脑电拓扑表征的黑盒困境与可解释性需求

目前在脑科学领域中，尽管有了许多突破性的进展，但我们对复杂大脑功能和认知原理和机制的理解仍然不完全。复杂网络作为描述复杂系统结构的概念，被引入脑科学以解释这些问题。本文介绍了复杂网络的相关概念，并回顾了复杂网络在脑科学中的体现和意义，以更好的推广这一工具在脑科学中的应用。

针对跨学科EEG情绪识别中个体差异显著和时频特征泛化不足的挑战，本文提出了一种多频时空注意力网络FA-TSception。该模型创新性地整合了多频率自适应机制和高效的通道注意力，构建了一个基于TSeption多尺度时空架构的三级处理框架。多频动态时间层通过参数化比例因子生成自适应卷积核组，以精确匹配Alpha、Beta、Gamma等情绪相关频带的时频特征；非对称空间层结合半球卷积核提取前额叶和时间区域的空间激活模式；集成了高效的信道注意力模块(ECA)，实现了多频特征的自适应校准。DEAP数据集上的跨学科实验表明，FA-TSception在唤醒和效价维度上的平均分类准确率分别达到62.73%和60.12%。与TSception相比，它提高了1.16%，仅增加了5.6%的模型参数计数。FA-TSception不仅提高了跨个体EEG情绪识别的准确性，而且通过引入有效的注意力机制，同时保持相对稳定的模型参数数量，增强了模型识别情绪相关特征的能力。

在工业5.0人本范式下，精密印刷质量控制亟需对操作人员注意力进行有效监测。本文提出一种基于实时EEG的注意力评估算法，将谱熵与传统频带特征相结合表征认知状态。基于同一脑电数据条件下的对比实验结果表明，所提出方法在注意力评估中相较传统频段功率指标表现出更高的稳定性与一致性，其输出差异在统计意义上达到显著水平，同时保持较低的计算成本与良好的可解释性。研究结果验证了谱熵特征在工业脑电注意力监测中的工程有效性，为工业5.0背景下将人因因素融入智能制造系统提供了可行的人本化技术路径。

记忆中不可避免的会存储一些令人厌恶或不愉快的信息，这些信息会对人们的社会适应和心理状态产生诸多消极影响。因此，人们通常会被动或主动的对不想要的记忆进行限制，这种限制可能体现在编码阶段，也可能体现在线索诱发提取阶段。本研究对遗忘过程中的脑电成分和节律活动特征进行概括，旨在找到遗忘的特异性电生理标志。我们发现，遗忘过程主要体现在N2振幅的增加，alpha能量的增强，晚正成分(LPC)减少，theta能量和同步性下降，以及alpha/beta频带间同步性下降，并且以上变化可以预测后续遗忘的成功发生。因此，遗忘可以通过特定的电生理机制来反映，进一步也可能从电生理现象反推遗忘对应的认知过程。

在人工智能技术飞速发展的当下，病理医学领域数据展现出多模态、高维度及复杂关联的特性，传统机器学习方法处理这类非欧几里得结构数据时困难重重。图神经网络(GNNs)作为新兴深度学习模型，能够高效处理图结构数据并获取深层次特征。医学领域中蛋白质相互作用网络、基因调控网络等应用场景与图结构天然适配，促使GNN模型在医学数据分析中的应用成为人工智能与医学交叉研究的热门方向。本文深入剖析近年来GNN模型在医学领域数据的应用，不仅阐述了其基本原理和模型架构，还从脑网络分析、疾病诊断与预测、药物发现与相互作用预测以及模型可解释性等多个维度，对相关研究进行详细解读与深度分析，提出模型设想，探讨未来发展方向与面临的挑战。目的在于为医学和人工智能领域的研究者提供全面且系统的参考，助力GNN模型在医学领域实现更广泛、更深入的应用与发展。

抑郁症已成为全球广泛流行的神经心理疾病，但传统的诊断手段因主观性影响，常导致较高的误诊率。鉴于此，开发一种更为客观且高效的抑郁症识别方法显得尤为重要。本研究构建了一种基于时空图卷积神经网络(ST-GCN)的脑电信号抑郁分类模型，将时间序列的脑电信号转化为脑拓扑图，捕捉大脑复杂的空间结构信息。同时，引入时空注意力机制，从时间和空间两个维度上有效提取关键信息。具体而言，时空图卷积神经网络结合了空间图卷积和时间卷积的优势，分别用于捕获脑电信号的空间布局特征和时间动态特性。实验结果显示，在公开的脑电数据集HUSM上，该模型的分类准确率、灵敏度以及特异度优于其他基线模型，充分验证了该模型在抑郁症识别方面的优越性能。

精神分裂症是一种严重的精神疾病，影响患者的思维、情感和行为。研究利用71名成年精神分裂症患者及对照组的脑电图信号和84名青少年精神分裂症患者及对照组的脑电图信号进行分析，基于定向传递函数构建了Delta、Theta、Alpha、Beta和Gamma频段的脑功能网络，通过数据预处理、网络构建、阈值选择、特征提取与分析及可视化，揭示患者与正常组脑功能网络特征差异。结果显示，在不同频段下，精神分裂症患者与正常组的脑功能网络特征存在显著差异，且成年患者与青少年患者之间也有显著差异，通过热图和圆环图的可视化结果也揭示了成年精神分裂症患者和青少年精神分裂症患者、成年精神分裂症患者和正常对照组、青少年精神分裂症患者和正常对照组在不同频段的脑网络连接模式异常，这为理解该疾病的神经机制提供了新的视角，并为临床诊断和干预提供了参考依据。

卷积神经网络(Convolutional neural networks, CNNs)已在运动想象(Motor Imagery, MI)脑机接口(Brain-Computer Interface, BCI)领域取得了广泛应用和良好效果。然而，基于单一尺度卷积结构的模型难以从脑电信号中充分挖掘多样化信息；而现有的多尺度CNN虽然能够提取不同尺度的特征，但通常仅通过简单拼接的方式进行融合，难以实现多尺度信息之间的深度协同建模。为了解决上述问题，相关研究提出了双尺度时空特征融合网络(Dual-Scale Spatiotemporal Feature Fusion Network, DSSFFN)。该方法通过双尺度卷积分支以提取脑电信号不同尺度的特征，并通过Transformer模块对来自不同尺度的特征进行融合，从而提升特征的判别性。本文采用BCI竞赛IV的2a数据集进行实验，同时与多个前沿的运动想象算法进行比较。结果显示，DSSFFN在数据集上平均准确率为72.65%，高于所有的对比模型，展现出了DSSFFN模型在运动想象脑电识别任务上的优异性能。此外，研究还通过消融实验分析了双尺度时空卷积分支与Transformer模块对整体性能的贡献，验证了各个关键模块的有效性。同时比较了不同融合模型对模型影响，进一步验证了Transformer模块融合双尺度的有效性。

基于运动想象的脑机接口(Brain Computer Interface, BCI)可以帮助残疾人控制机械手臂等外部设备，其中脑电信号解码是关键所在。但是不同个体间的脑电信号差异很大，使得传统的深度学习模型所采用的静态卷积很难自适应地提取脑电特征。为解决这个问题，本文提出了基于多维动态卷积的深度学习模型(Multidimensional Dynamic Convolution Net, MDconvnet)，该模型通过三层多维动态卷积来提取特征，并将提取的特征输入到全连接层来获取分类结果。其中多维动态卷积会依据输入的数据，生成卷积多维度的注意力权重，并将该权重与卷积参数相乘来动态地调节卷积参数，以便更好地挖掘数据时空特征。本文采用2023运动想象数据集RankA和数据集RankB对MDConvnet模型进行了测试，同时与多个经典的运动想象识别模型(FBCSP、EEGnet、EEGTCN、FBCnet、Tesecption、STASCNN、Deepconvnet和VIT)进行性能对比。结果显示MDConvnet模型在RankA和RankB数据集上的平均准确率分别为64.20%和67.04%，超过其他算法模型，展现出了MDConvnet模型在运动想象脑电识别任务上的优异性能，为残疾人通过脑机接口控制外部设备提供了有力的支持。

针对运动想象(Motor Imagery, MI)脑电信号(EEG)普遍存在的信噪比低、非平稳性强以及传统卷积神经网络在特征提取过程中易产生大量时空冗余信息的问题，本文提出了一种融合多尺度局部特征选择与特征重构机制的深度学习解码模型。首先，该模型在浅层特征提取后引入多尺度局部特征选择模块，通过并行的多尺度深度卷积捕获不同感受野下的特征，并利用可学习的通道注意力权重实现特征的自适应加权，以增强特征的判别性。其次，为了进一步抑制任务无关的冗余信息，模型引入了空间与通道重构卷积(SCConv)特征优化模块，通过空间重构单元(SRU)和通道重构单元(CRU)对特征图进行压缩与重组，从而显著提升特征表达的有效性。在大型公开数据集OpenBMI上的实验结果表明，该模型在运动想象任务中的平均准确率达到72.95%，优于EEGNet、Conformer等主流对比方法。消融实验进一步证实了多尺度特征选择模块与SCConv特征优化模块在提升模型鲁棒性和解码性能方面的关键作用。

疲劳驾驶是交通事故的重要诱因，其检测对于保障交通安全至关重要。脑电图(EEG)信号因能反映大脑活动状态而被广泛用于疲劳驾驶检测，但现有深度学习方法常面临对预处理的依赖或对EEG信号时空信息的处理不足等挑战。本文提出了一种基于脑电图信号的时空融合疲劳驾驶检测方法——KCL-STN (KAN-CNN-LSTM Spatio-Temporal information Network)。该方法巧妙地结合了卷积神经网络和长短期记忆网络，分别从原始脑电信号中提取空间和时间特征，并进行有效融合，实现了端到端的疲劳驾驶检测。针对脑电数据稀缺问题，本文还提出了一种脑电信号滑动窗口增强算法，以增加样本数量并提高模型训练的稳定性。在公开数据集上的实验结果表明，KCL-STN在分类准确度、召回率和精确率等指标上均优于多种现有方法，准确率达到86.05%。消融实验证实了关键组件KAN线性层和滑动窗口数据增强方法的有效性。跨被试实验也证明了模型良好的泛化性能和鲁棒性。研究结果表明，KCL-STN能够有效地从原始脑电信号中提取疲劳相关特征，是鲁棒且高性能的疲劳驾驶检测方法。

Due to the limited perceptual field, convolutional neural networks (CNN) only extract local temporal features and may fail to capture long-term dependencies for EEG decoding. In this paper, we propose a compact Convolutional Transformer, named EEG Conformer, to encapsulate local and global features in a unified EEG classification framework. Specifically, the convolution module learns the low-level local features throughout the one-dimensional temporal and spatial convolution layers. The self-attention module is straightforwardly connected to extract the global correlation within the local temporal features. Subsequently, the simple classifier module based on fully-connected layers is followed to predict the categories for EEG signals. To enhance interpretability, we also devise a visualization strategy to project the class activation mapping onto the brain topography. Finally, we have conducted extensive experiments to evaluate our method on three public datasets in EEG-based motor imagery and emotion recognition paradigms. The experimental results show that our method achieves state-of-the-art performance and has great potential to be a new baseline for general EEG decoding. The code has been released in https://github.com/eeyhsong/EEG-Conformer.

No abstract available

Brain-computer interface (BCI) technology decodes electroencephalography (EEG) signals to identify motor intentions associated with motor imagery (MI), offering assistive solutions for individuals with motor impairments. However, current deep learning methods often overlook the long-sequence nature of EEG-MI signals, leading to limited feature extraction and reduced decoding accuracy. To address this, we propose SATrans-Net, an end-to-end framework that models long-range dependencies in EEG-MI signals to enhance decoding performance. SATrans-Net uses two-dimensional depthwise separable convolution (2D-DSC) to extract spatiotemporal features and incorporates a Top-K Sparse Attention (TKSA) mechanism into the Transformer architecture, improving long-range modeling while reducing computational cost. By fusing local and global features, the model achieves accurate classification via a fully connected layer. For interpretability, Grad-CAM is applied to generate Class Activation Topography (CAT) maps, visualizing spatial attention over cortical regions. Cross-session evaluations show that SATrans-Net achieves average accuracies of 84.72%, 89.76%, and 96.79% on the BCI IV-2a, BCI IV-2b, and High-Gamma datasets, respectively, outperforming existing methods. Ablation studies further verify the critical role of the TKSA module. Overall, SATrans-Net demonstrates high decoding accuracy and strong interpretability, paving the way for the application of computational techniques in biomedical signal processing. Source Code:https://github.com/Jasmin-Tianhua/EEG-research_SATrans-Net

Trustworthy Graph Neural Networks (GNNs) for EEG emotion recognition should identify emotions accurately and elucidate corresponding rationales. Current GNNs have achieved notable performance by dynamically modeling emotional connections between EEG channels. However, these GNNs lack interpretability due to the absence of explicit rationale behind their predictions. This paper conducts a comprehensive identification of important EEG channels to enhance the interpretability of EEG emotion recognition from the perspective of mutual information. Specifically, an Adjacency-Explainable Graph Neural Network (AEG) for ante-hoc interpretability is proposed to capture genuine EEG emotional connections, which gives a theoretical guarantee to remove spurious connections. Moreover, a Channel-wise Adaptive Class Activation Mapping Explainer (CACA) for post-hoc interpretability is developed to locate the EEG channels that contribute most to predictions. Experimental results on three datasets, i.e., SEED, SEED-IV, and DREAMER, prove that imbuing training processes with enhanced interpretability ensures significant performance improvements in emotion recognition. Quantitative comparisons of post-hoc interpretability also demonstrate the superiority of CACA. Furthermore, this paper illustrates two potential applications of the proposed methodologies, showing their broader utility and significance.

EEG emotion recognition plays a significant role in various mental health services. Deep learning-based methods perform excellently, but still suffer from interpretability. Although methods such as Gradient-weighted Class Activation Mapping(Grad-CAM) can cope with the above problem, their coarse granularity cannot accurately reveal the mechanism to promote emotional intelligence. In this paper, fine-grained interpretability is proposed, called Concat-aided Grad-CAM. Specifically, the multi-level feature mapping before the fully connected layer is concatenated to obtain the gradients of the target concept so that the discriminant information can be directly located in the high-precision area. Unlike coarse-grained interpretability methods applied in EEG emotion recognition, it can accurately highlight the EEG channels related to emotion rather than an obscure area. In addition, a systematic brain functional network is proposed to reveal the relationship between those channels and to further improve emotion recognition performance. The channels with greater contributions are connected, and those connections are learned by dynamic graph convolutional networks, while the others are independent to eliminate interference. Experiments on four EEG emotion recognition datasets manifest that Concat-aided Grad-CAM can be interpreted by the fine-grained. In addition, it has been shown that the learned brain functional network can improve the performance of the baselines. Significantly, the experiment results achieve state-of-the-art performance in multiple experiments.

Emotion recognition plays a vital role in human-computer interaction, offering applications in healthcare, education, and entertainment. This research proposes a single modality paradigm for emotion detection leveraging EEG signals from the DREAMER dataset to predict the valence, arousal, and dominance. Unlike the other approaches that rely on using computationally expensive neural networks, we have utilized lightweight machine learning models and advanced feature engineering methods to achieve accuracies of 98.6%, 96.6%, and 95.7% respectively which are pretty close to the state of art model accuracies while having a relatively much lower inference time of just 9.78 milli seconds. Another highlight of our research is that we were able to determine which EEG channels and characteristics had the greatest influence on each emotional dimension, which further improves the interpretability of our research. Our approach highlights the possibility for scalable and explicable emotion identification systems utilizing EEG data alone by combining high accuracy, efficiency, and interpretability.

No abstract available

Predicting epileptic seizures effectively allows patients to take preventive measures in advance, reducing accident risk and enhancing safety. Several modeling challenges remain open: (1) The complex spatiotemporal dependency of EEG signals makes it challenging to design a model that efficiently extracts spatial and temporal features from multi-channel EEG signals to classify epileptic EEG signals. (2)While many studies have utilized machine learning and deep learning for seizure prediction, they often lack research on model interpretability. To tackle the challenges above, this paper proposes a novel approach—an epilepsy prediction framework that combines EEG-Inception Convolutional Neural Networks (EICNN) with Feature Pattern Interpretability Post-processing (FPIP). It yielded substantial performance improvements and achieved an average sensitivity of 94.6%, a false prediction rate of 0.29/h in the study involving 22 pediatric patients from the CHB-MIT database using the leave-one-out method. The proposed FPIP provides visual interpretations and significantly enhances performance in predicting epileptic seizures.

Abstract Electrophysiological studies are profiting from multivariate pattern analysis methods. However, these mostly rely on machine-learning algorithms that assume consistent response latencies across trials and individuals. Deep learning provides high performance without such assumptions, but often at the cost of interpretability of learned features. Here, we evaluated how the interpretability of deep learning for electroencephalography (EEG) data is affected by preprocessing choices, the network’s architecture, and the way the learned features are extracted and visualized. We trained two convolutional neural networks (CNN): (1) ResNet, a residual network, and (2) EEGNet, which leverages spatiotemporal properties of EEG. We trained these networks to decode single-trial EEG responses to three different visual stimuli (visual dataset) and to the presence of a sound (auditory dataset). We then extracted and visualized learned features with two gradient-based techniques: saliency and gradient-weighted activation maps (GradCam). Results showed that EEGNet and ResNet performed at a similar level. Yet, visualization of learned features revealed that different architectures learn different aspects of the data. Between the two CNNs, EEGNet features had a higher similarity to the EEG data than ResNet features. Moreover, the latency and distribution of important electrodes varied depending on the visualization technique. GradCam provided features more similar to EEG data than those with saliency, emphasizing the impact of the feature extraction method on interpretability. Our results call for careful consideration of network architecture and feature visualization methods to improve interpretability, which is a crucial step for advancing the use of deep learning in EEG research.

Cybersickness poses a serious challenge for users of virtual reality (VR) technology. Consequently, there has been significant effort to track its occurrence during VR use with brain activity through electroencephalography (EEG). However, a significant confound in current methods for detecting sickness from EEG is they do not account for the simultaneous processing of the sickening visual stimulus that is present in the brain data from VR. Using event-related potentials (ERPs) from an auditory stimulus shown to reflect cybersickness impacts, we can more precisely target EEG cybersickness features and use those to achieve better performance in online cybersickness classification. In this article, we introduce a method utilizing trained convolutional neural networks and transformer models and plot interpretability maps from integrated gradients and class activation to give a visual representation of what the model determined was most useful in sickness classification from an EEG dataset consisting of ERPs recorded during the elicitation of cybersickness. Across 12 runs of our method with three different neural networks, the models consistently pointed to a surprising finding: that amplitudes recorded at an electrode placed on the scalp near the left prefrontal cortex were important in the classification of cybersickness. These results help clarify a hidden pattern in other related research and point to exciting opportunities for future investigation: that this scalp location could be used as a tagged feature for better real-time cybersickness classification with EEG. We provide our code at: [anonymized].

Brain-Computer Interface (BCI) is a cutting-edge technology that facilitates human-computer interaction. Motor Imagery Electroencephalogram (MI-EEG) decoding technology has emerged as a significant direction in BCI research. Despite the remarkable advancements in deep learning for EEG signal decoding in recent years, two major challenges persist: the comprehensive representation and extraction of features, and the lack of interpretability. To address these issues, we propose a novel neurosymbolic framework termed MI-LTN (Motor Imagery Logic Tensor Network), incorporate logical constraints into the training model using the Logic Tensor Network (LTN) and employ Shapley values to evaluate and adjust the importance of channels. Our experimental results show that MI-LTN achieves classification accuracies of 86.00% and 88.84% on the BCI IV 2a and BCI IV 2b datasets, respectively. These results demonstrate the great potential of LTN in MI-EEG decoding.

Achieving both accurate and interpretable classification of motor-imagery EEG remains a key challenge in brain-computer interface (BCI) research. In this paper, we compare a transparent fuzzy-reasoning approach (ANFIS-FBCSP-PSO) with a well-known deep-learning benchmark (EEGNet) using the publicly available BCI Competition IV-2a dataset. The ANFIS pipeline combines filter-bank common spatial pattern feature extraction with fuzzy IF-THEN rules optimized via particle-swarm optimization, while EEGNet learns hierarchical spatial-temporal representations directly from raw EEG data. In within-subject experiments, the fuzzy-neural model performed better (68.58% +/- 13.76% accuracy, kappa = 58.04% +/- 18.43), while in cross-subject (LOSO) tests, the deep model exhibited stronger generalization (68.20% +/- 12.13% accuracy, kappa = 57.33% +/- 16.22). The study therefore provides practical guidance for selecting MI-BCI systems according to the design goal: interpretability or robustness across users. Future investigations into transformer-based and hybrid neuro-symbolic frameworks are expected to further advance transparent EEG decoding.

No abstract available

No abstract available

No abstract available

Patients with motor dysfunction show low subjective engagement in rehabilitation training. Traditional SSVEP-based brain-computer interface (BCI) systems rely heavily on external visual stimulus equipment, limiting their practicality in real-world settings. This study proposes an augmented reality steady-state visually evoked potential (AR-SSVEP) system to address the lack of patient initiative and the high workload on therapists. Firstly, we design four HoloLens 2-based EEG classes and collect EEG data from seven healthy subjects for analysis. Secondly, we build upon the conventional CNN-BiLSTM architecture by integrating a multi-head attention mechanism (MACNN-BiLSTM). We extract ten temporal-spectral EEG features and feed them into a CNN to learn high-level representations. Then, we use BiLSTM to model sequential dependencies and apply a multi-head attention mechanism to highlight motor-intention-related patterns. Finally, the SHAP (SHapley Additive exPlanations) method is applied to visualize EEG feature contributions to the neural network's decision-making process, enhancing the model's interpretability. These findings enhance real-time motor intention recognition and support recovery in patients with motor impairments.

BACKGROUND AND OBJECTIVE Electroencephalogram (EEG) signals record brain activity, with growing interest in quantifying neural activity through complexity analysis as a potential biological marker for schizophrenia. Presently, EEG complexity analysis primarily relies on manual feature extraction, which is subjective and yields varied findings in studies involving schizophrenia and healthy controls. METHODS This study aims to leverage deep learning methods for enhanced EEG complexity exploration, aiding early schizophrenia screening and diagnosis. Our proposed approach utilizes a three-dimensional Convolutional Neural Network (3DCNN) to extract enhanced data features for early schizophrenia identification and subsequent complexity analysis. Leveraging the spatiotemporal capabilities of 3DCNN, we extract advanced latent features and employ knowledge distillation to reintegrate these features into the original channels, creating feature-enhanced data. RESULTS We employ a 10-fold cross-validation strategy, achieving the average accuracies of 99.46% and 98.06% in subject-dependent experiments on Dataset 1(14SZ and 14HC) and Dataset 2 (45SZ and 39HC). The average accuracy for subject-independent is 96.04% and 92.67% on both datasets. Feature extraction and classification are conducted on both the re-aggregated data and the original data. Our results demonstrate that re-aggregated data exhibit superior classification performance and a more stable training process after feature extraction. In the complexity analysis of re-aggregated data, we observe lower entropy features in schizophrenic patients compared to healthy controls, with more pronounced differences in the temporal and frontal lobes. Analyzing Katz's Fractal Dimension (KFD) across three sub-bands of lobe channels reveals the lowest α band KFD value in schizophrenia patients. CONCLUSIONS This emphasizes the ability of our method to enhance the discrimination and interpretability in schizophrenia detection and analysis. Our approach enhances the potential for EEG-based schizophrenia diagnosis by leveraging deep learning, offering superior discrimination capabilities and richer interpretive insights.

No abstract available

No abstract available

No abstract available

Abstract The brain-computer interface (BCI) systems based on motor imagery typically rely on a large number of electrode channels to acquire information. The rational selection of electroencephalography (EEG) channel combinations is crucial for optimizing computational efficiency and enhancing practical applicability. However, evaluating all potential channel combinations individually is impractical. This study aims to explore a strategy for quickly achieving a balance between maximizing channel reduction and minimizing precision loss. To this end, we developed a spatio-temporal attention perception network named STAPNet. Based on the channel contributions adaptively generated by its subnetwork, we propose an extended step bi-directional search strategy that includes variable ratio channel selection (VRCS) and strided greedy channel selection (SGCS), designed to enhance global search capabilities and accelerate the optimization process. Experimental results show that on the High Gamma and BCI Competition IV 2a public datasets, the framework respectively achieved average maximum accuracies of 91.47% and 84.17%. Under conditions of zero precision loss, the average number of channels was reduced by a maximum of 87.5%. Additionally, to investigate the impact of neural information loss due to channel reduction on the interpretation of complex brain functions, we employed a heatmap visualization algorithm to verify the universal importance and complete symmetry of the selected optimal channel combination across multiple datasets. This is consistent with the brain’s cooperative mechanism when processing tasks involving both the left and right hands.

Electroencephalography (EEG)-based brain-computer interfaces (BCIs) pose a challenge for decoding due to their low spatial resolution and signal-to-noise ratio. Typically, EEG-based recognition of activities and states involves the use of prior neuroscience knowledge to generate quantitative EEG features, which may limit BCI performance. Although neural network-based methods can effectively extract features, they often encounter issues such as poor generalization across datasets, high predicting volatility, and low model interpretability. To address these limitations, we propose a novel lightweight multi-dimensional attention network, called LMDA-Net. By incorporating two novel attention modules designed specifically for EEG signals, the channel attention module and the depth attention module, LMDA-Net is able to effectively integrate features from multiple dimensions, resulting in improved classification performance across various BCI tasks. LMDA-Net was evaluated on four high-impact public datasets, including motor imagery (MI) and P300-Speller, and was compared with other representative models. The experimental results demonstrate that LMDA-Net outperforms other representative methods in terms of classification accuracy and predicting volatility, achieving the highest accuracy in all datasets within 300 training epochs. Ablation experiments further confirm the effectiveness of the channel attention module and the depth attention module. To facilitate an in-depth understanding of the features extracted by LMDA-Net, we propose class-specific neural network feature interpretability algorithms that are suitable for evoked responses and endogenous activities. By mapping the output of the specific layer of LMDA-Net to the time or spatial domain through class activation maps, the resulting feature visualizations can provide interpretable analysis and establish connections with EEG time-spatial analysis in neuroscience. In summary, LMDA-Net shows great potential as a general decoding model for various EEG tasks.

No abstract available

EEG-based recognition of activities and states involves the use of prior neuroscience knowledge to generate quantitative EEG features, which may limit BCI performance. Although neural network-based methods can effectively extract features, they often encounter issues such as poor generalization across datasets, high predicting volatility, and low model interpretability. Hence, we propose a novel lightweight multi-dimensional attention network, called LMDA-Net. By incorporating two novel attention modules designed specifically for EEG signals, the channel attention module and the depth attention module, LMDA-Net can effectively integrate features from multiple dimensions, resulting in improved classification performance across various BCI tasks. LMDA-Net was evaluated on four high-impact public datasets, including motor imagery (MI) and P300-Speller paradigms, and was compared with other representative models. The experimental results demonstrate that LMDA-Net outperforms other representative methods in terms of classification accuracy and predicting volatility, achieving the highest accuracy in all datasets within 300 training epochs. Ablation experiments further confirm the effectiveness of the channel attention module and the depth attention module. To facilitate an in-depth understanding of the features extracted by LMDA-Net, we propose class-specific neural network feature interpretability algorithms that are suitable for event-related potentials (ERPs) and event-related desynchronization/synchronization (ERD/ERS). By mapping the output of the specific layer of LMDA-Net to the time or spatial domain through class activation maps, the resulting feature visualizations can provide interpretable analysis and establish connections with EEG time-spatial analysis in neuroscience. In summary, LMDA-Net shows great potential as a general online decoding model for various EEG tasks.

With the deepening of neuroscience research, data mining of brain signals is becoming an emerging topic. Among various brain signals, electroencephalography (EEG) has attracted more and more attention due to its advantages of non-invasiveness, portability, and low cost. EEG modeling and analysis play a vital role in human healthcare. Although many machine learning algorithms have been successfully applied to data mining of EEG signals, few of them achieve a win-win in classification performance and interpretability. In this paper, we propose a Bayesian graph neural network named BayesEEGNet. Considering an electrical impulse between two nodes in the brain as a Poisson process, the countless electrical impulses generated by the brain in a period are represented as an infinite number of connection probability graphs. After coupling and transforming these probability graphs, we interpret the brain’s electrical activity state as the brain’s perceptual state. Benefiting from the joint optimization of Bayesian modules and deep neural networks, our model shows superior classification performance in sleep stage classification and emotion recognition tasks. Meanwhile, our model is able to learn interpretable functional connectivity relationships between EEG channels without any prior knowledge.

Drowsy driving has a crucial influence on driving safety, creating an urgent demand for driver drowsiness detection. Electroencephalogram (EEG) signal can accurately reflect the mental fatigue state and thus has been widely studied in drowsiness monitoring. However, the raw EEG data is inherently noisy and redundant, which is neglected by existing works that just use single-channel EEG data or full-head channel EEG data for model training, resulting in limited performance of driver drowsiness detection. In this paper, we are the first to propose an Interpretability-guided Channel Selection (ICS) framework for the driver drowsiness detection task. Specifically, we design a two-stage training strategy to progressively select the key contributing channels with the guidance of interpretability. We first train a teacher network in the first stage using full-head channel EEG data. Then we apply the class activation mapping (CAM) to the trained teacher model to highlight the high-contributing EEG channels and further propose a channel voting scheme to select the top N contributing EEG channels. Finally, we train a student network with the selected channels of EEG data in the second stage for driver drowsiness detection. Experiments are designed on a public dataset, and the results demonstrate that our method is highly applicable and can significantly improve the performance of cross-subject driver drowsiness detection.

A high cognitive load could significantly impair problem-solving skills. Electroencephalogram (EEG)-based real-time assessment of mental workload is feasible, and graph neural networks (GNN) can classify brain activity patterns during cognitively demanding tasks with high accuracy. However, previous GNN studies pertaining to mental workload classification lack explainability. This study utilized a state-of-the-art GNN variant with GNNexplainer to find relevant connectivity during mental arithmetic (MA) tasks. In this endeavor, MA EEG recordings were retrieved from an openaccess database. The signals were transformed to graph data through the envelope correlation and power spectral density (PSD), and subjected to GNN with hierarchical graph pooling with a structure learning model to classify MA and baseline (BL). The model accuracy was $85.57 \pm 6.27$ and $96.26 \pm 4.14$% for the connectivity dataset and the PSD and the connectivity feature, respectively. Among the connections between nodes identified as important by GNNExplainer, two notable edge patterns were found as 1) from the left centro-parietal region to left frontal regions, and 2) the frontoparietal connection. The results indicate 1) the GNN model performance could be improved using the connectivity and PSD feature together, and 2) characteristic patterns of the connectome and PSD could be important for MA classification. The connectivity analysis by the ‘‘explainable’’ GNN model could be beneficial in future brain activity pattern studies.

Understanding how local neurophysiological patterns interact with global brain dynamics is essential for decoding human emotions from EEG signals. However, existing deep learning approaches often overlook the brain's intrinsic spatial organization, failing to simultaneously capture local topological relations and global dependencies. To address these challenges, we propose Neuro-HGLN, a Neurologically-informed Hierarchical Graph-Transformer Learning Network that integrates biologically grounded priors with hierarchical representation learning. Neuro-HGLN first constructs a spatial Euclidean prior graph based on physical electrode distances to serve as an anatomically grounded inductive bias. A learnable global dynamic graph is then introduced to model functional connectivity across the entire brain. In parallel, to capture fine-grained regional dependencies, Neuro-HGLN builds region-level local graphs using a multi-head self-attention mechanism. These graphs are processed synchronously through local-constrained parallel GCN layers to produce region-specific representations. Subsequently, an iTransformer encoder aggregates these features to capture cross-region dependencies under a dimension-as-token formulation. Extensive experiments demonstrate that Neuro-HGLN achieves state-of-the-art performance on multiple benchmarks, providing enhanced interpretability grounded in neurophysiological structure. These results highlight the efficacy of unifying local topological learning with cross-region dependency modeling for robust EEG emotion recognition.

Methamphetamine (MA) is a highly addictive drug, and its abuse has become a global public health issue. Currently, most research on drug addiction severity is based on statistical scales, descriptions of drug abusers, and subjective judgments from rehabilitation doctors. In this study, participants were recruited and divided into two groups, mild and severe, based on suggestions from researchers and medical experts. An experimental paradigm was designed to collect EEG data. After conducting time-frequency analysis, significant differences around the occipital lobe between mild and severe drug abusers were discovered. Particularly, in the frequency domain analysis, the differences were most pronounced in the low-frequency range (delta band). In the paper, a new network model called AC-CNN is proposed based on a compact architecture of Convolutional Neural Network (CNN) for the classification of MA drug abusers. This model can adaptively learn the most relevant and useful features in the input data, thereby improving accuracy. Then, applying the Class Activation Mapping (CAM) method to interpret the classification results of the model is beneficial for understanding the decision-making process of the model in signal classification tasks, particularly in identifying the crucial regions that significantly contribute to classification results of the model.

Graph convolutional networks (GCNs) have shown great prowess in learning topological relationships among electroencephalogram (EEG) channels for EEG-based emotion recognition. However, most existing GCN-only methods are designed with a single spatial pattern, lacking connectivity enhancement within local functional regions and ignoring the data dependencies of EEG original data. In this article, hierarchical dynamic GCN (HD-GCN) is proposed to explore dynamic multilevel spatial information among EEG channels, with discriminative features of EEG signals as auxiliary information. Specifically, representation learning in topological space consists of two branches: one for extracting global dynamic information and one for exploring augmentation information in local functional regions. In each branch, a layerwise adjacency matrix is utilized to enrich the expressive power of GCN. Furthermore, a data-dependent auxiliary information module (AIM) is developed to capture multidimensional fusion features. Extensive experiments on two public datasets, SJTU emotion EEG dataset (SEED) and DREAMER, demonstrate that the proposed method consistently exceeds state-of-the-art methods. Interpretability analysis of the proposed model is performed, discovering the active brain regions and important electrode pairs related to emotion.

Current models on Explainable Artificial Intelligence (XAI) have shown a lack of reliability when evaluating feature-relevance for deep neural biomarker classifiers. The inclusion of reliable saliency-maps for obtaining trustworthy and interpretable neural activity is still insufficiently mature for practical applications. These limitations impede the development of clinical applications of Deep Learning. To address, these limitations we propose the RemOve-And-Retrain (ROAR) algorithm which supports the recovery of highly relevant features from any pre-trained deep neural network. In this study we evaluated the ROAR methodology and algorithm for the Face Emotion Recognition (FER) task, which is clinically applicable in the study of Autism Spectrum Disorder (ASD). We trained a Convolutional Neural Network (CNN) from electroencephalography (EEG) signals and assessed the relevance of FER-elicited EEG features from individuals diagnosed with and without ASD. Specifically, we compared the ROAR reliability from well-known relevance maps such as Layer-Wise Relevance Propagation, PatternNet, Pattern-Attribution, and Smooth-Grad Squared. This study is the first to bridge previous neuroscience and ASD research findings to feature-relevance calculation for EEG-based emotion recognition with CNN in typically-development (TD) and in ASD individuals.

Abstract Emotion recognition, a challenging computational issue, finds interesting applications in diverse fields. Usually, feature-based machine-learning methods have been used for emotion recognition. However, these conventional shallow machine learning methods often find unsatisfactory results as there is a tradeoff between feature dimensions and classification accuracy. Besides, extraction and selection of features from the spatial and frequency domains could be an additional issue. This work proposes a method that transforms EEG (electroencephalography) signals to topographic images that contain the frequency and spatial information and utilizes a convolutional neural network (CNN) to classify the emotion, as CNN has improved feature extraction capability. According to the proposed method, the topographic images are prepared from the relative power spectral density rather than power spectral density that shows remarkable improvement in classification accuracy. The proposed method is applied to the well-known SEED database and has given outperforming results than the current state-of-the-art.

Introduction Scalp electroencephalography (EEG) is commonly used to assist in epilepsy detection. Even automated detection algorithms are already available to assist clinicians in reviewing EEG data, many algorithms used for seizure detection in epilepsy fail to account for the contributions of different channels. The Fully Convolutional Network (FCN) can provide the model’s interpretability but has not been applied in seizure detection. Methods To address these challenges, a novel convolutional neural network (CNN) model, combining SE (Squeeze-and-Excitation) modules, was proposed on top of the FCN. The epilepsy detection performance for patient-independent was evaluated on the CHB-MIT dataset. Then, the SE module was removed from the model and integrated the model with Inception, ResNet, and CBAM modules separately. Results The method showed superior advancement, stability, and reliability compared to the other three methods. The method demonstrated a G-Mean of 82.7% for sensitivity (SEN) and specificity (SPE) on the CHB-MIT dataset. In addition, The contributions of each channel to the seizure detection task have also been quantified, which led us to find that the FZ, CZ, PZ, FT9, FT10, and T8 brain regions have a more pronounced impact on epileptic seizures. Discussion This article presents a novel algorithm for epilepsy detection that accurately identifies seizures in different patients and enhances the model’s interpretability.

Background Emotion recognition using EEG signals enables clinicians to assess patients’ emotional states with precision and immediacy. However, the complexity of EEG signal data poses challenges for traditional recognition methods. Deep learning techniques effectively capture the nuanced emotional cues within these signals by leveraging extensive data. Nonetheless, most deep learning techniques lack interpretability while maintaining accuracy. Methods We developed an interpretable end-to-end EEG emotion recognition framework rooted in the hybrid CNN and transformer architecture. Specifically, temporal convolution isolates salient information from EEG signals while filtering out potential high-frequency noise. Spatial convolution discerns the topological connections between channels. Subsequently, the transformer module processes the feature maps to integrate high-level spatiotemporal features, enabling the identification of the prevailing emotional state. Results Experiments’ results demonstrated that our model excels in diverse emotion classification, achieving an accuracy of 74.23% ± 2.59% on the dimensional model (DEAP) and 67.17% ± 1.70% on the discrete model (SEED-V). These results surpass the performances of both CNN and LSTM-based counterparts. Through interpretive analysis, we ascertained that the beta and gamma bands in the EEG signals exert the most significant impact on emotion recognition performance. Notably, our model can independently tailor a Gaussian-like convolution kernel, effectively filtering high-frequency noise from the input EEG data. Discussion Given its robust performance and interpretative capabilities, our proposed framework is a promising tool for EEG-driven emotion brain-computer interface.

Depression is a widespread mental health disorder affecting millions worldwide. While deep learning models have shown promise in detecting depression using neurophysiological data, lack of result interpretability hinders clinical adoption. In this work, a new method of combining a 2D Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM) depression diagnosis model based on EEG signals with Explainable AI (XAI) methods, specifically LIME. The integration provides transparency into decision-making of the model and mitigates the “black box” problem inherent in many AI systems. The results demonstrate the model’s high interpretability through clinically relevant explanations in model predictions in depression detection. This approach enhances the transparency and trustworthiness of AI-based depression diagnosis, potentially facilitating its integration into clinical practice. The paper discusses the implications of XAI in mental health diagnostics and its role in advancing responsible AI in healthcare.

In the context of electroencephalogram (EEG)-based driver drowsiness recognition, it is still challenging to design a calibration-free system, since EEG signals vary significantly among different subjects and recording sessions. Many efforts have been made to use deep learning methods for mental state recognition from EEG signals. However, existing work mostly treats deep learning models as black-box classifiers, while what have been learned by the models and to which extent they are affected by the noise in EEG data are still underexplored. In this article, we develop a novel convolutional neural network combined with an interpretation technique that allows sample-wise analysis of important features for classification. The network has a compact structure and takes advantage of separable convolutions to process the EEG signals in a spatial-temporal sequence. Results show that the model achieves an average accuracy of 78.35% on 11 subjects for leave-one-out cross-subject drowsiness recognition, which is higher than the conventional baseline methods of 53.40%–72.68% and state-of-the-art deep learning methods of 71.75%–75.19%. Interpretation results indicate the model has learned to recognize biologically meaningful features from EEG signals, e.g., alpha spindles, as strong indicators of drowsiness across different subjects. In addition, we also explore reasons behind some wrongly classified samples with the interpretation technique and discuss potential ways to improve the recognition accuracy. Our work illustrates a promising direction on using interpretable deep learning models to discover meaningful patterns related to different mental states from complex EEG signals.

Epilepsy is a common brain disease characterized by frequent occurrences. A reliable epilepsy prediction system can provide early warning, giving patients and healthcare professionals enough time to intervene to reduce patients’ suffering. Electroencephalograms(EEG) can record abnormal electrical activity in patients with epilepsy and are considered a powerful tool for the diagnosis and analysis of epilepsy. In our study, epilepsy prediction is modeled as a task of distinguishing the pre-seizure state from the normal brain state, and a classification model for epilepsy prediction based on EEG and deep learning is proposed. Traditional deep learning methods are mostly "black box" models, which do not explicitly explore the role of temporal and spatial features in EEG for seizure prediction tasks. In previous studies on epilepsy prediction, Limited attention has been given by researchers to the significance of different channels for each patient. Leading to the extraction and interference of useless information. To address the above issues and focus on extracting appropriate features, Our model extracts temporal and spatial features of the original EEG through two semi-isolated stages. Our model enhances more relevant features through attention transformation. Our proposed model achieves 97.6% accuracy and 98.5% sensitivity on the CHB dataset and 98.3% accuracy and 99% sensitivity on the SWEC dataset. It is superior to most existing epilepsy prediction models.

Deep learning models, particularly convolutional neural networks (CNNs), have shown promise in automated epileptic seizure detection from electroencephalogram (EEG). However, their “black-box” nature limits clinical adoption, as interpretability is critical for trust and validation in medical applications. A novel interpretability method for CNN-based seizure detection models, designed to uncover meaningful spatial and spectral EEG biomarkers, is proposed. The approach combines frequency- and spatial-domain interpretation to provide both global model behavior analysis and local, sample-specific explanations. It also accounts for the task-specific design, neurophysiological grounding and cross-framework validation — concepts often neglected by in many state-of-the-art methods. Results are represented as heatmap matrices of feature importance (5 frequency bands * 5 brain regions) with important features determined through statistical testing. Interpretation is based on neurophysiological alignment of these features. The method is validated on three CNN architectures, demonstrating how each leverages distinct frequency bands and brain regions for seizure identification. Global interpretation reveals that the highest-performing model utilizes complementary biomarkers across multiple frequency bands, while local interpretation captures dynamic intra-seizure spectral shifts. The results align with known neurophysiological mechanisms, such as thalamocortical interactions (theta-band) and default mode network suppression (alpha/beta-bands), while also suggesting new biomarkers for seizure detection. The method bridges the gap between deep learning and clinical EEG analysis, offering a tool for model validation and discovery of electrophysiological signatures in epilepsy.

Objective. This work presents xEEGNet, a novel, compact, and explainable neural network for electroencephalography (EEG) data analysis. It is fully interpretable and reduces overfitting through a major parameter reduction. Approach. As an applicative use case to develop our model, we focused on the classification of common dementia conditions, Alzheimer’s and frontotemporal dementia, versus controls. xEEGNet, however, is broadly applicable to other neurological conditions involving spectral alterations. We used ShallowNet, a simple and popular model in the EEGNet family, as a starting point. Its structure was analyzed and gradually modified to move from a ‘black box’ model to a more transparent one, without compromising performance. The learned kernels and weights were analyzed from a clinical standpoint to assess their medical significance. Model variants, including ShallowNet and the final xEEGNet, were evaluated using a robust nested-leave-n-subjects out cross-validation for unbiased performance estimates. Variability across data splits was explained using the embedded EEG representations, grouped by class and set, with pairwise separability to quantify group distinction. Overfitting was measured through training-validation loss correlation and training speed. Main results. xEEGNet uses only 168 parameters, 200 times fewer than ShallowNet, yet retains interpretability, resists overfitting, achieves comparable median performance (−1.5%), and reduces performance variability across splits. This variability is explained by the embedded EEG representations: higher accuracy correlates with greater separation between test-set controls and Alzheimer’s cases, without significant influence from the training data. Significance. The capability of xEEGNet to filter specific EEG bands, learns band specific topographies and use the right EEG spectral bands for disease classification demonstrates its interpretability. While big deep learning models are typically prioritized for performance, this study shows that smaller architectures like xEEGNet can be equally effective in pathology classification, using EEG data.

Deep learning has served pattern classification in many applications, with a performance which often well exceeds that of other machine learning paradigms. Yet, in general, deep learning has used computational architectures built, albeit partially, by ad hoc means, and its classification decisions are not necessarily interpretable in terms of knowledge relevant to the application it serves. This is often referred to as the black box problem, which in certain applications, such as epileptic seizure prediction, can be a serious impediment. The purpose of this study is to investigate an interpretable deep learning classifier for epileptic EEG-driven seizure prediction. This neural network is interpretable because its layers can be visualized and interpreted as a result of a novel architecture where the learned weights follow from signal processing computations such as frequency sub-band and spatial filters. Consequently, the extracted features are no longer abstract as they correspond to the features commonly used for decoding EEG data. In addition, the network uses layer-wise relevance propagation to reveal pertinent features which can further explain the computations leading to the decisions. In seizure prediction experiments using the CHB-MIT data set, the method produced classification results which improved on the state-of-the art, with first network layer filters corresponding to clinically relevant frequency bands, and the input channels in the brain location in which the seizure originates contributing most significantly to the network predictions.

With the rapid development of deep neural networks, significant progress has been made in the analysis of physiological signals such as EEG, greatly improving classification performance. However, the inherent “black-box” nature of these models limits our ability to understand how different brain regions coordinate to drive decisions-especially given that modern EEG architectures (e.g., EEGNet) use gating mechanisms that activate distinct computational pathways for different subjects. To address this, we introduce the MODET-based Interaction Backtracking Method (MIBM), which converts a standard EEGNet into a shallow, wide, and intrinsically interpretable Multi-Order Descartes Expansion Neural Network (MODENN). This “deep-to-broad” transformation makes high-order feature interactions explicit and traceable. Experimentally, the MIBMtransformed EEGNet preserves-or even surpasses-the original model's accuracy across subjects. By exposing each subject's gated pathways as a clear computation graph, MIBM substantially enhances transparency and trustworthiness, marking a paradigm shift from traditional single-feature attribution toward a holistic, network-level interpretability aligned with systems neuroscience.

Numerous deep learning-based methodologies have been proposed to facilitate automatic sleep stage classification tasks. Nevertheless, the black-box nature of these approaches is one of the skeptical factors hindering clinical application. Toward model interpretability, this study presents a novel interpretable sleep stage classification scheme based on layer-wise relevance propagation (LRP). We first adopt the short-time Fourier transform (STFT) to convert the raw electroencephalogram (EEG) signals to the time-frequency images, which could visually demonstrate EEG patterns of each sleep stage. Moreover, we introduce an efficient convolutional neural network (CNN)-based model, namely MSSENet, that assembles with the multiscale CNN (MSCNN) module and residual squeeze-and-excitation (R-SE) block for the image input. The LRP method is eventually applied to evaluate the contribution of each frequency pixel in the input time-frequency image to the model prediction. Experimental findings show that the MSSENet could outperform or achieve comparable performance to other state-of-the-art approaches on three polysomnography (PSG) datasets. Furthermore, through utilizing the heat mapping, the LRP-based explainability results validate the high relevance of specific EEG patterns to the prediction of the corresponding sleep stage, which is consistent with the sleep scoring guidelines.

Schizophrenia is an example of a rare mental disorder that is challenging to diagnose using conventional methods. Deep learning methods have been extensively employed to aid in the diagnosis of schizophrenia. However, their efficacy relies heavily on data quantity, and their black-box nature raises trust concerns, especially in medical diagnosis contexts. In this study, we leverage a deep learning-based method for the automatic diagnosis of schizophrenia using EEG brain recordings. This approach utilizes generative data augmentation, a powerful technique that enhances the accuracy of the diagnosis. Additionally, our study provides a framework to use when dealing with the challenge of limited training data for the diagnosis of other potential rare mental disorders. To enable the utilization of time-frequency features, spectrograms were extracted from the raw signals. After exploring several neural network architectural setups, a proper convolutional neural network (CNN) was used for the initial diagnosis. Subsequently, using Wasserstein GAN with Gradient Penalty (WGAN-GP) and Variational Autoencoder (VAE), two different synthetic datasets were generated in order to augment the initial dataset and address the over-fitting issue. The augmented dataset using VAE achieved a 3.0% improvement in accuracy, reaching 99.0%, and also demonstrated faster convergence. Finally, we addressed the lack of trust in black-box models using the Local Interpretable Model-agnostic Explanations (LIME) algorithm to determine the most important superpixels (frequencies) in the diagnosis process.

With an ever-growing demand for BCI-based systems, numerous algorithms and machine learning systems have been proposed over the past few decades. Although state-of-theart approaches have reached practically appropriate levels of accuracy, most are often regarded as black-box models, which need more explainability. However, for cases where a Neural Network is used as the base for a classifier, a Layerwise Relevance Propagation (LRP) approach can be utilized to analyze the decision boundaries considered by the network. By calculating the importance of the neuron in each layer, the LRP can also be used as an effective model complexity reduction technique through the inactivation (pruning) of the neural pathways. The following work investigates the usability of the LRP framework in the field of BCI. This study provides an example of the practical application of the LRP with respect to the EEG (ERP) dataset, along with visual heatmap and scalp map examples of the LRP. Furthermore, the work analyzes the impact of network pruning on heatmap visualization and the model’s accuracy while also practically determining the maximum cutoff range for pruning BCI models.

No abstract available

BACKGROUND Emotion recognition is increasingly essential for diagnosing mental disorders like depression and anxiety. Electroencephalography (EEG) is widely adopted for this purpose due to its high temporal resolution and non-invasive nature. However, existing EEG-based models often neglect the brain's dynamic neural connectivity, inadequately modeling spatial topology, and rely on extensive redundant data, increasing computational complexity and limiting performance. METHODS To address these limitations, we propose ELAI-SGCN, a lightweight and explainable framework for EEG analysis. ELAI-SGCN employs a trainable spiking encoder to transform raw EEG signals into sparse spike-based representations, preserving critical temporal dynamics in an event-driven manner to provide critical information for the subsequent Spiking Neural Network model. Simultaneously, the graph convolution module adaptively models inter-regional connectivity through efficient spike-based operations, enabling interpretable and resource-efficient EEG analysis. RESULTS We validated ELAI-SGCN on the DEAP and SEED emotion recognition datasets. On the DEAP dataset, the model achieved classification accuracies of 87.08 % for valence and 89.96 % for arousal, with only 60.48 K parameters and a computational cost of 0.36 M FLOPs. On the SEED dataset, it reached 94.63 % accuracy for three-class emotion recognition with 99.8 % fewer parameters and a more than 250-fold decrease in FLOPs compared to Dynamic Graph Convolutional Neural Networks. CONCLUSION ELAI-SGCN introduces a novel spiking-based dynamic graph convolutional framework that enables efficient and interpretable modeling of EEG spatiotemporal dynamics. ELAI-SGCN outperformed most existing approaches in both accuracy and computational efficiency, demonstrating its suitability for lightweight, real-time EEG-based emotion recognition. Its lightweight design supports deployment on bedside clinical devices, offering a promising solution for emotion recognition and laying a foundation for next-generation intelligent psychological assessment systems with broad clinical potential.

Deep learning-based emotion recognition using EEG has received increasing attention in recent years. The existing studies on emotion recognition show great variability in their employed methods including the choice of deep learning approaches and the type of input features. Although deep learning models for EEG-based emotion recognition can deliver superior accuracy, it comes at the cost of high computational complexity. Here, we propose a novel 3D convolutional neural network with a channel bottleneck module (CNN-BN) model for EEG-based emotion recognition, with the aim of accelerating the CNN computation without a significant loss in classification accuracy. To this end, we constructed a 3D spatiotemporal representation of EEG signals as the input of our proposed model. Our CNN-BN model extracts spatiotemporal EEG features, which effectively utilize the spatial and temporal information in EEG. We evaluated the performance of the CNN-BN model in the valence and arousal classification tasks. Our proposed CNN-BN model achieved an average accuracy of 99.1% and 99.5% for valence and arousal, respectively, on the DEAP dataset, while significantly reducing the number of parameters by 93.08% and FLOPs by 94.94%. The CNN-BN model with fewer parameters based on 3D EEG spatiotemporal representation outperforms the state-of-the-art models. Our proposed CNN-BN model with a better parameter efficiency has excellent potential for accelerating CNN-based emotion recognition without losing classification performance.

The brain age is a key indicator of brain health. While electroencephalography (EEG) is a practical tool for this task, existing models struggle with the common challenge of imperfect medical data, such as learning a ``normal''baseline from weakly supervised, healthy-only cohorts. This is a critical anomaly detection task for identifying disease, but standard models are often black boxes lacking an interpretable structure. We propose EVA-Net, a novel framework that recasts brain age as an interpretable anomaly detection problem. EVA-Net uses an efficient, sparsified-attention Transformer to model long EEG sequences. To handle noise and variability in imperfect data, it employs a Variational Information Bottleneck to learn a robust, compressed representation. For interpretability, this representation is aligned to a continuous prototype network that explicitly learns the normative healthy aging manifold. Trained on 1297 healthy subjects, EVA-Net achieves state-of-the-art accuracy. We validated its anomaly detection capabilities on an unseen cohort of 27 MCI and AD patients. This pathological group showed significantly higher brain-age gaps and a novel Prototype Alignment Error, confirming their deviation from the healthy manifold. EVA-Net provides an interpretable framework for healthcare intelligence using imperfect medical data.

Emotion recognition using electroencephalography (EEG) plays a critical role in affective computing applications such as mental health monitoring, brain-computer interfaces, and adaptive systems. While deep learning has advanced recognition performance, many existing approaches depend on hold-out evaluation strategies that limit generalization to unseen users. Additionally, computationally intensive models are often unsuitable for real-time or resource-constrained scenarios. To overcome these challenges, this study introduces a novel lightweight deep learning architecture, the Tiny Variational Residual Network (TVRN), designed for subject-independent EEG-based emotion recognition. TVRN operates directly on raw one-dimensional EEG feature vectors, enabling efficient end-to-end learning. Its architecture combines dense residual connections with a variational bottleneck, facilitating expressive yet compact feature representation. The model was evaluated using a leave-one-subject-out (LOSO) cross-validation scheme on two benchmark EEG datasets: GAMEEMO and LUMED. On the GAMEEMO dataset, TVRN achieved an accuracy of 80.40% for valence and 81.11% for arousal. On the LUMED dataset, it reached an accuracy of 93.22%. Compared to the baseline ResNet1D model, TVRN reduced the parameter count by over 75% and required only 306K FLOPs, underscoring its computational efficiency. These results highlight TVRN’s potential as a robust and scalable solution for EEG-based emotion recognition in real-world, low-resource environments.

Biometrics have attracted growing research interests as information security and safety gain increasing attention to date. As a kind of important biomedical signal, electroencephalogram (EEG) contains valuable information about identity, emotionality, personality, and so on. Thus, automatically distinguishing the identities based on EEG is beneficial to the development of biometrics, forensics, and informatics. Although deep learning has absorbed much research attention for the issue of EEG-based person identification, the performance enhancement of this methodology seems to have hit a bottleneck recently. Hence, by rethinking the problems haunting this issue, we plan to reinvigorate the conventional method pipeline and put forward a novel and effective tensorial scheme away from the deep learning mainstream. Specifically, the proposed tensorial scheme extracts the effective tensorial representation from multichannel EEG at first; then, the scheme performs the designed tensorial learning to improve the discriminability of the feature space; and finally, the scheme carries out the devised tensorial measurement in the learned metric space for classification. Experimental results have demonstrated the superiority of the proposed scheme over the related advanced approaches by means of the challenging benchmark databases: DEAP, SEED, and DREAMER.

Accurate classification of electroencephalogram (EEG) signals is crucial for brain activity understanding. However, EEG signals are characterized by data heterogeneity and label scarcity, which present a challenging low-data learning regime when building machine learning models. Existing methods tend to suffer from overfitting problem. To this end, we propose an adaptive Bayesian meta-learning framework for instance-specific learning and inference in EEG classification tasks. Specifically, first, a query set-driven dynamic parameter-based support set selection strategy is designed to adaptively fit the query set when constructing a meta-training task. Second, we employ an amortized variational inference network to generate task-specific adapted parameters given the support set, thereby achieving rapid model adaption for the inference of the data in the query set. Especially, a time- and frequency-aware representation learning encoder is leveraged to extract more task-relevant information guided by information bottleneck principle from time and frequency views, respectively, alleviating the low signal-to-noise ratio issue. Extensive experimental results on three public datasets demonstrate the superior effectiveness of our method.

No abstract available