EEG oscillations reveal neural correlates of evidence accumulation

Abstract

Recent studies have begun to elucidate the neural correlates of evidence accumulation in perceptual decision making, but few of them have used a combined modeling-electrophysiological approach to studying evidence accumulation. We introduce a multivariate approach to EEG analysis with which we can perform a comprehensive search for the neural correlate of dynamics predicted by accumulator models. We show that the dynamics of evidence accumulation are most strongly correlated with ramping of oscillatory power in the 4-9 Hz theta band over the course of a trial, although it also correlates with oscillatory power in other frequency bands. The rate of power decrease in the theta band correlates with individual differences in the parameters of drift diffusion models fitted to individuals' behavioral data.

Keywords: EEG; decision making; drift diffusion model; oscillations.

Figure 1

Sample regressors. Pictured are, from top to bottom: upramp, downramp, stimulus regressor, response regressor, and eyeblink regressor. X-axis represents time in samples, y-axis regressor amplitude in arbitrary units.

Figure 2

Schematic of the different models we used to create regressors from. Top row: RT model. Bottom row: DDM model. Left column: low-coherence condition. Right column: high-coherence condition. In the RT model, the ramp always starts at stimulus onset and ends always at the response, and it always has a height of one. It therefore has a different length for the slower and faster trials within a coherence condition. Conversely, the DDM model has a fixed shape for all trials within the low-coherence condition, and another shape for all trials within the high-coherence condition. This shape is determined by three DDM parameters: non-decision time (which determines ramp onset), decision threshold (which determines ramp height) and drift rate (which determines the slope of the ramp).

Figure 3

Stimulus and response-locked event-related potentials (ERPs) used for creating the stimulus and response regressors. (A) Grand average stimulus-locked ERP (i.e., average across all channels). (B) Topographical plot of ERP amplitude in the gray time window in the time course in (A) which represents the first stimulus-evoked peak. (C) Grand average response-locked ERP. (D) Topographical plot of ERP amplitude in the gray time window in the time course in (C) which represents the maximum response-related peak.

Figure 4

Graphical overview of the CCA/GLM method we developed. In the first step (A), the electrodes-by-time matrices are concatenated in the time dimension for all participants, where only a subset of the data of each participant is used (orange rectangles). This concatenated matrix with EEG data is then used together with the corresponding concatenation of regressors (purple; the DDM-inspired model time series) in the second step. (B) In this second step, the electrode-by-time matrix that contains data from all participants is correlated with the corresponding time course of the regressor (e.g., upramp) using CCA. This yields a correlation value, a weight on the regressor, and a set of weights on the electrodes (all in cyan). In the third step (C), the weight map on the electrodes is applied to the remainder of every participants’ EEG data (white) and the correlation (green) of this weighted EEG data to the regressor (purple) is compared to the correlation based on the group data in the random effects analysis (cyan). This whole procedure is done for the raw EEG data and separately for each frequency band (cf. Figure 9 below). The number of time points indicated above the matrices are just an indication.

Figure 5

Mean accuracy (A), response time (B), and coherences (C) across subjects for the low and high-coherence (difficult and easy) task conditions.

Figure 6

Basic spectrograms for electrode Cz (a representative electrode). Black line indicates the event to which the data are aligned (onset of dot-motion for the stimulus-locked graphs and the response for the response-locked graphs). Cyan line indicates the average response time and stimulus onset time, respectively. (A) stimulus-locked spectrograms for the high-coherence, low-coherence, and non-integration condition. There is a gradual decrease in oscillatory power over the course of the trial. (B) Difference spectrograms comparing low- and high-coherence conditions. Left column: stimulus-locked. Right column: response-locked. (C) Difference spectrograms indicating the contrast between integration and non-integration conditions. Left column: stimulus-locked. Right column: response-locked.

Figure 7

Basic electrophysiological data for electrode FPz. Black line indicates the event to which the data are aligned (onset of dot-motion for the stimulus-locked graphs and the response for the response-locked graphs). Cyan line indicates the average response time and stimulus onset time, respectively. These are difference spectrograms contrasting the integration and non-integration conditions. Left column: stimulus-locked. Right column: response-locked.

Figure 8

Topographical overviews of the 10% most significant electrodes for the (A) eye blink, (B) stimulus, and (C) response regressors. Red and blue reflect positive and negative regression weights, respectively, and the intensity of the shading indicates the magnitude of the regression weights. The most significant electrodes are in the locations from which the regressors were generated [indicated with arrows in (B,C)].

Figure 9

Canonical correlations as a function of frequency, with the DDM-modulated model in blue and additional correlation achieved by the RT-modulated model in red, shown separately for the ramp regressor of (A) dots (integration condition) and (B) arrows (non-integration condition). White dots indicate the 97.5th percentile of the distribution of canonical correlations expected based on random regressors. Letters indicate frequency bands: EEG = raw EEG, D = 2–4 Hz delta, T = 4–9 Hz theta, A = 9–14 Hz alpha, B = 14–28 Hz beta, G1 = 28–48 Hz low gamma, and G2 = 48–90 Hz high gamma.

Figure 10

Validation of the subset method: within-subject correlation of weighted regressors with EEG data divided by the across-subject canonical correlates. Perfect validity of the subset method would yield a fraction of one (within-subject correlations equal to across-subject canonical correlation). Each datapoint used to create this histogram reflects a single participant. This distribution has a mean that is not different from one.

Figure 11

Stimulus-locked (A) and response-locked (B) time courses of the canonical correlate in the 4–9 Hz theta band, correlated with the dots upramp, cyan line indicates RT in (A) and dots onset in (B). (C) Dots upramp topography (red indicates a positive correlation between oscillatory power and the regressor, blue a negative correlation). The shade of the color indicates the magnitude of the correlation. Note that there are no positive correlations for the dots upramp.