The EEG microstate approach models EEG signals as discrete, non-overlapping topographical patterns^[24,50,51], which are linked back to the original data through spatial correlation techniques^[52]. This method interprets EEG data as sequences of distinct topographies^[53], and has been effectively applied in analysing differences in patterns in a variety of applications. Notably, it adapts varying EEG channels by standardising datasets into a consistent feature set, enabling the model to accept heterogeneous electrode configurations without architectural modification.

Figure 1 presents the microstate extraction workflow. The raw EEG signals collected by an amplifier (after pre-processing steps as detailed in Section 3) are then processed with GFP^[54], computed for each time point using the following formula:

(1) $$ GFP(t) = \sqrt{\frac{\sum_{i=1}^{n} (v_i(t) - \bar{v}(t))^2}{n}}, $$

where $$ \bar{v} $$_i(t) represents the voltage at electrode i, v(t) is the average voltage across all electrodes, and n denotes the total number of electrodes. Topographical maps corresponding to the GFP peaks, where the signal-to-noise ratio (SNR) is maximised, are extracted and clustered using a modified k-means clustering algorithm. To quantify the similarity between maps, the GMD^[55] is employed, defined as:

(2) $$ GMD_{u,v} = \sqrt{\frac{1}{n} \sum_{i=1}^{n} \left( \frac{u_i}{GFP_u} - \frac{v_i}{GFP_v} \right)^2}, $$

where u and v are topographical maps, and GFP_u and GFP_v are their corresponding GFP values. Four standard microstates (A, B, C, D) are identified, associated with key neural networks: auditory, visual, salience, and attention^[56]. These microstates were then used to reconstruct the original EEG signal, forming the input sequence for the proposed model. It results in a sequence of 12,000 characters with A, B, C, D, or E (a small number of topographical maps that are unclassified among A, B, C, or D).

The microstate approach does not explicitly decompose EEG activity into anatomical brain regions (e.g., frontal vs. temporal). Instead, region-specific information is implicitly encoded in the topographic shape of each microstate: a frontally dominant topographic pattern will contribute positively to frontal-electrode GFP values and will be assigned to whichever microstate cluster best captures that spatial distribution. The four canonical microstates (A-D) are each associated with distinct distributed brain networks: auditory, visual, salience, and dorsal attention, respectively^[57], which inherently involve different combinations of cortical regions. This approach allows the model to operate consistently across different numbers of EEG channels: the same microstate labels and feature set can be derived regardless of whether 19, 64, or 128 electrodes are used, at the cost of not providing region-specific interpretability.

A key property of the microstate pipeline is its natural reconciliation of heterogeneous channel configurations. Unlike conventional approaches that require a fixed-size input tensor corresponding to a specific electrode montage, the microstate framework operates as follows: GFP is computed across all available electrodes [Equation (1)] at each timepoint, whether 19, 64, or 128. Topographic clustering then identifies the dominant spatial patterns within each dataset’s electrode space independently. The resulting microstate sequence (A, B, C, D, E) is invariant to the number of input channels because it describes temporal patterns of dominant topography rather than specific electrode amplitudes. The 40 extracted features are therefore structurally identical across all four datasets, enabling joint model training and evaluation despite the heterogeneous hardware configurations.

As shown in Figure 1, this research utilises four essential EEG microstate properties as authenticating features: occurrence, coverage, duration, and transition probabilities. Occurrence represents the average frequency at which a particular microstate class is observed per second. Coverage denotes the proportion of time, expressed as a percentage, that a specific microstate class occupies within a second. Duration corresponds to the mean length of time a particular microstate class persists during a single occurrence. Transition probabilities measure the likelihood of shifting from one microstate class to another, capturing the dynamic changes and interactions between different brain states over time. To illustrate, consider microstate A. It is characterised by three primary features, occurrence, coverage, and duration, alongside its transition probabilities, which include transitions from A to A, A to B, A to C, A to D, and A to E. Together, these yield eight distinct features for microstate A. When extended to all four microstates (A, B, C, D) and E, this results in a total of 40 features serving as inputs for user authentication.

Firstly, regarding the validation approach, as described in Section 3, the dataset consists of 1,015 samples collected from 203 participants, with each participant contributing 5 samples. To illustrate, these can be denoted as S₁, S₂, S₃, S₄, and S₅. For instance, participant I₁ has his/her own S₁, S₂, S₃, S₄, and S₅, and similarly for all other participants. We use ten validation splits so that all samples are evaluated.

In each split, two of the five samples from every participant are used for training, while the remaining three samples are reserved for testing. This ensures that the testing set in each split contains three of the five samples from every participant. Importantly, the assignment of samples to training or testing need not be identical across all participants. However, the ten splits must ensure that the training and testing samples for each participant differ across splits. For example, in the first split, S₁ and S₂ might be used for training while S₃, S₄, and S₅ are used for testing. In subsequent splits, this allocation changes, maintaining the consistent 2:3 ratio while introducing diversity.

Each split’s testing phase has three test sets: one using normal samples and two additional sets generated through adversarial attacks. Model performance metrics are explained in Section 5.2, and fairness metrics (Section 5.3) are calculated for all test sets in every split. Finally, the results from all ten splits, the average and the standard deviation of the metrics, are reported to provide a comprehensive evaluation of model performance and fairness.

The adversarial robustness evaluation targets a white-box feature-space attack scenario. Specifically, the threat model assumes an adversary who: (a) has full knowledge of the trained EchoMC meta-model and its parameters; (b) operates at the feature level, i.e., perturbs the extracted 40-dimensional microstate feature vectors rather than raw EEG signals; and (c) aims to cause misclassification by applying minimal perturbations within an ℓ_∞ ball of radius ϵ = 0.01, using fast gradient sign method (FGSM) and projected gradient descent (PGD). This simulates an attack targeting the feature representation rather than raw signals. Importantly, this threat model does not cover physical-domain attacks or hardware-level signal corruption. Non-adversarial stress tests (e.g., channel dropout, noise injection, site shift) were not conducted in this study but represent important future directions for validating robustness under realistic hardware variability.

To evaluate the developed model, five important metrics are employed: Accuracy, Recall (macro), Precision (macro), F1-score (macro), and expected calibration error (ECE) with a bin number of 10, following the reference studies^[64]. Their values range from 0 to 1. Higher values indicate better models, except for ECE, where a lower value means a better model.

Rather than simply calculating metrics across the entire test set and reporting average results, fairness evaluation focuses on the demographic characteristics of the samples within the test sets. These demographic attributes, referred to as sensitive attributes^[65], play a crucial role in assessing model fairness. In our research, sensitive attributes include Gender (male and female) and AD status (NC and AD), as we target to generalise the model for NC and AD users and gender-related biases in biometric model predictions are frequently observed^[21]. This issue is particularly evident in AI-driven BCI systems using EEG^[22].

To enable a more granular fairness evaluation, we move beyond examining a single sensitive attribute, such as comparing male vs. female or NC vs. AD independently. Instead, we consider multiple sensitive attributes simultaneously^[65], considering combinations of these attributes. The comparisons are structured into eight groups within each test set and across each split: G₁ compares NC and AD; G₂ distinguishes between Male and Female participants; G₃ contrasts Male NC with Male AD; G₄ examines Male NC against Female NC; G₅ compares Male NC with Female AD; G₆ differentiates Male AD from Female AD; G₇ contrasts Female NC with Male AD; and G₈ examines Female NC vs. Female AD.

For transparency, the approximate subgroup sample sizes per split test set (3 samples per participant × number of participants) are as follows: Male NC participants: approximately 201 (67 participants × 3), Female NC participants: approximately 240 (80 participants × 3), Male AD participants: approximately 63 (21 participants × 3), and Female AD participants: approximately 105 (35 participants × 3). Note that exact counts vary slightly across splits due to the randomised 2:3 sample allocation. These counts inform the interpretation of overall accuracy equality (OAE) and calibration using ECE (ΔECE) values, as smaller subgroup sizes yield wider variance in the reported metrics, reflected in the standard deviations of the fairness results reported in Section 6.1.

About the fairness metrics, OAE and ΔECE are utilised for comprehensive evaluation. For example, consider a scenario where the results of model predictions for a specific split are available. These results include the sample index, ground truth labels, predicted labels, confidence scores, gender (male or female), and AD status (NC or AD). For group comparison G₁, which compares sub-group A (NC) and sub-group B (AD), we first filter the data to calculate the Accuracy and ECE (as detailed in Section 5.2) for each sub-group separately. Next, we conduct a fairness evaluation by calculating the ratio of the difference between these two sub-groups using the following equations with results ranging from 0 to 1:

(3) $$ OAE = \frac{|\text{Accuracy}_A - \text{Accuracy}_B|}{\max(\text{Accuracy}_A, \text{Accuracy}_B)} \leq \gamma $$

(4) $$ \Delta ECE = \frac{|\text{ECE}_A - \text{ECE}_B|}{\max(\text{ECE}_A, \text{ECE}_B)} \leq \gamma $$

Lower OAE and ΔECE are better in fairness. This study adopts an acceptance threshold γ of 0.2, established by the “80% rule”^[66] specifying that the performance of one group must reach at least 80% of the performance achieved by the other group.

The hyperparameters of the EchoMC network, along with its meta-models, are detailed in Table 1. Notably, the hyperparameters for the meta-models and the models trained without the base model remain consistent across all methods used in this research. The selected meta-models were chosen based on their proven effectiveness in the literature^[5]. Regarding the adversarial attacks, ϵ is set at 0.01. The software and libraries used in this study include the following: EEG microstate extraction was performed using the Pycrostates library^[67]. Data visualisation was carried out using Matplotlib and Seaborn, while t-distributed stochastic neighbour embedding (t-SNE) was implemented using the scikit-learn library. To ensure full reproducibility, all hyperparameters for both base models and meta-models are reported in Table 1, and all software libraries used are listed above. No additional components beyond those described are required to replicate the experimental pipeline.

Detection of variations accounting for repeated measuresTo assess changes in feature distributions across time points (irrespective of AD status), we applied the non-parametric Friedman test, which is well-suited for repeated measures without assuming normality. This allowed for the identification of temporal effects across sessions at the group level. To control for the multiplicity of comparisons and assess the significance of broader feature domains, we grouped related features into four categories, consisting of occurrence, coverage, duration, and transition - and combined P-values within each group using established statistical techniques.

Three methods were used for P-value combination: Fisher’s method^[68], which aggregates P-values using the sum of their logarithms under the assumption of independence, yielding a chi-squared statistic; Stouffer’s method^[69], which transforms P-values into z-scores and combines them within the standard normal distribution framework; and the Bonferroni correction^[70], a conservative adjustment that scales individual P-values by the number of tests to control the family-wise error rate in the presence of potential dependencies.

Detection of variations based on AD statusTo further investigate differences within each group, we stratified the data by AD status and repeated the same analysis for each subgroup (AD and NC). This stratified evaluation enabled the identification of intra-group temporal variations and facilitated a clearer understanding of how each group individually evolves.

Detection of variations over time and between AD and NCTo examine both time-dependent changes and group-specific effects simultaneously, we conducted a two-way repeated-measures analysis of variance (ANOVA) with one within-subjects factor (time) and one between-subjects factor (AD status). This analysis assessed the main effects of time, group, and their interaction, providing insight into whether temporal trends differed significantly between AD and NC groups. P-values resulting from these ANOVAs were again combined within the four predefined feature groups using Fisher’s, Stouffer’s, and Bonferroni’s methods, as described above. All analyses were implemented in Python using the scipy.stats and pingouin libraries^[71].

To supplement statistical significance, effect sizes, specifically Kendall’s W for the Friedman tests and partial eta-squared (η_p²) for the two-way repeated measures ANOVA, are recommended as complementary measures of practical significance. Reporting these alongside the presented P-values in future extensions of this work would provide a standardised indication of effect magnitude independent of sample size, following best practice in statistical reporting^[72].

Global hypothesis testingThe objective of the P-value combination strategies was to test the global null hypothesis {H₀: ∩_i=1^kH_0i}, which assumes that all individual null hypotheses H_0i are true. Each p_i corresponds to a feature-specific hypothesis.

Fisher’s method combines P-values assuming p_i ~ $$ \mathcal{U} $$(0, 1). The test statistic is computed as:

$$ X^2 = -2 \sum_{i=1}^{k} \ln(p_i), $$

where X² ~ χ²(2k) under H₀. The combined P-value is obtained as p_c = 1 - F(X², 2k), where F(·, 2k) denotes the cumulative distribution function of the chi-squared distribution with 2k degrees of freedom.

Stouffer’s method combines P-values by converting them into z-scores:

$$ z = \frac{1}{\sqrt{k}} \sum_{i=1}^{k} \Phi^{-1}(1 - p_i), $$

where Φ^-1 is the inverse of the standard normal CDF, and z ~ $$ \mathcal{N} $$(0, 1) under H₀. The combined P-value is then p_c = 1 - Φ(z).

Bonferroni’s method adjusts for multiple testing by conservatively scaling the smallest P-value:

$$ p_c = \min\left(1, \min(p_1, \ldots, p_k) \times k\right). $$

This approach does not assume independence and effectively controls Type I error in the presence of correlated tests.

To supplement statistical significance, effect sizes are reported alongside P-values. For the two-way repeated measures ANOVA, partial eta-squared (η_p²) is computed as:

$$ \eta_p^2 = \frac{SS_{\text{effect}}}{SS_{\text{effect}} + SS_{\text{error}}}, $$

where values of 0.01, 0.06, and 0.14 correspond to small, medium, and large effects respectively. For the Friedman tests, Kendall’s W is reported as a measure of concordance across repeated samples, ranging from 0 (no agreement) to 1 (complete agreement). These effect size measures provide a standardised indication of practical significance independent of sample size.

To interpret the results of the two-way repeated measures ANOVA, we focused primarily on the interaction effect between time and AD status. A significant interaction indicates that the effect of time on feature distributions differs between the AD and NC groups. In such cases, interpreting the main effects independently may be misleading, as group differences may be driven by divergent temporal patterns.

If the interaction term was not significant, we interpreted the main effects separately. A significant main effect of time suggests that feature values changed consistently over time across both groups. A significant main effect of group (AD vs. NC) indicates an overall difference in feature values between the two groups, averaged across time points.

This analytic framework allows us to disentangle group effects, temporal dynamics, and their interactions, facilitating a nuanced understanding of how cognitive status modulates changes in EEG-derived features over time.

Regarding performance, Table 2 presents the results of various meta-models for the proposed EchoMC network. Among them, RF, along with XGBoost, DT, CatBoost, and LightGBM, achieved the highest accuracy and recall, both reaching 0.8446. Regarding F1-score, LightGBM leads with a score of 0.8820, indicating balanced performance across precision and recall. Notably, DT outperforms other models in precision, with a value of 0.9724, and achieves the lowest ECE of 5.92e-19, reflecting its high calibration performance. The most limited method is SVM, which has lower metrics than the others. However, it is vital to note that the gap between RF results and DT and LightGBM, achieving the highest precision and F1-score, is just 0.01% for both. When testing with adversarial attacks, with results in Table 3, the EchoMC network remains largely unaffected by these attacks.

It is important to be precise about what Table 4 demonstrates and what it does not. Table 4 shows that meta-models trained directly on the 40 microstate features, without the EchoMC base model, achieve accuracy below 0.35 across all meta-models, confirming that the stacked ensemble structure is essential and that neither the meta-model choice nor the microstate features alone are sufficient for strong performance. This provides evidence that the gain is attributable to the learned intermediate representations produced by the EchoMC base model. However, Table 4 does not isolate: (a) the contribution of microstate features relative to conventional EEG features such as band-power or connectivity measures; (b) the individual contributions of MHA, 1D-CNN, and ESN as standalone classifiers; or (c) the stacked architecture relative to a non-stacked deep baseline of equivalent capacity. These ablations are necessary to fully attribute the observed gains and are committed to as future work.

About fairness, as observed in Table 5 and Figure 2, all meta-models demonstrate fairness with the OAE metric, except GNB, k-NN, LR, and SVM. This trend is the same regardless of testing on regular or adversarial attacks [Table 6]. These four models exhibit the lowest performance metrics, as shown in Table 2. This suggests a potential relationship between lower predictive performance and increased unfairness in model outcomes. However, when evaluating fairness using calibration (ΔECE) with normal and adversarial attacks, as detailed in Table 7 with Figure 3 and Table 8, respectively, it reveals that only RF achieves fairness across all groups, with no group exceeding the acceptance threshold γ of 0.2. Therefore, although RF does not achieve the highest performance metrics, its ability to maintain fairness while delivering competitive performance results makes it the preferred meta-model among those evaluated as the meta-model for the proposed EchoMC network.

In comparison to existing studies, Table 9 provides a comprehensive summary of the proposed Bio-MEEG framework with existing methods. Although our framework may not achieve the highest accuracy, its reliability is strengthened by several key factors. As shown in Table 9, no prior study simultaneously includes both fairness evaluation and AD participants. Previous studies overlook AD users despite their differences from NC users (explained in Section 6.2). In contrast, our EchoMC network, a core component of Bio-MEEG, is trained on 203 NC and AD participants from diverse datasets across multiple countries, supporting multi-dataset evaluation across diverse acquisition settings, though subject-independent evaluation remains to be demonstrated. Thirdly, Bio-MEEG achieves fairness within acceptable thresholds in predictions using multiple sensitive attributes, making it a fair AI model for NC and AD participants. Additionally, adversarial attack assessments have demonstrated Bio-MEEG robustness, with results remaining almost unchanged under attacks.

Furthermore, while Bio-MEEG demonstrates promising authentication performance, its real-world deployment faces several practical constraints. First, resting-state EEG acquisition with wet electrodes requires trained personnel for electrode placement and gel application, typically taking 20-45 min per session. Second, the eyes-closed resting-state paradigm requires approximately 60 s of cooperative signal acquisition, significantly longer than fingerprint- or face-based authentication. Third, EEG systems with 19-128 channels represent clinical-grade equipment that is neither portable nor consumer-accessible in current form. For AD users specifically, the primary motivation for this framework, these constraints are partially mitigated by the clinical context: EEG systems are already present in neurological care facilities where AD patients regularly attend. Consumer-facing deployment would require adaptation to low-channel-count, dry-electrode, portable systems, for which the microstate approach is architecturally suited, given its channel-count agnosticism, but it would require dedicated validation.

The statistical feature analysis revealed several key findings. First, significant differences were observed across the five intra-participant samples (i.e., repeated measures), with all P-values for occurrence, coverage, and duration features consistently below 0.001 [Table 10]. Furthermore, 18 out of 25 transition features demonstrated P-values below 0.05. These results were further validated through P-value combination techniques applied to grouped features, confirming strong overall significance across feature categories [Table 11]. These findings suggest that samples collected from the same participant are not always stable, underscoring the necessity of accounting for intra-participant variability when designing and training models.

When stratifying the data by AD status, statistically significant differences were also observed within both the AD and NC groups. Across key feature types of occurrence, coverage, duration, and transitions, individual and grouped P-values consistently indicated significant within-group temporal variability [Table 11]. This highlights that even within homogeneous clinical groups, temporal fluctuations in EEG-derived features are evident and relevant for downstream model performance.

Considering both intra-subject variability (time) and inter-subject differences (AD status), results from the two-way repeated measures ANOVA [Table 12] identified several significant interaction effects. These findings indicate that changes in feature distributions over time were modulated by clinical group, i.e., AD and NC participants exhibited distinct temporal trajectories. When P-values were combined across grouped features [Table 13], interaction effects remained significant for coverage, duration, and transition features, demonstrating that AD status meaningfully shaped the observed longitudinal patterns.

These results collectively demonstrate that temporal changes in EEG-based features cannot be interpreted independently of AD status. The significant interaction effects confirm that group-specific neurophysiological dynamics influence how features evolve. Therefore, when building classification or regression models, it is critical to account for both within-subject variability and between-group differences, as ignoring either factor could obscure meaningful patterns or lead to overfitting. The feature distribution visualisations in Figures 4-6 further illustrate the within-participant diversity and between-group separation across the extracted EEG features.