Differential modulation of behavior by infraslow activities of different brain regions

Dataset

We used the publicly available dataset downloaded from https://openneuro.org/datasets/ds003517/versions/1.1.0 (Cavanagh & Castellanos, 2021). The dataset contains two behavioral tasks: “Exemplar tasks” and “Escape from Asteroid Axon”. We only used data of the “Escape from Asteroid Axon” behavioral task (video game). A 63-channel EEG (500-Hz sampling rate) was recorded when 17 participants played the video game. The experiment on the “Escape from Asteroid Axon” behavioral task lasted 45 min. The player in the video game aimed to shoot missiles to hit enemies while avoiding obstacles (walls and enemies). The player’s six behaviors were logged as follows: “Missile launch button” (the player shoots missiles.), “Collect star” (The player fills the health bar.), “Collect ammo box” (The player fills the missiles.), “Crash into wall” (The player hits the wall. Then, the health bar decreases.), “Crash into enemy” (The player hits the enemy. Then, the health bar decreases.), and “Missile hit enemy” (Missiles fired by the player hit the enemy.) Participants’ behaviors were logged using a Logitech F310 gamepad (Cavanagh & Castellanos, 2016).

Monto et al. (2008) used a somatosensory discrimination task which needs perceptual sensitivity. To extend these findings to other brain functions, we chose this “Escape from Asteroid Axon” task which needs a function of visuo-motor control (“Missile launch button”, “Crash into wall”, “Crash into enemy”, and “Missile hit enemy”) and an overall sustained attention to a behavioral task (“Collect star” and “Collect ammo box”). The purpose of this study was to investigate whether these behaviors, which are known to be related to different brain functions, are modulated by EEG ISA in the same or different brain regions.

Behavioral data analysis

Since the six behaviors could be closely related to each other in playing the game, we analyzed the dependencies of the different behaviors. One type of these dependencies is the causal relationship between behaviors: If a missile fired by the player hits the enemy (“Missile hit enemy”), the player must have fired the missile before that (“Missile launch button”). Another type of dependencies to consider is probabilistic antecedents: When the player fires a missile (“Missile launch button”), the missile won’t always hit the enemy (“Missile hit enemy”), but it will hit the enemy with a certain probability (“Missile hit enemy”). In this sense, “Missile launch button” is a probabilistic antecedent behavior of “Missile hit enemy”. The reverse is also possible: If the player continues to fire multiple missiles, the player may have continued to fire missiles (“Missile launch button”) after the first of those missiles hit the enemy (“Missile hit enemy”). In this case, “Missile hit enemy” is a probabilistic antecedent behavior of “Missile launch button”. As another example, if the player hits a wall and the health bar is reduced (“Crash into wall”), the player will try to fill the health bar by collecting stars (“Collect star”). In this case, “Crash into wall” is a probabilistic antecedent behavior of “Collect star”. We measured these probabilistic antecedents that reflect the behavioral task structure. The reason for measuring these probabilistic antecedents is that if there is a strong antecedent relationship between the behaviors, the timing of occurrence of an ensuing behavior is likely to be mainly modulated by the antecedent behavior, not the EEG ISA. If there is no strong antecedent relationship between behaviors, our hypothesis is that the behaviors will be modulated by EEG ISA.

We quantified how a behavior (e.g., Missile hit enemy) depended on the preceding occurrence of another behavior (e.g., Missile launch button) by calculating the ratio of the actual to the pseudorandom durations after the onset of the preceding behavior before the onset of behavior 1. We called this quantification as “the timing relative to random” (see Fig. 1C).

Specifically, let t be the time of occurrence of behavior 1 and let s_t be the duration after the onset of the preceding behavior before the onset of behavior 1. For simplicity, we downsampled the timing data for the occurrence of all behaviors from 500 Hz to 1 Hz. Let r_t be the pseudo-random latest occurrence timing of the preceding behavior before the occurrence of behavior 1. r_t was obtained from the sequence of Poisson random numbers generated once per second. The mean parameter of the Poisson random numbers was estimated from the actual frequency of the preceding behavior. Let $〈r_t〉$ be the average of the 10,000-independently generated r_t’s. The timing relative to the random is given by: (1)${\displaystyle \text{Timing relative to random}=100{〈\frac{s_t}{〈r_t〉}〉}_1}$

where ${〈\cdot 〉}_1$ is an average over all occurrences of behavior 1.

To quantify how the occurrence of one behavior depends on the occurrence of another behavior, we assessed how random the closeness of the dependency (timing relative to random; Eq. (1)) was (100%). If the timing relative to random was markedly below 100%, it indicated that the occurrence of behavior 1 was dependent on the occurrence of the preceding behavior.

We also determined whether the dependencies of some behaviors were more significant than those of others. To quantify this, we measured how much the dependency on the preceding behavior (timing relative to random) differed for each of five preceding behaviors. We expressed this as the ratio of the variability of the dependencies between preceding behaviors to the variability of the dependencies within each preceding behavior, which is the F-statistics as a quantification indicator. For each behavior, we calculated the timing relative to random values for the behavior and the other five behaviors for each participant using Eq. (1). We calculated the F-statistics of the timing relative to random among the five behaviors. A high F-statistic (i.e., a large difference in the timing relative to random for the other five behaviors) indicated that the corresponding behavior was highly dependent on the other specific behavior(s), while a low F-statistic indicated that the corresponding behavior was less dependent on the occurrence of specific behaviors.

EEG ISA analysis

To determine whether the EEG data showed specific frequency peaks in the infraslow frequency band, which would indicate that there was a dominant oscillation in the ISA, the magnitude for each of the frequencies of the EEG data was computed. The magnitude of raw EEG data at each frequency was measured by the absolute value of the Fourier transformed EEG data, which was computed by the fast Fourier transform (FFT) algorithm (see Fig. 1A).

To extract ISA from the raw EEG data, we performed bandpass filtering at 0.01–0.1 Hz. We also applied the Hilbert transform to estimate the oscillation phase of the ISA. Note that the oscillation phase at 0 rad was located at the maximum amplitude of the ISA, and the oscillation phase at −π = π rad was located at the minimum. The oscillation phase at $\left[-\pi ,0\right]$ rad was the rising phase, and the oscillation phase at $\left[0,\pi \right]$ rad was the falling phase. After the Hilbert transform, we downsampled the EEG ISA phase data from 500 Hz to 1 Hz to place the phase data at the same time scale as the behavioral data (see above).

To investigate the association between behavior and EEG ISA, the ISA phase was divided into eight oscillation phase bins: i = 1, …, 8. The central phase of each bin was $\left(2i-1\right)\pi /8$ and the width of each bin was π/4. The occurrence of each behavior in each participant was counted for each phase bin: b_i, i = 1, …, 8. The normalized occurrence of a behavior with respect to the ISA phase is calculated as (2)${\displaystyle n_i=\frac{b_i/p_i}{\sum _{j=1}^8{b}_j/p_j}}$

where p_i is the measure of the oscillation phase duration for each phase bin i = 1, …, 8. Instead of assuming that the oscillation phases of the EEG ISA occur equally in all phase bins, we normalized to the amount that actually occurs.

To quantify how significantly the occurrence of a behavior varied with the EEG ISA phase, the behavior occurrence frequency relative to the chance level, m_i, at an oscillation phase bin i (see Fig. 2A), was computed as follows: (3)${\displaystyle m_i=100\left(8n_i-1\right).}$

Note that because ${\sum }_{i=1}^8{n}_i=1$ and ${\sum }_{i=1}^8{m}_i=0$, Eq. (3) indicates the zero-centered deviation of behavioral occurrence on the percentile scale, if n_i = 1/8 (chance level), m_i = 0, and a positive (or negative) value of m_i indicates a more (or less) frequent occurrence of behavior i.

These m_i values for Eq. (3) were first obtained for each participant and averaged over all the participants. This behavior occurrence frequency relative to the chance level (m_i) was the basis for further analyses after averaging for all participants. The peak-to-trough difference in the behavior occurrence frequency over the phase bins, d_{peak-to-trough}, was computed as follows: (4)${\displaystyle d_{\text{peak-to-trough}}=m_{\alpha }-m_{\beta }}$

where α = argmax_im_i and β = argmin_im_i. To identify the EEG channels that showed more significant behavioral modulation, a normalized peak-to-trough difference in behavior occurrence over the phase bins was obtained by normalizing d_{peak-to-trough} over all the EEG channels. We used MATLAB for all our analyses.

Dependencies among behaviors

The frequency-magnitude curve of EEG showed log–log linearity without a clear peak at the 0.01–0.1 Hz infraslow frequency band (Fig. 1A). This result was consistent with that of a previous study by Monto et al. (2008). The frequency-magnitude curve in Fig. 1A was reminiscent of 1/f noise, giving rise to the doubt that slow EEG may be simply composed of noise with no brain activity. However, as we demonstrated below, the ISA of EEG data used in this study revealed associations with behavior, indicating that ISA at 0.01–0.1 Hz carried information related to brain activity, which was not illustrated in the frequency-magnitude curve.

The frequency of occurrence of each behavior during game play was different. Among the six behaviors (“Missile launch button”, “Collect star”, “Collect ammo box”, “Crash into wall”, “Crash into enemy”, and “Missile hit enemy”), the occurrence frequencies of “Missile launch button” and “Collect star” were relatively high whereas those of “Collect ammo box” and “Crash into enemy” were relatively low (Fig. 1B).

We observed dependencies among behaviors such that the occurrence of some behaviors depended on the preceding occurrence of others. The timing relative to random of “Missile launch button” before the occurrence of the behaviors of “Crash into enemy” and “Hit enemy by missile” was significantly lower than 100% (t-test, p = 10⁻⁸, N = 17), showing that the “Missile launch button” behavior was likely to precede the other two behaviors (Fig. 1C). Similarly, “Missile launch button” occurred after “Missile hit enemy”, “Collect star” occurred after “Crash into wall”, and “Missile hit enemy” occurred after “Missile launch button”. On the other hand, the times of the occurrences of “Collect ammo box”, “Crash into wall”, and “Crash into enemy” were independent of the occurrence of other behaviors, showing a timing relative to random close to 100% (Fig. 1C) and lower F-statistic values (Fig. 1D, see Methods for the interpretation of F-statistics).

Different behaviors are associated with the EEG ISA phases for different brain regions

The degree to which the occurrence of a behavior was modulated by the ISA phase and the phase at which the occurrence frequency was maximized varied with the behaviors and EEG channel locations. For instance, the variation in the occurrence frequency (m_i in Eq. (3) in Methods) of “Crash into wall” was most pronounced at channel CP5, where this behavior occurred most frequently at the ISA phase of $\left[\pi /4,2\pi /4\right]$ rad with the max.–min. occurrence difference (d_{peak-to-troug}) of 98.51%. In contrast, the min. occurrence difference (d_{peak-to-troug}) for “Missile launch button” was 22.73% and maximized at channel F7, showing the most frequent occurrence at $\left[-\pi /4,0\right]$ rad (Fig. 2A). The representative modulation of the occurrence frequency of other behaviors by the ISA phase is also depicted in Fig. 2.

Overall, a stronger modulation by the ISA phase was observed for the “Collect ammo box”, “Crash into wall”, and “Crash into enemy” behaviors relative to others (Fig. 2). It is noteworthy that these three behaviors were less dependent on the occurrence of other behaviors (Fig. 1) and showed stronger associations with the phase of EEG ISA than the other three behaviors more dependent on other behaviors.

Among the three behaviors that showed stronger associations with the ISA phase, “Collect ammo box” was more likely to occur at the maximum phase (i.e., around 0 rad) in the right hemispheric regions. In addition, “Crash into wall” was more likely to occur at the falling ISA phase over the left posterior region, and “Crash into enemy” was more likely to occur at the maximum ISA phase over the left hemispheric region (Figs. 3A and 3B). To represent this tendency of the ISA phase modulation for each behavior, we created a vector of a phase at which the behavior occurred most frequently (Fig. 3B) and scaled the length of the vector by normalized peak-to-trough difference (Fig. 3A). In other words, the angle of the phase vector of a channel is the oscillation phase at maximum behavior occurrence and the magnitude of the phase vector of a channel is the normalized peak-to-through difference of that channel. We collected the phase vectors from all channels (marked by black crosses in Fig. 3C) and calculated the vector sum which was the mean resultant vector. The angle of this vector sum (marked by the red line in Fig. 3C) indicates the dominant phase at which each behavior is modulated. It demonstrated again that “Crash into wall” was mainly modulated at the falling phase whereas the other two behaviors were mainly modulated at the maximum phase (Fig. 3C).