations reflect background neural processes, a growing body of research demonstrates that these oscillations can yield novel insights into neurocognitive processes, beyond what can be gleaned from ERPs (Fell et al., 2004; Freeman, Holmes, Burke, & Vanhatalo, 2003; Makeig et al., 2002; Basar, Schurmann, Demiralp, Basar-Eroglu, & Ademoglu, 2001; Salinas & Sejnowski, 2001). This is supported by theoretical work and empirical research in animals, which suggests that field potential oscillations support neural communication and memory encoding, among other cognitive processes (Axmacher, Mormann, Fernandez, Elger, & Fell, 2006; Steriade, 2006; Fries, 2005; Jensen & Lisman, 2005; Buzsaki & Draguhn, 2004; Chrobak & Buzsaki, 1998; O'Keefe & Recce, 1993).
Oscillation power and cross-trial phase coherence in the theta band (4–8 Hz) over and within the medial frontal cortex have been linked to a wide range of mental operations including working memory, attention, action selection, and feedback processing (Cohen et al., 2007; Tsujimoto, Shimazu, & Isomura, 2006; Onton, Delorme, & Makeig, 2005; Wang, Ulbert, Schomer, Marinkovic, & Halgren, 2005; Basar-Eroglu & Demiralp, 2001; Kubota et al., 2001; Ishii et al., 1999; Klimesch, 1999; Basar-Eroglu, Basar, Demiralp, & Schurmann, 1992). More relevant to decision making, recent work demonstrated that evaluation of decision feedback is accompanied by changes in spectral power ranging from 4 to 30 Hz over medial frontal electrode sites (Cohen et al., 2007; Sobotka, Davidson, & Senulis, 1992). In particular, enhanced power in the theta range (4–8 Hz), and decreased activity around 20–30 Hz, has been observed following negative compared to positive feedback (Cohen et al., 2007; Marco-Pallares et al., 2007; Sobotka et al., 1992). We are not aware of any work on the oscillatory characteristics of other components of decision making, such as the decision itself or anticipation of the feedback of the decision.
In principle, oscillations in different frequency bands have the ability to act independently. However, several kinds of coupling between different frequency bands have been observed (Jensen & Colgin, 2007). One prominent type of cross-frequency coupling is termed phase–amplitude coupling, that is, when the amplitude of power in one frequency band is modulated by the phase of a lower frequency band (Canolty et al., 2006; Isomura et al., 2006; Buzsaki & Draguhn, 2004; Chrobak & Buzsaki, 1998). Phase–amplitude coupling may reflect integration or gating processes across populations of neurons, and thus, may be correlates of neural information processing or encoding (Steriade, 2006; Jensen, 2005; Lisman, 2005; Buzsaki & Draguhn, 2004). Recent evidence suggests that the prefrontal cortex exhibits phase–amplitude coupling, and that this coupling might be related to information encoding (Canolty et al., 2006; Demiralp et al., 2006; Jones & Wilson, 2005a, 2005b; Siapas, Lubenov, & Wilson, 2005). Whether phase–amplitude coupling exists during, and is related to, decision making is, to our knowledge, unknown.
Here, we examined medial frontal EEG in humans during a decision-making task in order to investigate (1) oscillatory activity during decision making, feedback anticipation, and feedback receipt; and (2) phase–amplitude coupling during decision making, and how phase–amplitude coupling might differ between task conditions (e.g., loss and win). These analyses have the potential to shed new light into the functions of the medial frontal cortex during decision making, and how the medial frontal cortex might encode feedback valence information.
METHODS
Previous nonoverlapping results from this dataset are reported elsewhere (Cohen & Ranganath, 2007), and readers are referred to the original article for additional details of the experimental design and time-domain analyses.
Subjects and Experimental Design
Fifteen right-handed subjects (9 men, aged 21–28 years) from the University of Bonn community participated. The study was approved by the local ethics review committee. In the experiment, subjects played a competitive game against a computer opponent. On each trial, the subject and a computer opponent chose a target on the left or right side of a computer monitor, using buttons controlled by the left or right index finger. If both players chose the same target, the subject lost one point, and if players chose opposite targets, the subject won one point (see Figure 1A). The computer opponent searched for patterns in the subjects' previous six decisions (e.g., left–right–left–right–left), and if it detected a pattern, it chose the target on the next trial that completed the pattern. On each of 1020 trials, subjects first saw the two targets, a fixation dot, and “you!” on the screen, and pressed the left or right button with the left or right index finger on a response box to indicate their decision, which they were instructed to make as quickly as possible (“decision” period). A green box surrounded their chosen target for 400 msec (“response period”), followed by a 1000-msec delay (“anticipation” period), followed by the computer opponent's choice highlighted in violet and “−1” or “+1” displayed above the targets for 1000 msec (“feedback” period) (see Figure 1A). We refer to these events as “loss” and “win,” respectively. Because of the competitive nature of the experiment, and because the number of wins and losses was similar, it was generally not possible for subjects to anticipate whether they would win or lose on any given trial. An intertrial interval of 1500 msec separated each trial.
Overview of experimental design and oscillation waveforms. (A) Visual display and timing of one trial. Boxes on top depict what subjects saw on the screen throughout a single trial. The response period (resp.) began when subjects depressed a response button indicating their decision (left, in this example), which was highlighted by a green square. Middle row labels each period, and bottom row displays the duration of each event. (B) Time course of frequency band power through trial events. During feedback, time courses are split into wins (dotted lines) and losses (solid lines). (C) Time course of frequency band cross-trial phase coherence. Color conventions are as in B. (D) Changes in ERP voltage potential (filtered 0.1 to 49 Hz).
Overview of experimental design and oscillation waveforms. (A) Visual display and timing of one trial. Boxes on top depict what subjects saw on the screen throughout a single trial. The response period (resp.) began when subjects depressed a response button indicating their decision (left, in this example), which was highlighted by a green square. Middle row labels each period, and bottom row displays the duration of each event. (B) Time course of frequency band power through trial events. During feedback, time courses are split into wins (dotted lines) and losses (solid lines). (C) Time course of frequency band cross-trial phase coherence. Color conventions are as in B. (D) Changes in ERP voltage potential (filtered 0.1 to 49 Hz).
EEG Recording
EEG data were recorded in a sound and electromagnetically shielded room at 1000 Hz (with an anti-aliasing low-pass filter set at 300 Hz) from 23 scalp electrodes spread out across the scalp, and four ocular (2 horizontal electrooculogram and 2 vertical electrooculogram) electrodes. Here we focus on data from channel FCz because (1) it exhibited the largest loss–win effect (see Cohen & Ranganath, 2007, for topographic maps), and (2) it is consistent with previous studies that use FCz or Fz for oscillatory analyses and feedback-related ERP analyses (Cohen & Ranganath, 2007; Cohen et al., 2007; Hajcak, Moser, Holroyd, & Simons, 2006; Nieuwenhuis, Slagter, von Geusau, Heslenfeld, & Holroyd, 2005; Miltner et al., 1997).
EEG Oscillation Analyses
There are several ways to extract estimates of frequency band-specific power and phase from EEG data, including the fast-Fourier transform, wavelet transform, and the Hilbert transform in combination with digital filtering. These different methods provide qualitatively similar information (Quian Quiroga, Kraskov, Kreuz, & Grassberger, 2002; Le Van Quyen et al., 2001). The approach used here—extracting analytic amplitude and phase values from the Hilbert transform of frequency band-filtered EEG data—has been termed “clinical mode decomposition” (Freeman, 2004), and is useful for examining oscillatory dynamics in a priori selected frequency bands. Thus, the following steps were taken in Matlab 6.5 (Mathworks, Sherborn, MA).
After extracting power and phase values for each time point, the data were segmented into the following experiment periods: (1) Decision period—500 msec prior to the button press to the onset of the button press. We selected 500 msec because average response times were slightly over this period. Thus, decision making and preparation processes leading up to the response are captured in this period. (2) Response period—Onset of the button press and simultaneous visual confirmation of the selected target to offset of visual stimulus (400 msec duration). (3) Anticipation period—Offset of visual stimulus until onset of feedback stimulus (1000 msec duration). (4) Feedback period—The duration of the feedback stimulus (1000 msec duration). Power values and EEG data were averaged across all trials for each subject to create ERPs and event-related power perturbations. EEG data were baseline corrected using the average EEG voltage potential from −200 msec until the onset of the feedback stimulus. Trials in which the maximum EEG deflection exceeded |100| μV were excluded from analyses.
To calculate phase probability maps, we first categorized each time point in each trial as being in one of 30 phase bins (described in more detail in the Phase–Amplitude Coupling section below), and then calculated the average probability of phase values being in a particular bin at each time point across trials. Then, for each subject, we calculated additional maps containing resorted phase values taken from the observed phase distribution. Averaged over all trials for each subject, this has the effect of creating a phase probability map that would be expected given oscillations that have random trial-to-trial phase. To compare the difference between the observed and randomized phase probability maps, we calculated the decibel (dB) change from the randomized to the observed map [10 * log10(observed/random)] at each point in time–phase space. Resulting dB values thus indicate that a specific value of phase is more (light colors) or less (dark colors) likely to occur than expected by chance at that point in time. Finally, pixels with dB values less than |1| were masked out and appear gray. The time-course plots shown underneath the graphs were created by taking the average absolute value across all phase bins for each time point. These time courses therefore reflect the increased likelihood of oscillation phases taking on specific values over randomized phases. This analysis is complementary to, but not redundant with, the cross-trial phase coherence analysis because this analysis identifies the specific values of phase that are more or less likely to occur at each time point, and compares this probability distribution with that expected by chance; in contrast, cross-trial phase coherence is an absolute measure that indicates the consistency of phase values over trials. Thus, this analysis is useful in determining whether oscillation phases are likely to have specific values in specific time ranges (e.g., the number of oscillation cycles).
Statistics
Statistics were conducted by first averaging power values across trials within each subject, and entering results from the subject averages into repeated measures analyses of variance (ANOVAs). As dependent measures, we used the average power or cross-trial phase coherence values for each experiment period (decision, response, anticipation, feedback). We conducted two sets of ANOVAs, one including the factors experiment period (decision, response, anticipation, feedback) and frequency band (delta, theta, alpha, beta), and one including the factors feedback valence (win, loss) and frequency band. In the former set of ANOVAs, we ignored feedback valence because subjects could not know prior to the feedback period whether they would receive a reward or not. For the latter set of ANOVAs, we included only the feedback period for the same reason. ANOVAs were conducted in SPSS 14 software (SPSS, Chicago, IL). Greenhouse–Geisser corrections were used, and adjusted degrees of freedom are reported where appropriate. Because cross-trial phase coherence values are not normally distributed (they are limited to a range of 0 and 1), we applied the Fisher transform to cross-trial phase coherence values (0.5 * log([1 + PC]/[1 − PC]), where PC is the cross-trial phase coherence value) for statistical tests; untransformed values are plotted in the figures. The Shapiro–Wilk test was used to confirm normality by testing the null hypothesis that the data were drawn from a normally distributed population; all phase variables had nonsignificant p values after Bonferroni correction, indicating that the null hypothesis cannot be rejected (i.e., the data are normally distributed). Power values were normalized within each frequency band to allow a direct comparison across frequencies. This was done by z-transforming, separately for each frequency band and subject, all values across time from all trials. This ensured that the normalization was not skewed by non-task-relevant events, such as breaks between trials. Note that here, negative values do not imply a suppression of activity; merely a decrease from the average activity levels across all time points and all trials.
Phase–Amplitude Coupling
We use the term phase–amplitude coupling to refer to a statistical relationship between power in activity in certain frequency bands and the phase of a lower frequency band oscillation. Assessing phase–amplitude coupling involves examining the distribution of power values from one frequency band over the phase space of a lower frequency band and testing the null hypothesis that the distribution is uniform. We examined phase–amplitude coupling separately for the theta and alpha phase. To do this, we first binned each time point of each trial according to its theta or alpha phase value, divided into 30 equally spaced ranges (−3.1416 to −2.9322, −2.9323 to −2.7227, etc.). Next, we separately averaged power values in each frequency band for each time–phase bin. This was performed in two separate windows: 667 (600) msec prior to and lasting until the onset of the feedback stimulus for the theta (alpha) band, and the onset of the feedback stimulus until 667 (600) msec later. The latter window was analyzed separately for losses and wins. These windows correspond to 4 (6) cycles of the theta (alpha) center frequency (6 or 10 Hz), and were selected to allow us to distinguish between phase–amplitude coupling during the anticipation and feedback periods. We did not assess phase–amplitude coupling during the decision and response periods because those conditions were shorter than the other analysis windows. As in the phase-locking analyses, we used the same number of trials for wins and losses, which ensured that no biases due to differences in the number of trials were introduced.
In a further set of analyses, we examined whether there were significant differences between the power–phase relationships following wins and losses. This was done by testing whether the difference between loss and win distributions diverged from a uniform distribution. In this case, the null hypothesis—that losses and wins have the same distribution across theta or alpha phase—would mean that their difference values have a uniform distribution. To examine the direction of this effect—whether losses or wins elicited more theta phase coupling, we subtracted the magnitude of mean projection vectors; the feedback type associated with a larger vector is relatively more entrained to phase. For these tests, we used an alpha of .025 for theta tests, which corresponds to a family-wise error rate of .05 with two comparisons (frequency band).
RESULTS
Power
Figure 1B displays the time course of the power in each frequency band over time. We first examined whether power values changed as a function of task period (decision vs. response vs. anticipation vs. feedback). Here we found main effects of period [F(2.1, 30.4) = 14.04, p < .001], frequency band [F(1.9, 27.6) = 9.0, p < .001], and a Period × Frequency band interaction [F(4, 56) = 30.2, p < .001]. This interaction was driven by opposing effects in delta and theta versus alpha and beta bands. Specifically, in both delta and theta bands, power was lowest during anticipation compared to decision, response, or feedback periods. This was confirmed by a significant quadratic relationship between period and power values in the delta and theta bands (both ps < .001) (see Figures 1B and 2A). In contrast, power in the alpha and beta bands showed the inverse relationship, being greater during anticipation compared to decision, response, or feedback. This quadratic relationship was highly significant in the beta band [F(1, 14) = 23.66, p < .001], but was not statistically significant in the alpha band. Thus, during feedback anticipation, there was a shift in the dominant oscillation power from lower frequency bands (delta and theta) to relatively higher frequency bands (alpha and beta).
Next, we examined whether power values during the feedback period differed between losses and wins (valence). Here we found a main effect of frequency band [F(2.7, 24.9) = 34.42, p < .001] and a main effect of valence [F(1, 14) = 9.16, p = .009]. However, this latter main effect is qualified by a significant Frequency band × Valence interaction [F(3.5, 35.4) = 7.36, p = .001]. This crossover interaction occurred because there was greater power following losses compared to wins in the theta, alpha, and beta bands (p values: .018, .002, .002, respectively). But, in contrast, there was greater power following wins compared to losses in the delta band. The simple effect of wins versus losses in the delta band was not statistically significant [F(1, 14) = 1.69, p = .214], although, as seen in Figure 1B, this may have arisen because the win > loss effect in delta power reversed from around 400 to 800 msec. Thus, it is possible that the statistical test was clouded by a reversed effect later in time. We thus conducted a post hoc follow-up analysis in which we analyzed the base-to-peak power value difference instead of the average power values. Here we found significantly greater delta power following wins compared to losses [t(14) = −1.99, p = .033].
In summary (see Figure 1B), we found that the decision, response, and feedback periods were characterized by enhanced delta and theta power, and the anticipation period by enhanced beta and alpha power. Further, we found that losses elicited more power than did wins in all frequency bands except delta, for which the opposite occurred.
Cross-trial Phase Coherence
Figure 1C displays the time course of cross-trial phase coherence in each frequency band over time. We first examined whether average coherence values changed as a function of task period (decision, response, anticipation, feedback). Here we found a main effect of period [F(1.8, 25.6) = 40.38, p < .001] and a Period × Frequency band interaction [F(3.4, 48.1) = 18.84, p < .001]. The main effect of period was driven by cross-trial phase coherence values being lower in the anticipation period for all frequency bands (all ps < .01). The interaction was driven by this effect being larger in lower frequency bands compared to higher frequency bands.
Next, we examined whether average cross-trial phase coherence values during the feedback period were influenced by feedback valence. There was no main effect of feedback valence [F(1, 14) = 0.96, p = .343], but a significant Valence × Frequency band interaction [F(2.0, 28.5) = 6.09, p = .006]. This was driven by a crossover interaction between theta cross-trial phase coherence, which was greater during losses than wins [F(1, 14) = 6.56, p = .023], and delta cross-trial phase coherence, in which coherence values were greater during wins compared to losses. As with power values, the win > loss difference in the average coherence values in the delta range was not significant [F(1, 14) = 2.87, p = .112], although this may have been due to the reversal of this effect later in time (see Figure 1C). Coherence values were numerically but not significantly greater during losses than wins for the alpha and beta bands (ps = .112 and .076, respectively).
It has been questioned whether cross-trial phase coherence measures such as the one we used truly measure a reset of ongoing phase values, or whether it could reflect transient additive oscillations or even temporally aligned nonoscillatory bursts. For example, Yeung, Bogacz, Holroyd, Nieuwenhuis, and Cohen (2007) and Yeung, Bogacz, Holroyd, and Cohen (2004) showed that isolated bursts of a half-oscillation that is time-locked to an event (e.g., an erroneous response) can give rise to similar ERPs as ongoing oscillations that are phase-reset by the event. To address this issue, we calculated the probability that the EEG would contain specific phase values at each point in time (see Methods). Phase reset of ongoing oscillations would imply that phase values are more likely to take on specific values for several oscillation cycles; in contrast, if a burst of nonoscillatory activity occurred independent of background oscillations, the phase would not be expected to continue being aligned across trials for several oscillation cycles. As seen in Figure 2, phase values in the delta, theta, and alpha frequency bands were indeed more likely to take on specific values than expected by chance. The graphs display the probability densities of phase values (ranging from pi to −pi on the y-axis) over time (x-axis) for each frequency band. The null hypothesis that there is no concentration of phase values across trials would result in all-gray maps (i.e., no difference between a map with randomized phase values; see Methods). In contrast, we found that specific phase values in the delta, theta, and alpha frequency bands were more likely than chance to take on specific values for several oscillation cycles than expected by chance. For example, delta phase values were most likely to take values close to pi/−pi at the onset of the feedback period (Figure 2A), and delta, theta, and alpha phase concentrations lasted around 2 to 4 cycles. Because the anticipation period was of a fixed time delay, it was possible for subjects to anticipate when feedback about their decision would be available. Note that this analysis does not prove that ongoing oscillations were reset; it only provides complimentary support because phases were more likely than chance to be at specific values over several oscillation cycles.
Phase–time probability space, depicting the increased or decreased likelihood over chance that frequency band oscillation phases would take on specific values (in radians; y-axis) at specific time points throughout the trial (x-axis). Pixels with dB values less than |1| are set to zero and appear gray for ease of viewing. Time-course plots underneath each graph show the average absolute value of the dB changes over time.
Phase–time probability space, depicting the increased or decreased likelihood over chance that frequency band oscillation phases would take on specific values (in radians; y-axis) at specific time points throughout the trial (x-axis). Pixels with dB values less than |1| are set to zero and appear gray for ease of viewing. Time-course plots underneath each graph show the average absolute value of the dB changes over time.
In summary, we observed enhanced cross-trial phase coherence during the decision and response periods in delta and theta frequency bands. There was little cross-trial phase coherence during anticipation with the exception of delta, which showed increasing coherence as the feedback period approached, and with delta phase values shifting toward pi at the onset of visual feedback. Finally, during feedback, there were sharp increases in cross-trial phase coherence in all frequency bands. Cross-trial phase coherence was greater for losses compared to wins in all frequency bands except delta, for which the opposite occurred. Phase values took on specific values at specific points in time, providing support that ongoing oscillations were reset by certain task events.
Timing of Power and Cross-trial Phase Coherence Peaks
From inspection of Figure 1, it appears that power in different frequency bands peaked at different times, whereas cross-trial phase coherence peaked at roughly the same time. To assess whether these differences were significant, we conducted a 2 (measure: power, cross-trial phase coherence) × 2 (feedback) × 4 (frequency band) ANOVA on the postfeedback peak time of each subject's response. We found a significant main effect of frequency band [F(2.4, 29.4) = 6.36, p = .005], which was qualified by a Frequency band × Measure interaction [F(2, 28.4) = 3.06, p = .05]. The Frequency band × Measure interaction arose because there were no significant effects in peak times for cross-trial phase coherence (all ps > .07), but power showed a significant main effect of frequency band [F(2, 28) = 6.09, p = .006]. There was also a main effect of feedback [F(1, 14) = 5.38, p = .036], such that power peaks were earlier in time following wins than following losses. There was no Frequency band × Feedback interaction (p = .30).
Phase–Amplitude Coupling
In our final set of analyses, we examined how power values in different frequency bands might be synchronized with the theta and alpha phase. This was done by calculating the average power values for each binned range of theta or alpha phase, and testing whether this power–phase distribution diverged from a uniform distribution (see Methods).
Coupling with the Theta Phase
Alpha power was significantly entrained to theta phase during anticipation (p < .001), and following losses (p < .001) and wins (p = .003) (Figure 3A). The loss–win difference was also significantly related to theta phase (p < .001), with losses having a larger projection vector than wins. Finally, we found that beta power was entrained to theta phase during anticipation (p < .001) and during wins (p < .001) and losses (p = .011) (Figure 3B). During anticipation and feedback, preferred theta phase was around pi/2. Losses and wins were not differentially distributed (p = .097).
Cross-frequency coupling with the theta and alpha phase. Graphs in the left-hand column show the average power for each level of phase, binned into 30 categories (see text), plotted separately for anticipation (gray), loss (dashed black line), and win (solid black line). Note that because phase values are circular, the x-axis wraps around such that pi = −pi. Polar plots in the middle column illustrate mean projection vectors onto the theta or alpha phase, plotted separately for anticipation, loss, and win. The next column depicts the difference between losses and wins across phase. Positive values indicate increased power for losses compared to wins; negative vales indicate the opposite. (A) Alpha–theta phase coupling, (B) beta–theta phase coupling, (C) beta–alpha phase coupling. Finally, polar plots in the right-hand-most column depict the preferred angle of the difference between losses and wins.
Cross-frequency coupling with the theta and alpha phase. Graphs in the left-hand column show the average power for each level of phase, binned into 30 categories (see text), plotted separately for anticipation (gray), loss (dashed black line), and win (solid black line). Note that because phase values are circular, the x-axis wraps around such that pi = −pi. Polar plots in the middle column illustrate mean projection vectors onto the theta or alpha phase, plotted separately for anticipation, loss, and win. The next column depicts the difference between losses and wins across phase. Positive values indicate increased power for losses compared to wins; negative vales indicate the opposite. (A) Alpha–theta phase coupling, (B) beta–theta phase coupling, (C) beta–alpha phase coupling. Finally, polar plots in the right-hand-most column depict the preferred angle of the difference between losses and wins.
Coupling with the Alpha Phase
Beta power was entrained to alpha phase during anticipation (p < .001) and during wins (p = .008) and losses (p < .001) (Figure 3C). Here, the loss–win difference was also significantly related to alpha phase (p < .001). During anticipation and feedback, preferred alpha phase approached 0, with power being more entrained to alpha phase during losses than during wins.
In summary, we found that power in alpha and beta frequency bands was modulated by the phase of slower oscillations. In all cases, in which phase–amplitude coupling significantly differed between losses and wins, relatively more oscillation power was observed during slower oscillation troughs during wins, and relatively more oscillation power was observed during slower oscillation peaks following losses.
DISCUSSION
We analyzed EEG activity over the medial frontal cortex during decision making, and found significant changes in oscillatory power and cross-trial phase coherence in specific frequency bands during different cognitive processes related to decision making. We observed the following sets of findings. First, the decision and feedback periods were characterized by enhanced delta and theta power and cross-trial phase coherence that may have reflected a reset of ongoing oscillations. In contrast, feedback anticipation was associated with a relative shift in the dominant power structure to alpha and beta frequency bands. Second, we found phase–amplitude coupling of power fluctuations in relatively higher frequency bands with the phase of relatively slower oscillations. This phase–amplitude coupling differed between losses and wins, suggesting that the medial frontal cortex at least partly encodes feedback valence information in the relations among activities in different frequency bands. Many of these findings, including phase–amplitude coupling and oscillations that are not highly phase-locked to a stimulus, are not observable in time-domain analyses, demonstrating the usefulness of frequency decompositions of EEG data.
Oscillatory Characteristics of Distinct Components of Decision Making
The different periods—decision, response, anticipation, feedback—were characterized by different patterns of oscillatory and ERP activity. The decision period, leading up to the button press indicating which decision was chosen, was characterized by moderate enhancements in delta and theta frequency band power, strong enhancements of delta and theta frequency band cross-trial phase coherence, and a positive–negative deflection cycle in the ERP. The anticipation period was characterized by relative decreases in delta and theta power, and a shift to increased alpha and beta power, as well as an increase in delta cross-trial phase coherence, leading up to the time of feedback. The feedback period was characterized by strong and transient (0–600 msec) increases in power in delta and theta bands, enhanced cross-trial phase coherence in all frequency bands, and a strong positive deflection in the ERP. Interestingly, in contrast to the other frequency bands, alpha power showed an early enhancement of power, and a later and more sustained decrease in power (desynchronization), likely reflecting enhanced cognitive processing (Klimesch, Sauseng, & Hanslmayr, 2007; Klimesch, 1999). This decrease in alpha power was coupled with a concomitant increase in cross-trial phase coherence, suggesting a feedback-induced phase reset (Makeig, Debener, Onton, & Delorme, 2004; Makeig et al., 2002).
The divergences between peak times in the power and cross-trial phase coherence domains suggest that power and phase alignment may reflect different neural processes. Specifically, power enhancements are thought to reflect synchronized and rhythmic activation of neural populations (e.g., Klimesch et al., 2007). On the other hand, an increase in cross-trial phase coherence, when temporally dissociated from power increases, might reflect a “reset” of ongoing oscillations (Makeig et al., 2004), such that the temporal sequence of neurocognitive processes is similar on each trial. Our findings suggest that feedback processing causes a reordering of phases in multiple frequency bands at around 220 msec, which roughly corresponds to the time when the differences between loss and win ERPs emerge. Because the feedback was given visually, it is possible that lower-level visual processes contributed to this phase reset. However, this explanation cannot entirely account for the results for two reasons. First, the alpha and gamma cross-trial phase coherence peaks following feedback were larger than they were following the button press (which was also accompanied by visual stimulation). Second, and more importantly, there were significant differences between losses and wins, despite the similarity of the visual stimulation. Thus, cognitive, and feedback valence-related, processes also contributed to the reset of feedback-induced ongoing oscillations.
It remains debated whether cross-trial phase coherence measures reflect a reset (i.e., an abrupt change) of oscillation phase values, or whether they reflect stimulus-locked additive oscillations or nonoscillatory bursts of neural activity (Yeung et al., 2007; Mazaheri & Jensen, 2006; Fell et al., 2004; Makeig et al., 2002). Although our analyses do not prove either side of this argument, several considerations suggest that ongoing oscillations were phase-reset in our task. First, a nonoscillatory burst of activity would be expected to increase phase coherence measures only over a brief period (Yeung et al., 2004), but we found enhanced cross-trial phase coherence over several oscillation cycles, and the duration of the enhanced phase coherence was related to the duration of the oscillation cycles (i.e., it was longer in the delta than in the alpha range; Figure 2A and C). Second, in the case of alpha, the time courses of cross-trial phase coherence and power fluctuations were dissociated. In contrast, stimulus-locked additive oscillations or nonoscillatory bursts of neural activity would be expected to induce simultaneous phase coherence and power increases. Third, the direction of the feedback valence effect was in opposite directions for different frequency bands (delta vs. theta, alpha, and beta). If the cross-trial phase coherence effect was driven by a nonoscillatory burst that was of larger amplitude for losses compared to wins, it would be expected to incite broadband changes in the same direction; our findings suggest that brain oscillations in different frequency bands respond differently to feedback valence. We note that these analyses do not preclude the possibility of transient additive oscillations or nonoscillatory bursts; these may have occurred in addition to a phase reset of ongoing oscillations. It is also possible that band-pass filtering a truly phasic burst of activity (e.g., a sine half-wave), which would result in ringing artifacts around the edges of the burst, would produce results similar to what we displayed in Figure 2 (see also Yeung et al., 2004, 2007). However, we would like to note that decision and feedback-related theta oscillations can even be recognized in the raw EEG data (see Figure 1D), in particular, for the loss trials.
Phase–Amplitude Coupling in the Medial Frontal Cortex
The second set of findings was that power in alpha and beta frequency bands was coupled with the theta and alpha phase, and that these power–phase relationships changed in different conditions. Phase–amplitude coupling might reflect communication within neural populations (Buzsaki & Draguhn, 2004; Varela, Lachaux, Rodriguez, & Martinerie, 2001) and a mechanism for information encoding (Axmacher et al., 2006; Jensen & Lisman, 1998, 2005). The idea is that information is coded in the brain by simultaneous coactivation of a population of neurons (leading to coherent oscillations in that network); to disambiguate different kinds of information (e.g., loss vs. win) signaled by spatially overlapping populations of neurons, signals can be embedded in a specific phase range of a slower oscillation (Lisman, 2005). In the context of the present findings, the medial frontal cortex might differentiate losses and wins partly by the relation between relative increases in oscillation power and the phase of a slower oscillation. This does not imply that wins and losses are processed only during specific phases of slower oscillations, but rather that some kinds of processing (e.g., related to the valence or reinforcement value) might preferentially occur during oscillation troughs and peaks for wins and losses, respectively.
Phase–amplitude coupling may also reflect functional connectivity between disparate brain regions. For example, medial frontal gamma-band activity (30–50 Hz) can be induced via transcranial magnetic stimulation of the cerebellum in humans (Schutter, van Honk, d'Alfonso, Peper, & Panksepp, 2003), and via electrical stimulation of the hippocampus in rats (Izaki, Nomura, & Akema, 2003; Izaki, Takita, Nomura, & Akema, 2002). Medial frontal oscillations are also synchronized with coherent activity in the ventral tegmental area (Hernandez-Gonzalez, Navarro-Meza, Prieto-Beracoechea, & Guevara, 2005) and basal forebrain structures (Lin, Gervasoni, & Nicolelis, 2006). Thus, phase–amplitude coupling in the medial frontal cortex in our study might be related to neural communication between subcortical and cortical structures. This may be in addition to, or an integral part of, feedback valence encoding.
Encoding of Feedback Valence in the Medial Frontal Cortex
How are wins and losses processed in the medial frontal cortex? Based on the current and previous findings, we can begin to develop a cohesive set of electrophysiological characteristics of the medial frontal cortex during feedback processing. First, feedback-locked ERPs show a strong modulation of activity by feedback valence, which typically begins around 200 msec and lasts until around 300–400 msec, and is more positive for wins than for losses (Cohen & Ranganath, 2007; Frank et al., 2005; Nieuwenhuis et al., 2004; Gehring & Willoughby, 2002; Holroyd & Coles, 2002; Miltner et al., 1997). This loss–win difference is thought to reflect activation of a medial frontal learning system that evaluates outcomes of behavior to guide decision making (Nieuwenhuis et al., 2004; Holroyd et al., 2003; Holroyd & Coles, 2002). We and others (Cohen et al., 2007; Luu & Tucker, 2001) have demonstrated that losses also induce significantly more theta power and cross-trial phase coherence than do wins. Event-related activity in other frequency bands (e.g., beta, lower gamma) exists as well, and although the timing of activity peaks and troughs seems to be consistent across studies, their modulation by feedback valence might be sensitive to the particular task (see below). In the present study, these changes in power were accompanied by a broadband phase reset, which was greater for losses than for wins in all frequency bands except delta, for which the opposite was observed. In our previous study (Cohen et al., 2007), we found that, similar to the present findings, feedback-locked power in different frequency bands had different time courses, but the timing of cross-trial phase coherence effects were similar across frequency bands. Together, these findings demonstrate that decision feedback incites increased broadband cross-trial phase coherence around 200 msec and dynamic changes in multiple frequency bands. In most cases, power and cross-trial phase coherence is greater for losses compared to wins.
In two previous studies (Cohen et al., 2007; Marco-Pallares et al., 2007), it was found that wins elicited more activity in the lower gamma band (20–30 Hz) compared to losses. Based on these findings, one might expect a similar reversal in the beta range, which we did not see. One possible reason for this discrepancy might be related to differences in frequency band selection (21–29 or 20–30 Hz previously and 13–25 Hz in this study). To assess whether this discrepancy accounted for the differences between this and our previous study, we reanalyzed the feedback data using a frequency range of 21–29 Hz. However, losses still had increased power compared to wins [t(14) = 2.4, p = .03] and cross-trial phase coherence did not differ [t(14) = 0.44]. However, although the direction of the feedback valence modulation differed, the power and cross-trial phase coherence time courses were similar to those in previous studies (Cohen et al., 2007; Marco-Pallares et al., 2007), with a later peak than observed in lower frequency bands, and a temporally dissociated peak in cross-trial phase coherence. Thus, it seems that although the general involvement of beta/gamma oscillations is consistent across experiments and subjects, the specific effects of feedback valence might be sensitive to task demands.
An important direction for future research is to link medial frontal oscillations and phase–amplitude coupling (or any other measure of cross-frequency coupling) to dopamine system functioning. Holroyd and Coles (2002) proposed that the FRN (i.e., the loss–win ERP difference) is driven by the impact of the dopamine response on the anterior cingulate cortex. Might dopamine functioning be related to medial frontal oscillations and cross-frequency coupling? Some indirect evidence comes from Silberstein et al. (2005), who found enhanced cross-channel synchronization in beta (13–24 Hz) and gamma (60–80 Hz) frequency bands, largely in medial frontal sites, when Parkinson's patients were on, compared to off, their dopamine medication. Also, Demiralp et al. (2007) have shown that oscillation power in the gamma range (30–80 Hz) over the medial frontal cortex differs across individuals with different dopamine receptor genotypes. Thus, the dopamine system appears to modulate medial frontal oscillatory activity, and dopamine is heavily involved in reinforcement learning (Frank, Seeberger, & O'Reilly, 2004; Montague, Hyman, & Cohen, 2004; Dayan & Balleine, 2002; Holroyd & Coles, 2002; Schultz, 1998), so one would expect the dopamine system to be involved in generating or modulating oscillations and cross-frequency coupling during decision making and reinforcement learning.
Conclusions
We have demonstrated changes in frequency power, cross-trial phase coherence, and phase–amplitude coupling in the medial frontal cortex during decision making, anticipation, and feedback processing. These findings highlight the oscillatory nature of activity in the medial frontal cortex and suggest that different cognitive processes are supported by oscillations in different frequency bands. Further, our data provide evidence that loss and win information may be at least partly encoded as the relation between oscillation power in the alpha and beta frequency bands and the phase of theta and alpha frequency bands. To our knowledge, this study is the first to apply such analytic techniques to the neurocognitive mechanisms underlying decision making. This helps to lay groundwork for future studies to continue examining the oscillatory nature of neural activity during decision making.
Acknowledgments
We thank Charan Ranganath for his help designing the experiment, and Beate Newport and Nicole David for technical assistance. This research was supported by DAAD and an NRSA from NIDA awarded to M. X C.
Reprint requests should be sent to Michael X Cohen, Department of Psychology and Center for Neuroscience, University of California, Davis, CA 95616, or via e-mail: mcohen@ucdavis.edu.