The decoding of imagined speech from EEG has been investigated using both traditional machine learning and deep learning approaches. Early work by DaSalla et al. [7] demonstrated the feasibility of classifying two imagined vowels (/a/ vs. /u/) using EEG with CSP and Support Vector Machines (SVM), achieving approximately 70% binary classification accuracy within a single session. Brigham and Kumar [8] extended this to imagined syllables using spectral features and achieved above-chance classification for binary tasks. Nguyen et al. [9] employed Filter Bank Common Spatial Pattern (FBCSP) combined with regularized linear discriminant analysis for multi-class imagined speech, reporting accuracies in the range of 35–50% for 4–5 classes within-session.

Power Spectral Density (PSD) features in theta (4–8 Hz), alpha (8–13 Hz), and beta (13–30 Hz) frequency bands have been identified as informative for imagined speech discrimination [23]. Wavelet-based features have also shown utility, capturing transient spectral dynamics associated with the temporal unfolding of imagined speech [24]. However, these feature-engineering approaches require domain expertise, may not capture complex nonlinear interactions in the EEG data, and have shown limited generalization to new sessions or subjects.

Deep learning methods have more recently been applied to Speech Imagery EEG. Saha et al. [24] applied convolutional neural networks to classify imagined speech and demonstrated improvements over traditional approaches but noted substantial performance degradation in cross-subject evaluation. Lee et al. [25] explored Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks for temporal modeling of imagined speech EEG, showing that sequential modeling could capture temporal dynamics but at the cost of computational complexity. EEGNet [13] has been widely adopted as a baseline architecture for various EEG classification tasks, including Speech Imagery, owing to its compact design and cross-paradigm versatility. Attention mechanisms have been explored by Nieto et al. [26], who demonstrated that temporal attention could selectively weight informative time segments in imagined speech trials, improving classification of imagined words. Despite these advances, the state-of-the-art for multi-class (4–6 classes) imagined speech classification remains at 30–50% within-session accuracy and 25–40% cross-session, highlighting the fundamental difficulty of this decoding task.

The problem of distributional shift between EEG recording sessions has been extensively studied in the motor imagery BCI literature, though far less so for Speech Imagery. Wu et al. [27] provided a comprehensive survey of transfer learning methods for EEG-based BCIs, categorizing approaches into instance-based, feature-based, and model-based transfer. He and Wu [28] proposed Euclidean Alignment (EA), a simple but effective method that aligns the mean covariance matrix of EEG trials across sessions or subjects to a common reference, demonstrating substantial improvements in cross-session motor imagery classification. Barachant et al. [29] developed Riemannian geometry-based approaches that operate on the manifold of symmetric positive definite matrices (EEG covariance matrices), providing geometrically principled alignment and classification that is inherently robust to affine transformations of the EEG data.

Deep domain adaptation methods have also been applied to BCI transfer. Ozdenizci et al. [30] employed adversarial domain adaptation for cross-subject EEG transfer, training a domain discriminator alongside the main classifier to learn domain-invariant features. Long et al. [31] introduced deep adaptation networks using MMD for domain adaptation in general machine learning, which has subsequently been adopted for EEG applications. Zhao et al. [32] proposed multi-source domain adaptation for motor imagery BCI, aggregating information from multiple source subjects to improve transfer to a new user. Session-to-session style transfer and normalization techniques [33], [34] have shown promise in reducing cross-session variability, but these methods have been predominantly validated on motor imagery tasks with relatively strong and focal neural signals. The application of domain adaptation to Speech Imagery EEG, where the signals are weaker and more distributed, remains largely unexplored.

NAS automates the process of designing neural network architectures, replacing the conventional approach of manual architecture engineering. The field was catalyzed by Barret and Le Quoc [16], who used reinforcement learning to train a controller RNN to generate neural architecture descriptions, achieving competitive performance on image classification benchmarks albeit at enormous computational cost (800 GPU-days). Subsequent work dramatically reduced this cost: DARTS [17] reformulated architecture search as a continuous optimization problem by relaxing the discrete architecture space, enabling gradient-based search that completes in a single GPU-day. Real et al. [18] demonstrated that simple evolutionary methods, particularly aging evolution, which removes the oldest individuals from the population regardless of fitness, could match or exceed the performance of reinforcement learning-based NAS with better computational efficiency and simplicity.

Hardware-aware NAS methods [27], [35] introduced multi-objective optimization to jointly optimize accuracy and hardware efficiency metrics (latency, FLOPs, parameter count), producing models suitable for deployment on resource-constrained devices. One-shot NAS methods [36], [37] amortize the cost of architecture evaluation by training a single supernet that subsumes all candidate architectures, enabling rapid architecture scoring without individual training. These advances in general NAS have not been systematically translated to the EEG/BCI domain.

NAS for EEG remains nascent. Sun et al. [19] proposed NASIL for motor imagery EEG classification, applying a simplified NAS procedure with a limited search space, but did not address cross-session transfer or Speech Imagery. Zhu et al. [20] introduced AutoEEG, an automated model design framework for general EEG classification that employed differentiable search, but focused on within-subject evaluation and did not incorporate domain adaptation or cross-session objectives. To our knowledge, no NAS framework has been proposed for Speech Imagery EEG decoding, and no NAS approach in the BCI domain has incorporated cross-session generalization as an explicit optimization target or included searchable domain adaptation modules within the architecture space.

Metaheuristic optimization encompasses a family of nature-inspired, population-based algorithms that search for optimal solutions in complex, high-dimensional spaces without requiring gradient information. Key algorithms include Genetic Algorithms (GAs) [38], particle swarm optimization [39], DE [40], simulated annealing [41], and ant colony optimization [42]. These methods suit the combinatorial nature of the architecture search problem, where the search space is discrete, non-differentiable, and multi-modal.

Neuroevolution, the application of evolutionary algorithms to optimize neural network architectures, weights, or both, has a rich history [43]. NeuroEvolution of Augmenting Topologies (NEAT) [44] evolved both topology and weights of neural networks through complexification. More recently, Large-Scale Evolution of image classifiers [45] and AmoebaNet [18] demonstrated that evolutionary methods can discover competitive architectures at scale. In the BCI domain, metaheuristic optimization has been applied primarily for feature selection and hyperparameter tuning: GAs for CSP filter selection [46], particle swarm optimization for SVM hyperparameters [47], and DE for optimizing neural network weights in EEG classification [48]. However, full neural architecture optimization via metaheuristics for EEG decoding remains unexplored, particularly with EEG-specific architectural components and cross-session objectives. MetaNAS-SI bridges this gap by casting the cross-session Speech Imagery architecture design problem as a multi-objective metaheuristic optimization problem with an EEG-tailored search space.

Let a Speech Imagery EEG dataset consist of recordings from S sessions for a given subject, where each session s contains N labeled trials of K-class imagined speech. Each trial is represented as a matrix X ∈ ℝC×T, where C denotes the number of EEG channels, and T denotes the number of temporal samples. The associated class label is y ∈ {1, 2, …, K}.

The objective of MetaNAS-SI is to discover an architecture A* and corresponding trained weights W* that maximize cross-session classification performance. We formalize this through a LOSO protocol:

A* = arg max (1/S) Σ S Acc(A, W , D s | D −s), (1)

A ∈ s=1 A test train

where D −s = {D1, …, Ds−1, Ds+1, …, DS} denotes the training data comprising all sessions except session

train

s, and D s denotes the test data from the held-out session s. This LOSO cross-session accuracy serves as

test

the primary fitness function for the NAS procedure, directly rewarding architectures that generalize across sessions.

Simultaneously, we seek to minimize the model parameter count |A| (for efficient real-time BCI deployment) and inference latency L(A). The complete multi-objective formulation is:

maximize {AccLOSO(A), −|A|, −L(A)}. (2)

Eq. (2) yields a Pareto front of non-dominated architectures, from which deployments can select based on application-specific constraints (Section 3.6).

A central contribution of MetaNAS-SI is the design of a search space that incorporates neurophysiologically informed operations tailored to the characteristics of EEG signals. The search space consists of four categories of operations, a macro-architectural skeleton, and the SIFAM module.

MetaNAS-SI employs a two-level hybrid evolutionary search strategy that decomposes the architecture optimization into macro-architecture search (discrete, combinatorial) and micro-architecture optimization (continuous, numerical), each addressed by an appropriate metaheuristic algorithm.

The SIFAM is a key architectural component that performs unsupervised domain adaptation between EEG sessions within the network. Unlike post-hoc domain adaptation methods that are applied independently of the classification architecture, SIFAM is integrated within the candidate architecture, and its internal structure is jointly optimized by the NAS alongside the main classifier, enabling the search process to discover the optimal configuration for aligning session-specific distributions while preserving class-discriminative information.

SIFAM operates on intermediate feature representations. During training, let F = {f src, …, f src} denote the

src 1 n

features extracted from source session(s) at the SIFAM insertion point, and F = {f tgt, …, f tgt} denote

tgt 1 m

features from the target session (available as unlabeled data during training). SIFAM minimizes the distributional discrepancy between these feature sets using the MMD criterion:

LSIFAM = λMMD · MMD²(Fsrc, Ftgt), (3)

where the squared MMD is computed as:

MMD²(P, Q) = (1/n²) Σi,j k(pi, pj) − (2/nm) Σi,j k(pi, qj) + (1/m²) Σi,j k(qi, qj). (4)

The kernel function k is selected from three options as part of the NAS search: 1) a linear kernel k(x, y) = xTy, 2) a Gaussian RBF kernel k(x, y) = exp(−||x − y||² / 2σ²) with bandwidth σ determined by the median heuristic (σ = median of pairwise distances), and 3) a multi-kernel combination:

kmulti(x, y) = Σb αb · kRBF(x, y; σb). (5)

With bandwidths σb selected as multiples of the median heuristic ({0.5σ, σ, 2σ}) and uniform weights αb = 1/3. The total training loss combines classification and alignment objectives:

Ltotal = Lclassification + LSIFAM, (6)

where Lclassification is the standard categorical cross-entropy loss. The alignment strength λMMD is searchable during the NAS process from {0.01, 0.05, 0.1, 0.5, 1.0}, as is the number of alignment layers within SIFAM (1, 2, or 3 fully connected layers that transform features before MMD computation) and the position of SIFAM within the main architecture (after cell 2, 4, or 6).

By making SIFAM’s configuration part of the architecture search, MetaNAS-SI discovers the optimal domain adaptation strategy jointly with the classification architecture, a capability absent from all prior work that treats domain adaptation as a separate, manually configured post-processing step.

All EEG data undergo a standardized preprocessing pipeline before entering the network. Continuous EEG recordings are band-pass filtered between 0.5 and 45 Hz using a 4th-order zero-phase Butterworth filter to remove DC drift and high-frequency noise while preserving the physiologically relevant EEG frequency bands. A Common Average Reference (CAR) is applied to reduce common-mode noise. Epochs are extracted from −0.5 to 2.0 seconds relative to the Speech Imagery cue onset, yielding 640-sample epochs at 256 Hz. Baseline correction is performed by subtracting the mean amplitude of the pre-stimulus period (−0.5 to 0 seconds) from each channel. Trials with peak-to-peak amplitude exceeding 100 μV on any channel are rejected as artifacts. Finally, z-score normalization is applied independently per channel per session to partially reduce session-level amplitude differences.

Data augmentation is applied during training to increase the effective training set size and improve generalization. Four augmentation strategies are employed, with their respective strengths optimized as continuous hyperparameters by the DE search in Level 2:

I. Gaussian noise injection: additive Gaussian noise with standard deviation σnoise, searched in the range [0.01, 0.1], is added to each trial independently.

II. Temporal jittering: random temporal shifts of ±5 samples (approximately ±20 ms) are applied to each trial, simulating slight variations in imagery onset timing.

III. Channel dropout: 1–3 channels are randomly zeroed in each trial during training, encouraging the network to be robust to individual channel noise or loss.

IV. Sliding window cropping: random temporal crops of 80–100% of the trial length are extracted and zero-padded to the original length, providing temporal variability.

After the two-level NAS completes, the set of evaluated architectures yields a Pareto front of non-dominated solutions across three objectives: cross-session LOSO accuracy, model parameter count, and inference latency. An architecture is Pareto-optimal (non-dominated) if no other architecture is simultaneously better on all three objectives. The Pareto front provides a menu of architectures from which practitioners can select based on deployment requirements:

I. Maximum accuracy: the architecture with the highest cross-session accuracy, suitable for offline analysis applications where computational resources are unconstrained.

II. Balanced (knee-point): the architecture at the knee point of the Pareto front, defined as the point with minimum normalized Euclidean distance to the ideal point (maximum accuracy, minimum parameters, minimum latency). Such a selection represents the best trade-off for general-purpose BCI applications. III. Minimum latency: the architecture with the lowest inference latency that still exceeds a minimum accuracy threshold (chance level + 10 percentage points). Such an architecture is suitable for real-time BCI applications requiring <50 ms inference.

The Pareto-optimal selection is performed using the non-dominated sorting procedure from NSGA-II [55], applied as a post-processing step after all architecture evaluations are complete.

The total computational cost of MetaNAS-SI is summarized as follows:

All experiments are conducted on an NVIDIA RTX 3090 GPU (24 GB VRAM). The 90 GPU-hour NAS cost is a one-time investment per dataset and Speech Imagery paradigm. Once the optimal architecture is discovered, it is fixed and can be trained for new subjects in approximately 1.5 hours, making MetaNAS-SI practical for clinical BCI deployment.

MetaNAS-SI was evaluated on three Speech Imagery EEG datasets spanning different languages, channel configurations, and class counts to assess generalization across diverse experimental conditions.

Two evaluation protocols are employed, with cross-session LOSO serving as the primary metric:

I. Within-session (baseline comparison): 5-fold stratified cross-validation within each session. Performance is averaged across folds and sessions. Such an evaluation serves as an upper-bound reference representing the best-case scenario without session transfer.

II. Cross-session LOSO (primary metric): LOSO evaluation. For a subject with S sessions, the model is trained on S−1 sessions and tested on the held-out session. This is repeated for each session fold, and accuracy is averaged across all folds and subjects. This protocol directly measures cross-session generalization capability. III. Cross-subject LOSO (secondary analysis): leave-one-subject-out evaluation. The model is trained on data from all subjects except one and tested on the held-out subject. Such an evaluation assesses cross-subject generalization, which is a more challenging transfer scenario.

We report the following metrics: classification accuracy (mean ± standard deviation across subjects), Cohen’s kappa coefficient (κ), macro-averaged F1-score, and per-class confusion matrices for the best-performing architecture.

MetaNAS-SI is compared against 10 baseline methods spanning traditional machine learning, hand-designed deep learning, and transfer learning approaches:

All models are implemented in PyTorch 2.0 and trained on an NVIDIA RTX 3090 GPU. MetaNAS-SI search parameters: Aging evolution population P=25, tournament size S=5, 200 generations, proxy training=20 epochs. DE parameters: NDE=20, F=0.5, CR=0.9, 50 generations, full training=100 epochs. Training uses the Adam optimizer [59] with learning rate, batch size, and weight decay optimized by DE from the ranges {1×10−4, 5×10−4, 1×10−3}, {16, 32, 64}, and {1×10−4, 1×10−3, 1×10−2}, respectively. Early stopping with patience of 15 epochs is applied during all training runs. SIFAM MMD kernel bandwidths are set using the median heuristic computed on the training data. All baseline methods are reproduced using publicly available implementations with hyperparameters optimized via grid search on a validation split to ensure fair comparison.

Statistical significance of accuracy differences between MetaNAS-SI and each baseline is assessed using the paired Wilcoxon signed-rank test across subjects at a significance level of α = 0.05, with Bonferroni correction for multiple comparisons (10 comparisons). Effect size is quantified using Cohen’s d. Significance levels are denoted as: * p < 0.05, ** p < 0.01, *** p < 0.001.

Table 4 presents the cross-session LOSO classification accuracy for all methods on the KaraOne dataset (5- class, 12 subjects, 3 sessions per subject). MetaNAS-SI achieved a mean cross-session accuracy of 45.8 ± 4.2%, substantially outperforming all 10 baseline methods. The best baseline, EEGNet + EA, achieved 36.4 ± 4.8%, yielding a 9.4 percentage-point absolute improvement (25.8% relative improvement). The improvement over the standard EEGNet baseline was 14.6 percentage points (46.8% relative). All pairwise comparisons between MetaNAS-SI and baselines were statistically significant at p < 0.001 after Bonferroni correction, with large effect sizes (Cohen’s d > 1.5 for all comparisons).

Among the deep learning baselines, EEG-Conformer (34.1%) outperformed EEGNet (31.2%), suggesting that the attention mechanism captures temporal dependencies relevant to Speech Imagery. However, the improvement from EEG-Conformer over EEGNet (+2.9 pp) was modest compared to MetaNAS-SI’s improvement over EEG-Conformer (+11.7 pp), indicating that architectural innovation alone is insufficient without cross-session optimization. Traditional methods (CSP+LDA, FBCSP+SVM, PSD+RF) performed near or below 30%, consistent with the known limitations of hand-crafted features for the weak and diffuse signals of Speech Imagery.

The within-session comparison (Table 5) reveals a complementary pattern. MetaNAS-SI achieved 62.4 ± 5.1% within-session accuracy, compared to 56.8 ± 5.7% for EEG-Conformer and 54.3 ± 6.2% for EEGNet. The within-session improvement (5.6–8.1 pp) is notably smaller than the cross-session improvement (9.4–14.6 pp), confirming that MetaNAS-SI’s cross-session fitness function specifically optimizes for session-transfer robustness rather than merely improving general classification capability.

MetaNAS-SI also exhibits the smallest within-to-cross-session accuracy drop (−16.6 pp) compared to EEGNet (−23.1 pp) and EEG-conformer (−22.7 pp), demonstrating superior session-transfer robustness. EEGNet + EA reduces the drop to −17.9 pp through EA, approaching MetaNAS-SI’s robustness, but at substantially lower absolute accuracy.

Table 6 presents the cross-session LOSO results on the Coretto imagined speech dataset (4-class vowel classification, 15 subjects). MetaNAS-SI achieved 52.3 ± 3.8%, compared to 43.8 ± 3.9% for the best baseline (EEGNet + EA), representing an 8.5 percentage-point improvement (19.4% relative). The higher absolute accuracies across all methods on Coretto compared to KaraOne are expected given the fewer classes (4 vs. 5) and the phonological distinctiveness of the four vowels (/a/, /e/, /i/, /u/).

Notably, MetaNAS-SI achieves a Cohen’s kappa of 0.364, substantially above the 0.251 of EEGNet + EA, indicating better agreement between predictions and true labels beyond chance. The lower standard deviation (3.8%) across 15 subjects also suggests that MetaNAS-SI provides more consistent performance across individuals.

Table 7 presents results on the FA-IS dataset, the most challenging evaluation due to the higher class count (6 classes, chance = 16.7%) and the longest session span (5 sessions over 2 weeks). MetaNAS-SI achieved 48.1 ± 4.5%, compared to 35.2 ± 4.6% for the best baseline (EEGNet + EA), a 12.9 percentage-point absolute improvement (36.6% relative).

The larger improvement on FA-IS compared to KaraOne and Coretto can be attributed to the 5-session design providing more cross-session variability for the NAS fitness function to exploit. With more session pairs available during architecture evaluation, MetaNAS-SI can more reliably identify architectures that generalize across session-specific confounds. Analysis by native language revealed comparable performance: MetaNAS-SI achieved 49.3 ± 4.1% for Farsi speakers and 46.9 ± 4.9% for Arabic speakers. The difference was not statistically significant (p = 0.31, Wilcoxon), suggesting that the discovered architecture captures language-invariant Speech Imagery representations.

Analysis of the Pareto-optimal architectures discovered by MetaNAS-SI reveals several notable patterns. Table 8 summarizes the three representative architectures from the Pareto front, corresponding to maximum accuracy, balanced, and minimum latency configurations.

Several observations emerge from the discovered architectures:

I. Multi-scale temporal convolutions are consistently preferred. The highest-accuracy architecture (A) selects TC-Multi (parallel multi-scale convolution), while the balanced architecture (B) selects two specific temporal convolutions (TC-Alpha and TC-Beta). These findings confirm the neuroscientific understanding that Speech Imagery engages multiple frequency bands simultaneously and that architectures capturing this multi-scale temporal structure outperform those restricted to a single temporal resolution.

II. CSP-inspired spatial filters outperform standard approaches. Architecture A selects SC-CSP over SC-Full, demonstrating that incorporating prior knowledge from the decades-long CSP tradition in BCI engineering into the learnable architecture yields better spatial feature extraction. The SC-CSP operation effectively bridges classical BCI signal processing with modern deep learning.

III. SIFAM is optimally placed in middle layers. Across all Pareto-optimal architectures, SIFAM is placed after cell 2 or cell 4, never at the earliest or latest cells. These findings suggest that early layers should retain session-specific low-level information (which SIFAM would prematurely remove), while late layers should produce session-invariant, class-discriminative representations (which SIFAM has already facilitated in the middle layers).

IV. Discovered architectures are remarkably compact. With parameter counts ranging from 8K to 42K, the MetaNAS-SI architectures are 4–50× smaller than EEGNet (~170K parameters) and 50–260× smaller than DeepConvNet (~2.1M parameters), while achieving substantially higher cross-session accuracy. This compactness suggests that Speech Imagery decoding does not benefit from overparameterized architectures but rather requires the right inductive biases, which MetaNAS-SI discovers automatically.

While cross-session transfer is the primary focus of this work, we also assessed cross-subject generalization using the more challenging leave-one-subject-out protocol. Table 9 presents results.

Cross-subject accuracy is substantially lower than cross-session accuracy for all methods, reflecting the greater distributional shift between subjects (anatomical differences, individual cognitive strategies, neurophysiological variability) compared to between sessions of the same subject. Nevertheless, MetaNAS-SI achieves the largest improvement in cross-subject settings as well (4.7–7.1 pp over best baseline), suggesting that SIFAM and the cross-session fitness function also confer partial robustness to subject-level distribution shifts.

To quantify the contribution of each component, we conducted a systematic ablation study on the KaraOne dataset, removing one component at a time from the full MetaNAS-SI framework. Results are presented in

The two most impactful components are the cross-session fitness function (−10.6 pp when removed) and the aging evolution search algorithm (−8.0 pp when replaced with random search). The cross-session fitness function is the single most important innovation: Without it, the NAS discovers architectures that overfit to within-session patterns, and the resulting architecture achieves only 35.2% cross-session accuracy, comparable to hand-designed baselines. These findings confirm our central hypothesis that explicitly optimizing for cross-session generalization during architecture search is critical.

Removing SIFAM (−7.4 pp) demonstrates that domain adaptation is a crucial complementary mechanism: even with cross-session fitness, the absence of explicit feature alignment substantially degrades performance. The EEG-specific search space contributes −5.7 pp; replacing it with generic operations (standard convolutions with arbitrary kernel sizes, no CSP-inspired filters) yields an architecture that fails to exploit the oscillatory structure of EEG signals. Data augmentation (−4.2 pp) and DE micro-optimization (−3.5 pp) provide smaller but meaningful contributions.

Interestingly, removing the multi-objective constraint (allowing the search to optimize only for accuracy without penalizing model size) yields a marginal accuracy increase (+0.3 pp) but produces a model 5× larger (~210K parameters). These findings confirm that the Pareto-optimal architecture selection imposes minimal accuracy cost while dramatically reducing model size.

To examine the consistency of cross-session transfer across different session pairs, Table 11 presents a per-session breakdown for a representative subject (subject 3) from the KaraOne dataset.

MetaNAS-SI achieves not only higher mean accuracy but also lower variance across session pairs (SD = 2.5 vs. 2.9–3.0 for baselines), indicating more consistent and reliable cross-session transfer. The S1→S3 transfer (largest temporal gap between sessions) is the most challenging for all methods, but MetaNAS-SI maintains a relatively small drop compared to adjacent-session transfers.

Table 12 summarizes the computational characteristics of MetaNAS-SI and baselines.

All MetaNAS-SI architectures achieve inference latency well below the 50 ms threshold required for real-time BCI operation at a 256 Hz sampling rate. The minimum-latency architecture (C) achieves 3 ms inference, enabling deployment on edge devices with limited computational resources. While the one-time NAS search cost of ~90 GPU-hours is substantial, it is amortized across all future subjects and sessions and replaces months of manual architecture engineering effort.

To provide neurophysiological interpretability, we analyzed the contribution of each temporal convolution kernel in the discovered Architecture A using gradient-based feature importance (integrated gradients [60]). Table 13 presents the relative importance of each temporal convolution operation.

The beta band (13–30 Hz) emerged as the most important frequency range, contributing 32.1% of the total feature importance. This finding is consistent with the well-established role of beta oscillations in motor planning and speech production: Beta desynchronization in sensorimotor and frontal regions accompanies both overt and imagined speech [49], [61]. The alpha band (8–13 Hz, 24.5% importance) likely reflects alpha event-related desynchronization associated with cortical activation during imagery. The gamma band (30–100 Hz, 18.7% importance) is linked to high-level language processing, with gamma synchronization in Broca’s and Wernicke’s areas during language tasks [50]. The theta band (4–8 Hz, 15.2% importance) reflects working memory and semantic processing demands inherent to imagined speech [51]. The delta band contributes the least (9.5%), consistent with its primary association with deep sleep and autonomic processes rather than active cognitive tasks. These findings demonstrate that MetaNAS-SI discovers architectures whose learned features align with established neuroscientific understanding of speech-related cortical oscillations, enhancing the interpretability and trustworthiness of the discovered architectures.

The ablation study reveals that the cross-session fitness function is the single most impactful component of MetaNAS-SI, contributing 10.6 percentage points to cross-session accuracy on KaraOne. Standard NAS approaches optimize within-session accuracy, inadvertently selecting architectures that exploit session-specific confounding patterns in the EEG that are consistent within a session but do not transfer across sessions. These confounds include session-specific electrode impedance profiles that create characteristic spatial noise patterns, environmental electromagnetic interference that differs between recording days, and the subject’s momentary cognitive strategy or arousal level. By evaluating candidate architectures using a LOSO protocol, MetaNAS-SI applies selection pressure that directly penalizes reliance on such non-transferable patterns and rewards extraction of session-invariant neural features. This simple but powerful shift in the optimization target fundamentally changes which architectures are discovered, favoring those with the right inductive biases for cross-session robustness.

SIFAM contributes 7.4% points to cross-session accuracy by explicitly minimizing the distributional distance between session-specific feature representations at an intermediate network layer. By forcing the MMD between source and target session features toward zero, SIFAM encourages the network to learn representations that are invariant to session-specific confounds while preserving class-discriminative information (because the classification loss is optimized simultaneously). The finding that SIFAM is optimally placed in middle layers (cell 2–4) rather than at the input or output has an intuitive interpretation: Early layers should retain session-specific low-level information (e.g., signal amplitude scaling, channel-specific noise patterns) that is useful for the subsequent alignment process, while late layers should already operate on session-invariant representations. Placing SIFAM too early would force alignment of raw signals before meaningful features are extracted, while placing it too late would miss the opportunity to guide the network toward invariant intermediate representations.

The 5.7% point improvement contributed by the EEG-specific search space (compared to generic operations) validates the hypothesis that constraining the NAS to neurophysiologically plausible operations yields better architectures more efficiently. The physiologically-motivated temporal convolution kernels ensure that the network can efficiently extract features from canonical EEG frequency bands without needing to learn appropriate filter lengths from scratch. The CSP-inspired spatial filters embed decades of BCI engineering knowledge into the learnable architecture, providing a structured inductive bias that standard spatial convolutions lack. This approach represents a productive middle ground between fully hand-designed architectures (which may miss novel architectural possibilities) and unconstrained NAS (which wastes search budget exploring neurophysiologically implausible operations).

The remarkably compact size of discovered architectures (8K–42K parameters vs. 170K–2.1M for baselines) challenges the assumption that larger models are needed for EEG decoding. EEG signals, particularly from Speech Imagery, have relatively limited information content compared to natural images or audio, and the small training datasets typical in BCI (tens to hundreds of trials per class) cannot support highly overparameterized models without severe overfitting. MetaNAS-SI discovers that the critical ingredients for Speech Imagery decoding are multi-scale temporal feature extraction, spatially informed filtering, and domain adaptation, not model depth or width. This finding has practical implications: Compact models are faster to train, require less memory, and can be deployed on embedded BCI hardware.

The 45.8% accuracy for 5-class imagined speech classification in cross-session evaluation on KaraOne represents, to our knowledge, the highest reported cross-session performance on this dataset. While this accuracy is not yet sufficient for reliable communication (which typically requires >70% accuracy for practical BCI communication systems [1]), it represents a meaningful step toward clinical viability. The cross-session robustness of MetaNAS-SI is particularly relevant for clinical deployment, where daily recalibration is burdensome for patients with severe motor impairments. Combining MetaNAS-SI with improved recording protocols (e.g., standardized electrode placement using 3D digitization), larger training datasets (enabled by multi-day recordings), and complementary modalities (e.g., Functional Near-Infrared Spectroscopy (fNIRS)) could push accuracy toward clinically useful levels. The Pareto-optimal architecture selection also enables clinicians to choose deployment configurations appropriate for the clinical context: Maximum accuracy for research-grade offline analysis, balanced for general-purpose BCI use, or minimum latency for real-time feedback applications.

Several limitations of the current work should be acknowledged. First, the NAS search cost of approximately 90 GPU-hours, while a one-time investment, may be prohibitive for researchers without access to GPU computing infrastructure. Transfer of discovered architectures across datasets could reduce this cost but requires further validation. Second, evaluation was conducted on relatively small datasets (12–15 subjects, 3– 5 sessions), and the generalizability of the discovered architectures to larger populations and longer-term studies remains to be established. Third, the search space, while EEG-specific, is manually designed based on our domain knowledge; automatic search space generation through meta-learning could further improve results. Fourth, MetaNAS-SI discovers a fixed architecture after the search process; online adaptation of the architecture during BCI use (e.g., adapting to changing neural states within a session) is not addressed. Fifth, only discrete imagined words and phonemes were evaluated; continuous imagined speech decoding, which is a substantially harder problem requiring temporal sequence models, is beyond the scope of this work. Sixth, the current framework uses EEG as the sole input modality; multimodal fusion with complementary signals such as fNIRS or Electromyography (EMG) from articulatory muscles could provide additional discriminative information.

The limited existing NAS work in BCI [19], AutoEEG by Zhu et al. [20], focuses exclusively on motor imagery classification using within-subject evaluation. Neither incorporates cross-session objectives, searchable domain adaptation modules, or EEG-specific operations. MetaNAS-SI’s cross-session fitness function and SIFAM module represent novel contributions to the NAS-BCI literature. Furthermore, the hybrid aging evolution + DE strategy addresses both the discrete and continuous aspects of the architecture optimization problem more naturally than purely differentiable or purely evolutionary approaches.

Several promising research directions emerge from this work. Extension to online BCI with real-time architecture adaptation could enable the system to adjust its processing pipeline as the user’s neural signals change within and across sessions. Transfer of MetaNAS-SI to intracranial EEG Electrocorticography (ECoG) could leverage higher signal quality for more accurate speech decoding while still benefiting from automated architecture optimization. Federated NAS across multiple BCI centers would enable architecture discovery from larger and more diverse datasets without centralizing sensitive neural data. Integration of meta-learning for few-shot session adaptation could reduce the calibration data required for a new session from hundreds of trials to tens. Continuous imagined speech decoding, where the user imagines speaking full sentences rather than isolated words, represents the ultimate goal for silent speech interfaces and could benefit from MetaNAS-SI’s framework by incorporating sequence-to-sequence architectures into the search space. Finally, integration with large language models as post-processing could leverage linguistic priors to correct classification errors, effectively implementing a neural speech “autocorrect” system.