---
title: EEG Microstate Analysis Framework (EEG-MSAF)
url: https://www.emergentmind.com/topics/eeg-microstate-analysis-framework-eeg-msaf
type: topic
---

# EEG Microstate Analysis Framework (EEG-MSAF)

Searching arXiv for the specified paper and closely related work to ground the article.
Searching for arXiv:2509.02568.
EEG Microstate Analysis Framework (EEG-MSAF) is an end-to-end pipeline for dementia-oriented electroencephalography analysis that uses EEG microstates—brief (\(\sim 80\)–\(120\) ms), quasi-stable scalp topographies—to identify biomarkers and distinguish dementia (DEM), mild cognitive impairment (MCI), and normal cognition (NC) [2509.02568]. The framework comprises three stages: automated microstate feature extraction, classification with machine learning, and feature ranking using Shapley Additive Explanations (SHAP) to highlight key biomarkers. In the reported formulation, EEG-MSAF is designed to combine accuracy, generalizability, and interpretability, while addressing the temporal and biological-interpretability limitations of conventional spectral or connectivity analyses and many deep learning approaches [2509.02568].

## 1. Conceptual basis

EEG microstates are described as discrete, quasi-stable topographies that are thought to index the momentary activation of large-scale resting-state networks (RSNs) [2509.02568]. In the framework, four canonical microstates, \(A\), \(B\), \(C\), and \(F\), are functionally linked to auditory, visual, salience/attention, and anterior default-mode networks, respectively. This network-level interpretation is central to EEG-MSAF because Alzheimer’s disease (AD) and its prodromal stage, MCI, are characterized by early disruption of these RSNs.

The framework is motivated by the claim that conventional EEG analyses often average over hundreds of milliseconds or seconds, thereby missing fast transitions between brain states and offering limited spatial specificity [2509.02568]. EEG-MSAF instead targets microstate parameters such as duration, occurrence, and coverage, which are presented as direct, physiologically interpretable markers of RSN integrity and as complements to fMRI findings on default-mode network decoherence.

A second motivation concerns model class and sample regime. EEG-MSAF feeds 20 microstate features plus global field power into lightweight, tabular machine-learning models—SVM, RF, and XGB—that are reported to perform well with moderate sample sizes [2509.02568]. The interpretability layer is provided by SHAP, which quantifies each feature’s contribution to multiclass decisions. This suggests a deliberate trade-off: rather than maximizing representational complexity, the framework prioritizes neurophysiological interpretability while retaining strong classification performance.

## 2. Pipeline architecture and preprocessing

EEG-MSAF is organized into three stages: automated microstate feature extraction, machine-learning classification, and SHAP-based feature ranking [2509.02568]. The preprocessing sequence for each subject consists of band-pass filtering at \(0.5\)–\(40\) Hz, average reference, \(z\)-scoring, and cropping of edge artifacts.

The extraction procedure begins with the computation of Global Field Power (GFP) at each time \(t\):
$$
\mathrm{GFP}(t)=\sqrt{\frac{1}{K}\sum_{i=1}^{K}[V_i(t)-\bar V(t)]^2},
$$
where \(V_i(t)\) is electrode \(i\)’s potential and \(\bar V(t)\) is the average across \(K\) channels [2509.02568]. GFP peaks are then identified, and their topographies are extracted.

Microstate segmentation is performed by modified \(k\)-means clustering, using Euclidean distance on normalized maps, at the subject level. The number of classes is fixed at \(k=4\), corresponding to \(A\), \(B\), \(C\), and \(F\), and selected via maximal total Global Explained Variance (GEV) [2509.02568]. Subject-level maps are pooled, and \(k\)-means is rerun to derive group-level templates \(m_1,\ldots,m_4\). Continuous EEG is subsequently back-fitted by assigning each time point to the nearest template according to highest spatial correlation. A temporal-smoothing constraint is applied by merging segments shorter than \(10\) ms into neighboring segments.

The resulting four group-level centroids \(m_k \in \mathbb{R}^K\) define the canonical microstate topographies. Within the framework, these centroids are not merely clustering artifacts; they serve as the reference states from which subject-level temporal, spatial, and occurrence-based descriptors are derived.

## 3. Feature representation

For each microstate class \(k \in \{A,B,C,F\}\), EEG-MSAF computes a fixed set of metrics over \(T\) total time points and \(S_k=\{t \mid \mathrm{state}(t)=k\}\) [2509.02568]. The first is Global Explained Variance:
$$
\mathrm{GEV}_k=\frac{\sum_{t=1}^{T}\mathrm{GFP}(t)^2\cdot r_k(t)^2}{\sum_{t=1}^{T}\mathrm{GFP}(t)^2},
$$
where \(r_k(t)=\mathrm{corr}(x_t,m_k)\).

Mean spatial correlation is defined as
$$
\mathrm{mean\_corr}_k=\frac{1}{|S_k|}\sum_{t\in S_k}\mathrm{corr}(x_t,m_k),
$$
and time coverage as
$$
\mathrm{coverage}_k=\frac{|S_k|}{T}.
$$

Mean duration is computed by
$$
\mathrm{mean\_dur}_k=\frac{1}{N_k}\sum_{i=1}^{N_k} d_i,
$$
where \(N_k\) is the number of continuous segments of class \(k\) and \(d_i\) their durations in ms. Occurrence rate is defined as
$$
\mathrm{occurrence}_k=\frac{N_k}{T\cdot \Delta t},
$$
where \(\Delta t\) is the sampling interval in seconds.

Two additional descriptors are specified. The transition probability matrix is optional:
$$
p_{ij}=\frac{\#\text{transitions from }i\text{ to }j}{\sum_j \#\text{transitions from }i},
$$
and inter-microstate correlation is
$$
\mathrm{corr}_{kl}=\mathrm{corr}(m_k,m_l), \qquad k\neq l.
$$

The reported tabular representation consists of \(4\) classes \(\times 5\) metrics plus \(1\) global GFP, yielding \(21\) features per subject [2509.02568]. The framework’s emphasis on mean correlation and occurrence is especially consequential, because these features are later identified as the most informative biomarkers. A plausible implication is that EEG-MSAF treats both topographic fidelity and recurrence structure as primary signatures of neurodegenerative change.

## 4. Machine-learning and explanation modules

EEG-MSAF evaluates three model families: Support Vector Machine (SVM) with Radial Basis Function kernel, Random Forest (RF), and eXtreme Gradient Boosting (XGB) [2509.02568]. For one-vs-rest class \(k\), the SVM decision function is
$$
f_k(x)=\sum_{i=1}^{N}\alpha_i y_i K(x_i,x)+b,
$$
with
$$
K(x,x')=\exp(-\gamma \|x-x'\|^2),
$$
and hyperparameters \(C \in \{0.1,1,10,100\}\) and \(\gamma \in \{1\mathrm{e}{-4},1\mathrm{e}{-3},5\mathrm{e}{-2}\}\).

The RF model is described as an ensemble of \(M\) trees with majority vote, with grid search over \(\#\mathrm{estimators}\in\{100,200,300\}\) and \(\mathrm{max\_depth}\in\{5,10,15\}\). XGB is formulated as an additive tree model minimizing
$$
L^{(q)}=\sum \ell(y_i,\hat y_i^{(q)})+\sum_{k=1}^{q}\Omega(f_k),
$$
with
$$
\Omega(f)=\gamma \cdot Q+\frac{\lambda}{2}\sum_j w_j^2,
$$
and tuned over \(\#\mathrm{rounds}\in\{100,200\}\), \(\eta\in\{1\mathrm{e}{-4},1\mathrm{e}{-3},5\mathrm{e}{-2}\}\), and \(\mathrm{max\_depth}\in\{3,6,10\}\) [2509.02568].

Evaluation uses five-fold stratified cross-validation, ensuring that no subject appears in both train and test splits; hyperparameters are selected via inner cross-validation. Performance metrics are accuracy,
$$
\mathrm{Accuracy}=\frac{1}{n}\sum I(\hat y_i=y_i),
$$
together with sensitivity (recall), specificity, precision, and F1-score per class and macro-averaged.

Interpretability is provided by SHAP. For a model \(f:\mathbb{R}^d\to\mathbb{R}\) and instance \(x\), the contribution \(\phi_i\) of feature \(i\) is given by
$$
\phi_i=\sum_{S\subseteq N\setminus\{i\}} \frac{|S|!(d-|S|-1)!}{d!}\cdot \left[f_{S\cup\{i\}}(x_{S\cup\{i\}})-f_S(x_S)\right].
$$
For tree-based models, TreeSHAP computes exact \(\phi_i\) in \(O(T\cdot L\cdot D^2)\) time, whereas for SVM, KernelSHAP approximates \(\phi_i\) via weighted linear regression over perturbed feature subsets [2509.02568]. Across both datasets, SHAP consistently ranks \(\mathrm{mean\_corr}_C\), \(\mathrm{occurrence}_C\), \(\mathrm{mean\_corr}_F\), and \(\mathrm{occurrence}_F\) among the top biomarkers, with global GFP occasionally emerging as a summary marker.

## 5. Empirical evaluation and comparative performance

The framework is evaluated on two resting-state EEG datasets [2509.02568]. CAUEEG is a public dataset with 21 channels at 200 Hz and includes 459 NC, 416 MCI, and 311 DEM subjects, with mean age \(70.8\pm 9.9\) years; Fisher’s exact test confirmed age differences across groups (\(p<0.05\)). The Thessaloniki dataset is a clinical cohort with 19 channels, resampled to 100 Hz, comprising 29 CN, 23 FTD, and 36 AD subjects, recorded under eyes-open resting conditions with independent clinical labels.

The reported classification performance is as follows:

| Model | CAUEEG (NC/MCI/DEM) | Thessaloniki (CN/FTD/AD) |
|---|---|---|
| CeedNet (DL baseline) | Acc \(=0.746\) | – |
| EEGConvNeXt (baseline) | – | Acc \(=0.957\) |
| EEG-MSAF-RF | Acc \(=0.65\pm 0.01\), F1 \(=0.65\pm 0.01\) | Acc \(=0.86\pm 0.01\), F1 \(=0.86\pm 0.01\) |
| EEG-MSAF-XGB | Acc \(=0.78\pm 0.02\), F1 \(=0.77\pm 0.01\) | Acc \(=0.73\pm 0.02\), F1 \(=0.71\pm 0.02\) |
| EEG-MSAF-SVM | Acc \(=0.89\pm 0.01\), F1 \(=0.88\pm 0.01\) | Acc \(=0.95\pm 0.01\), F1 \(=0.96\pm 0.01\) |

On CAUEEG, EEG-MSAF-SVM improves accuracy by approximately \(19.3\%\) over CeedNet; on the Thessaloniki dataset, it reaches \(95\%\pm 0.01\) accuracy, comparable to EEGConvNeXt [2509.02568]. Theta-band (\(4\)–\(8\) Hz) features yield the highest multiclass accuracy (\(\approx 0.89\)). All reported improvements are statistically significant according to pairwise McNemar’s test (\(p<0.05\)).

These results are presented as evidence for both generalizability and competitive performance across a public dataset and an independent clinical cohort. Since the same feature family underlies the cross-dataset analysis, a plausible implication is that the discriminative signal is not limited to a single acquisition environment.

## 6. Biomarkers, default-mode decoherence, and early neurodegeneration

The statistical analysis of microstate features identifies \(\mathrm{mean\_corr}_C\) and \(\mathrm{occurrence}_C\) as the strongest differentiators of NC, MCI, and DEM [2509.02568]. Specifically, Kruskal–Wallis tests yield \(\chi^2=81.75\) with \(p<0.0001\) for \(\mathrm{mean\_corr}_C\), and \(\chi^2=180.04\) with \(p<0.0001\) for \(\mathrm{occurrence}_C\). Post-hoc Dunn–Bonferroni comparisons further show that microstate \(C\) and \(F\) metrics are the only ones to distinguish MCI from NC (\(adjusted\ p<0.05\)), marking them as early-stage biomarkers.

The reported neurophysiological interpretation centers on two complementary trajectories. Microstate \(C\), associated with the salience/attention network, shows a progressive decrease in occurrence and spatial coherence (\(\mathrm{mean\_corr}\)) from NC \(\rightarrow\) MCI \(\rightarrow\) DEM, reflecting salience-network breakdown [2509.02568]. Microstate \(F\), linked to the anterior default-mode network, exhibits a complementary “push” in occurrence but decline in coherence, consistent with compensatory hyperactivity followed by network decoherence in AD.

This opposing trajectory is described as mirroring fMRI findings of RSN imbalance in early AD and is termed “default-mode decoherence” [2509.02568]. SHAP ranks \(F\_\mathrm{mean\_corr}\) and \(F\_\mathrm{occurrence}\) among the top features for classifying MCI versus NC, indicating that anterior DMN disruptions can be detected in the prodromal stage. The framework therefore treats mean correlation as an index of microstate coherence and occurrence as an index of engagement frequency. This suggests that EEG-MSAF captures both functional degradation and compensatory reorganization within large-scale brain networks.

## 7. Clinical and methodological significance

EEG-MSAF is presented as a fast, cost-effective, and interpretable tool for early dementia screening, suitable for longitudinal monitoring in outpatient or primary-care settings [2509.02568]. Its clinical rationale derives from the conjunction of three reported properties: automated extraction of well-studied microstate features with established neurophysiological meaning, strong multiclass performance with lightweight tabular models, and transparent SHAP-based explanations.

From a methodological standpoint, the framework is positioned against two limitations: the temporal coarsening of conventional EEG analyses and the black-box character of many deep-learning systems [2509.02568]. Its response is to preserve transient state structure through microstate segmentation while keeping the feature space compact and interpretable. The prominence of microstates \(C\) and \(F\) in both statistical tests and SHAP rankings further ties predictive performance to a biologically articulated account of salience-network breakdown and anterior default-mode disruption.

A common misconception in EEG biomarker research is that interpretability necessarily entails a large loss in predictive accuracy. The reported results do not support that assumption for this framework: the SVM-based configuration attains \(0.89\pm 0.01\) accuracy on CAUEEG and \(0.95\pm 0.01\) on Thessaloniki while maintaining explicit feature-level attribution through SHAP [2509.02568]. Conversely, another potential misconception is that classification alone establishes mechanism. The framework does not eliminate that distinction; rather, it provides neurophysiologically meaningful biomarkers that support hypothesis-driven follow-up, including targeted cognitive interventions or amyloid PET. In that sense, EEG-MSAF functions simultaneously as a diagnostic classifier and as a structured lens on early neurodegenerative network dynamics.

Source: https://www.emergentmind.com/topics/eeg-microstate-analysis-framework-eeg-msaf