Microstate Processing

Spectral analysis tells you what the brain is doing in terms of oscillatory power. Connectivity tells you how regions are coupled. Microstates tell you when the brain switches between distinct spatial configurations—and how long each configuration persists before giving way to the next.

A microstate is a brief period (typically 60–120 ms) during which the brain’s scalp topography remains quasi-stable—the spatial pattern of voltage across electrodes holds roughly steady before rapidly transitioning to a different pattern. These stable periods are not arbitrary segments; they correspond to moments of maximal cortical synchronization, and their temporal dynamics (duration, frequency of occurrence, transition patterns) are altered in a wide range of clinical conditions.

The Coherence Workstation computes microstates using pycrostates, a Python library built on MNE-Python that implements the standard microstate analysis pipeline: GFP peak extraction, modified k-means clustering, canonical map sorting, back-fitting, and temporal statistics.

GFP Peaks: When to Sample

Not every time point in the EEG is equally informative for microstate analysis. At moments of low Global Field Power (GFP), the scalp topography is transitional—the brain is between states, and the spatial pattern is poorly defined. At moments of high GFP, the topography is well-organized—the brain is in a distinct state, and the spatial pattern is clear and classifiable.

Microstate analysis therefore begins by identifying GFP peaks—local maxima in the GFP time series. Each GFP peak corresponds to a moment of maximal topographic organization. The scalp voltage map at each peak is extracted and used as the input to the clustering algorithm. This dramatically reduces the data (from thousands of time points to hundreds or thousands of GFP peaks) while preserving the most informative snapshots of the brain’s spatial dynamics.

Modified K-Means Clustering

The extracted GFP peak maps are clustered into a small number of representative maps using modified k-means. This is similar to standard k-means clustering but with a critical modification: polarity is ignored. A microstate map and its polarity-inverted version (all positive values become negative and vice versa) are considered the same state, because the polarity of scalp EEG depends on the orientation of the underlying cortical dipole rather than on the functional identity of the source. Two time points with identical topography but opposite polarity reflect the same cortical configuration.

The number of clusters is typically set to 4, corresponding to the four canonical microstate classes (A, B, C, D) that have been consistently identified across thousands of subjects in the microstate literature. These four classes explain 65–80% of the variance in resting-state EEG topography—a remarkable degree of regularity given the complexity of the underlying neural dynamics.

The Four Canonical Maps

The four canonical microstate classes have characteristic topographies and functional associations:

Map A shows a left-right diagonal pattern with maximum positivity over the left posterior and maximum negativity over the right anterior regions (or the polarity inverse). Map A is associated with auditory and phonological processing. Its duration and occurrence are altered in conditions affecting language function and auditory attention.

Map B shows the complementary right-left diagonal pattern—maximum positivity over the right posterior and maximum negativity over the left anterior regions. Map B is associated with visual processing and spatial attention. It often shows increased occurrence during visual tasks and altered dynamics in visual processing disorders.

Map C shows a strong anterior-posterior pattern with maximum positivity at frontal midline and maximum negativity at posterior midline (or vice versa). Map C is associated with the default mode network and salience processing. Its temporal dynamics are among the most clinically sensitive—altered Map C duration and occurrence have been reported in depression, schizophrenia, and attention disorders.

Map D shows a frontocentral pattern with maximum positivity at the vertex. Map D is associated with executive function, attention, and cognitive control. Its occurrence tends to increase during focused attention and decrease during mind-wandering.

Map Sorting

The clustering algorithm produces cluster centroids that may not align with the canonical A/B/C/D ordering. Map sorting resolves this by correlating each computed centroid with a set of reference canonical maps and assigning each centroid to its best-matching canonical class. This ensures that “Map A” always refers to the same functional configuration across subjects, enabling meaningful comparisons.

Back-Fitting

After clustering identifies the canonical maps, the pipeline assigns every time point in the original EEG (not just the GFP peaks) to its best-matching microstate class. This is back-fitting—the process of labeling each sample with the canonical map it most closely resembles, using spatial correlation with polarity ignored.

The result is a time series of microstate labels—a sequence like A-A-A-C-C-B-B-B-D-D-A-A—that describes the brain’s moment-to-moment transitions between spatial configurations. This label sequence is the basis for all temporal statistics.

Back-fitting uses a winner-take-all assignment: each time point is assigned to the class with the highest absolute spatial correlation. Time points with very low GFP (below the mean minus one standard deviation) are sometimes left unlabeled, as their topography is too poorly defined for reliable classification.

Temporal Metrics

From the back-fitted label sequence, the pipeline computes four families of temporal statistics:

Duration

The mean duration of each microstate class is the average number of consecutive milliseconds that the brain remains in that state before transitioning. Typical values for resting-state EEG are 60–120 ms, depending on the class and the individual. Longer durations suggest the brain is maintaining that spatial configuration—the underlying neural process is sustained. Shorter durations suggest rapid switching, either adaptive (flexible state transitions) or pathological (failure to maintain stable processing).

Occurrence

The occurrence (or frequency) of each class is how many times per second the brain enters that state. Higher occurrence means the state is frequently visited; lower occurrence means it’s rarely engaged. Occurrence and duration are partially independent—a state can be visited frequently but briefly (high occurrence, short duration) or rarely but for extended periods (low occurrence, long duration).

Coverage

Coverage is the percentage of total recording time spent in each state. It’s the product of occurrence and duration, normalized to the recording length. Coverage tells you how much of the brain’s resting-state activity is dominated by each spatial configuration. In healthy adults, the four canonical states together typically cover 65–80% of the recording, with Maps C and D often showing the highest coverage.

Global Explained Variance (GEV)

GEV measures how well the canonical maps capture the actual EEG topography across the entire recording. It’s the proportion of total spatial variance explained by the assigned microstate labels. Higher GEV (>0.65) indicates that the four-class solution is a good summary of the brain’s spatial dynamics. Lower GEV suggests either that the brain’s topography is more complex than four states can capture, or that the recording is contaminated with artifact that doesn’t fit any canonical pattern.

Transition Probabilities

The transition probability matrix describes how likely the brain is to move from one state to each other state. Each row of the matrix sums to 1.0, and each cell represents the probability of transitioning from state i to state j. Non-random transition patterns—for example, a strong tendency to go from A to B but rarely from A to D—reveal sequential structure in the brain’s spatial dynamics. These transition patterns are altered in various clinical conditions and may reflect the brain’s preferred processing sequences.

Clinical Significance

Microstate temporal dynamics are among the most sensitive EEG biomarkers available. The research literature has reported altered microstate parameters in schizophrenia (reduced Map C duration), depression (altered Map C and D coverage), ADHD (altered transition probabilities), Alzheimer’s disease (reduced overall duration), and many other conditions.

The Coherence Workstation reports microstate metrics as part of the comprehensive analysis but does not generate diagnostic claims from them. Like all analysis layers, microstates are one source of structural information that gains clinical meaning only when integrated with the spectral, connectivity, and ERP findings—and with the clinician’s knowledge of the patient.