Real-Time Detection and Feedback of Canonical Electroencephalogram Microstates: Validating a Neurofeedback System as a Function of Delay

Abstract

Recent neurotechnology has developed various methods for neurofeedback (NF), in which participants observe their own neural activity to be regulated in an ideal direction. EEG-microstates (EEGms) are spatially featured states that can be regulated through NF training, given that they have recently been indicated as biomarkers for some disorders. The current study was conducted to develop an EEG-NF system for detecting "canonical 4 EEGms" in real time. There are four representative EEG states, regardless of the number of channels, preprocessing procedures, or participants. Accordingly, our 10 Hz NF system was implemented to detect them (msA, B, C, and D) and audio-visually inform participants of its detection. To validate the real-time effect of this system on participants' performance, the NF was intentionally delayed for participants to prevent their cognitive control in learning. Our results suggest that the feedback effect was observed only under the no-delay condition. The number of Hits increased significantly from the baseline period and increased from the 1- or 20-s delay conditions. In addition, when the Hits were compared among the msABCD, each cognitive or perceptual function could be characterized, though the correspondence between each microstate and psychological ability might not be that simple. For example, msD should be generally task-positive and less affected by the inserted delay, whereas msC is more delay-sensitive. In this study, we developed and validated a new EEGms-NF system as a function of delay. Although the participants were naive to the inserted delay, the real-time NF successfully increased their Hit performance, even within a single-day experiment, although target specificity remains unclear. Future research should examine long-term training effects using this NF system.

Keywords: EEG microstates; control; delay; neurofeedback; sense of agency.

FIGURE 1

Schematics of EEGms neurofeedback system. The referenced bipolar signals are processed every 100 ms (10 Hz) into an epoch-averaged spatial pattern. This current state is compared with four EEGms templates on the basis of spatial similarity (Pearson’s correlation). When the largest value among the four absolute correlation coefficients is greater than 0.8 (for example) threshold, the display suggests a green circle in the corresponding area. The targeted state is suggested at the same time using a blue frame. If the green circle overlaps the blue frame, the participants receive a ringing sound as a reward.

FIGURE 2

Canonical EEGms templates as the targeted states in neurofeedback. The microstate maps obtained from external participants under the eyes-closed resting state are shown. (A) Four maps are identified through a typical microstate analysis (upper) based on the GFP (global field power) peak dataset (lower). The sequences of microstate classes were determined by back-fitting to the data with the highest topographical correlation (see text for details). (B) These canonical templates were spatially congruent with previous studies, regardless of the EEG measurement or analysis tools. (C) Polarity-ignored spatial similarity (Pearson’s | r|) with normative templates in which both configurations were interpolated onto 67 × 67 grids.

FIGURE 3

Detected canonical msABCD templates. (A) The exemplified time series of four correlation coefficients with template msABCD at 10 Hz is shown. The yellow circle indicates the detected canonical microstates in real-time (when r_thr = 0.8). (B) If the system is processed at 100 Hz (for comparison), a continuous state-transition dynamic can be observed. In this sense, our time-averaged 10-Hz system detects only spatio-temporally robust microstates.

FIGURE 4

Outputs of the neurofeedback system. (A) A typical scatter plot of a 5-min session (left) showing visually informed (circle feedback) data points only (middle) and auditory-informed “Hits” only (right). (B) If the threshold is lowered (e.g., r_thr = 0.1 for comparison), participants receive too much feedback (visual for the middle panel and auditory for the right panel) to learn.

FIGURE 5

Within- and between-session conditions. (A) A session consists of five within-session target conditions × 3 repetitions. (B) Participants completed 6 sessions (3 between-session delay conditions × 2 repetitions). The order of conditions was randomized.

FIGURE 6

Questionnaire ratings for controllability and the raw number of Hits. (A) Participants reported their subjective feelings regarding “controllability” over the targeted EEGms and their sleepiness on the five-point Likert scale. (B) Participants’ raw Hits performance for each EEGms target. (C) A plot between the number of raw Hits and the Likert rating, including the delay conditions. Error bars indicate ± SE.

FIGURE 7

Participants’ Hit performance as a function of delay. (A) The relative magnification of total Hits in comparison to the individual baseline (no-target condition). (B) The same indices are calculated, respectively, for msA, B, C, and D. Error bars indicate ± 1 SE. *p < 0.05 in ANOVA.

FIGURE 8

System outputs as an asymmetrical matrix. (A) The raw counts for an exemplified session (left) and the summarized matrix for the target specificity are shown in contrast to possible baselines (right). (B) Actual participants’ averages as occupation ratios for each delay condition.

FIGURE 9

Participants’ relative performance for target specificity. (A) The baseline is the no-target period (“task-ready”) under the no-delay condition. (B) The baseline is the session average (the column-means for “task-general”) for each delay condition. The graphs show the relative Hits (self-recursive arrows for target specificity) and Misses (other arrows) where the nodes (A, B, C, and D) indicate the target. (C) The bar plot summarizes (B) with ± 1 S.E. error bars.