Aug 2026· IEEE Transactions on Cybernetics· Vol PP, pp. 1-13· 0 citations
Medicine
TL;DR
A novel training framework that enhances the generalization ability of the model by reducing the influence of noisy samples during training, without additional hyperparameter optimizations is proposed, which improved the generalization ability of the model across two motor imagery datasets and one sleep stage dataset.
Abstract
As deep learning (DL) performs remarkably in pattern recognition from complex data, it is used to interpret user intentions from electroencephalography (EEG) signals. However, the DL models trained on EEG datasets have low generalization ability owing to numerous noisy samples in datasets. Therefore, prior research has focused on distinguishing and eliminating noisy samples from datasets. One intuitive solution is based on the property of noisy samples during the training phase. Noisy samples are located near the decision boundary after model training. Therefore, they can be detected using a gradient-based adversarial attack. However, the limitations of usability exist because the intuitive solution requires additional hyperparameter optimizations, resulting in a tradeoff between accurateness and efficiency. In this article, we proposed a novel training framework that enhances the generalization ability of the model by reducing the influence of noisy samples during training, without additional hyperparameter optimizations. We designed the proximity-to-boundary score (PBS) to continuously measure the data closeness to the decision boundary. As a result, the proposed framework improved the generalization ability of the model across two motor imagery datasets and one sleep stage dataset. Specifically, training the model with the proposed framework resulted in performance improvements ranging from a minimum of 1.43% to a maximum of 6.66% on two motor imagery datasets and from a minimum of 0.72% to a maximum of 2.85% on a sleep stage classification dataset. We qualitatively confirmed that data with low PBS are indeed noisy samples and degrade the model training. Hence, we demonstrated that employing the proposed framework accurately and efficiently mitigates the influence of noisy samples, enhancing the model's generalization capabilities.
Goal: Deep learning-based motor imagery EEG classification is limited by data scarcity, which constrains model generalization and performance. Methods: We propose a dual-cascade generative adversarial network (dcGAN) framework with a variable focused attention (VFA) module for MI-EEG data augmentation. The first stage learns latent frequency-domain priors from random noise through an adversarial training scheme; the second stage then synthesizes artificial EEG samples with a U-Net generator conditioned on these priors, augmented by the VFA module and a time-domain consistency loss. A VFA-enhanced EEGNet is subsequently trained on the combination of real and generated samples for classification. Results: On the BCI Competition IV 2a and 2b datasets, the proposed method achieves classification accuracies of 84.92% and 91.79%, with Cohen’s Kappa coefficients of 0.79 and 0.81, respectively, outperforming baseline methods. Conclusions: The integration of structured frequency-domain priors and attention mechanisms improves the fidelity of generated EEG samples, which in turn enhances downstream classification performance.
The growing use of electroencephalography (EEG) motivates automated analysis that is accurate, transferable, and deployable on constrained hardware. Recent EEG foundation models learn general representations from large-scale pretraining, but their size and computational cost limit edge and wearable deployment. We introduce BRIDGE-EEG, an efficient multi-task EEG classification pipeline that preserves the benefits of pretraining while reducing model size. A unified preprocessing scheme maps heterogeneous recordings with different channel counts, montages, and sampling rates to a device-agnostic 62-channel time--frequency representation. We pretrain an SE-ResNet18 teacher (11.84 M parameters) with SimCLR on unlabeled EEG from five heterogeneous datasets, then compress it into SE-ResNet8 (1.56 M) and SE-ResNet4 (0.48 M) students using task-agnostic and task-specific distillation. We evaluate six benchmarks spanning abnormality detection, motor imagery, and emotion recognition. For abnormality detection and emotion recognition, the students achieve accuracy comparable to or better than several recent EEG foundation models with 10--1,000$\times$ more parameters. Motor imagery shows a remaining representation gap, highlighting the importance of pretraining diversity. Inference profiling on a server GPU, desktop CPU, and NVIDIA Jetson Orin Nano shows up to 3.0$\times$ lower edge energy per inference (15.64 mJ vs. 46.67 mJ). The compact models further support future deployment on MCU-class wearables.
Meghna Roy Chowdhury, Cheng-Wei Zhou, Hao-Tian Yu et al.· 0 citations
Electroencephalography (EEG) models used for epilepsy are often limited to specific datasets and tasks. This limited approach can make it challenging to apply these models across different datasets or in various situations. However, recent studies in foundation models and self-supervised learning suggest that an adaptable EEG backbone could support a range of EEG related tasks. In this study, we have developed a multimodal EEG foundation model that combines a raw signal encoder based on the Mamba architecture, a Vision Transformer (ViT)-style encoder for time-frequency data, and a lightweight encoder for text, all within a shared embedding space. The pretraining process relies on several innovative techniques, such as masked modeling, cross-view contrastive alignment, and temporal consistency losses. These methods are designed to create rich, seizure-relevant representations without requiring labeled data. To assess the efficacy and generalization of our pretrained model, we fine-tuned it on the canonical CHB-MIT seizure detection benchmark and additional seizure detection datasets, and conducted extensive experiments comparing different model variants. On the standard CHB-MIT split, our best single model achieved an AUROC of 0.874, and an ensemble variant reached 0.878 AUROC, representing state-of-the-art performance on this benchmark. In addition to standard train-test splits, we evaluated performance under a leave-one-subject-out (LOSO) protocol, which is rarely reported in prior EEG seizure modeling work and highlights the difficulty of patient-independent seizure detection, with a mean LOSO balanced accuracy of 0.558 across 19 subjects. Across datasets and evaluation settings, our multimodal foundation model enabled robust seizure detection and straightforward adaptation to new seizure detection scenarios, while also supporting interpretable seizure localization.
Targol Bakhtiarvand, J. Kalita, Adham Atyabi· arXiv.org· 0 citations
Detecting abnormal events in physiological signals such as EEG and ECG is critical for early diagnosis of neurological and cardiovascular disorders. However, existing unsupervised anomaly detection methods often suffer from limited representation capacity and weak generalization across diverse signal domains. To address these challenges, we propose MAGE, a novel unsupervised anomaly detection framework that integrates multi head memory gating, self-supervised learning, and adversarial training within a unified convolutional autoencoder architecture. The proposed memory-augmented gating mechanism selectively preserves and adaptively integrates salient features, improving discriminability and robustness over prior memory-based approaches. To further enhance representation learning, a self-supervised auxiliary task based on multiple signal transformations is introduced to encourage structure-aware feature extraction. In addition, transformation-aware adversarial perturbations are incorporated during training to enhance robustness against distribution shifts. Extensive experiments on multiple benchmark EEG and ECG datasets demonstrate that MAGE consistently outperforms state-of-the-art baselines in standard within-dataset evaluation settings, achieving over 98% detection accuracy and superior F1-scores. These results highlight the effectiveness and clinical potential of MAGE for early anomaly detection and continuous health monitoring under realistic deployment conditions. Code available at https://github.com/wzxmodel/MAGE.
Zexin Wang, Lin Shi, Shoukun Xu et al.· IEEE journal of biomedical a...· 0 citations
In this paper, we discuss the recognition of electroencephalographic (EEG) signals, which is crucial in order to improve the performance of non-invasive brain-computer interfaces (BCIs). Although deep learning (DL) has achieved considerable advancements in the decoding of EEG signals, it frequently encounters difficulties related to noisy data and non-stationarity challenges. We discuss geometric learning which offers a more robust way to handle EEG signals by leveraging the mathematical structure of the data. We extended the existing Manifold Attention Network (MAtt), a novel deep learning model that applies a manifold attention mechanism to better capture the spatiotemporal patterns of EEG signals. Instead of using traditional Euclidean space, we mapped the data onto a Riemannian symmetric positive definite (SPD) manifold, which allows for more effective feature extraction. One major issue with existing SPD-based deep learning approaches is that they rely on fixed Riemannian metrics, which can be suboptimal. To solve this, we integrate Adaptive Log-Euclidean Metrics (ALEMs)---a learnable metric framework into the MAtt network that adapts to the specific structure of EEG data, improving model performance with minimal extra computation. AMAtt achieves 63.19% accuracy on BCIC-IV-2a and 46.00% on MAMEM-SSVEP-II, outperforming the MAtt baseline by 5.55% and 4.60% respectively. This adaptive geometric approach opens new possibilities for robust, subject-independent EEG decoding, paving the way towards practical BCI systems.
S. Shaihan· European Journal of Clinical...· 0 citations
Wearable EEG systems may expose sensitive information beyond their intended health function, creating substantial risks to neuroprivacy. In this work, we show that commonly used EEG features can reveal participant identity and demographic attributes in addition to supporting the intended cognitive task. Wearable EEG is increasingly being explored for cognitive monitoring, neurological assessment, and longitudinal digital-health applications, yet many systems assume that transmitting compact spectral or spatial features instead of raw EEG provides sufficient privacy protection. Using EEGMAT as a motivating case study, we find that compact EEG features achieve a balanced accuracy of 0.788 for cognitive-state classification while enabling gender, age, and subject-identity inference with balanced accuracies of 0.858, 0.789, and 0.692, respectively. We further show that privacy-aware representation learning preserves task performance at 0.781 while reducing these inference accuracies to 0.563, 0.467, and 0.206. These findings motivate purpose-limited representations and explicit privacy auditing in wearable neurohealth systems.
Sarmistha Sarna Gomasta, Bhawana Chhaglani, Prashant J. Shenoy· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
