Table 1_DeepAttNet: deep neural network incorporating cross-attention mechanism for subject-independent mental stress detection in passive brain–computer interfaces using bilateral ear-EEG.docx

IntroductionElectroencephalography (EEG)-based mental stress detection has the potential to be applied in diverse real-world scenarios, including workplace safety, mental health monitoring, and human–computer interaction. However, most previous passive brain–computer interface (BCI) studies have employed EEG recorded during the performance of specific tasks, making the classification results susceptible to task engagement effects rather than reflecting stress alone. To address this limitation, we introduce a rest-versus-rest paradigm that compares resting EEG recorded immediately after exposure to a stressor with that recorded after meditation, thereby isolating mental stress from the task-related confounds. EEG recording setups were designed under the assumption of bilateral ear-EEG, a compact and discreet form factor suitable for real-world applications. Furthermore, we developed a novel subject-independent deep learning classifier tailored to model interhemispheric neural dynamics for enhanced mental stress detection performance.
MethodsThirty-two adults participated in the experiment. To classify mental stress status in a subject-independent manner, we proposed DeepAttNet, a deep learning model based on cross-attention and pointwise temporal compression, specifically designed to effectively capture left and right hemispherical interactions. Classification performance was assessed using eight-fold subject-level cross-validation against conventional deep learning models, including EEGNet, ShallowConvNet, DeepConvNet, and TSception. Ablation studies evaluated the impact of the cross-attention and/or pointwise compression modules.
ResultsDeepAttNet achieved the highest average accuracy and macro-F1 values, with performance declining when either the cross-attention or pointwise compression module was removed in the ablation studies. Explainability analyses indicated lower cross-attention entropy with stronger directional ear-to-ear asymmetry under stress, and temporal occlusion identified mid–late windows supporting stress decisions. Moreover, six of seven canonical scalp-EEG markers were FDR-significant for post-stressor vs. post-relaxation rest.
ConclusionThe proposed rest-versus-rest paradigm and DeepAttNet enabled robust, subject-independent mental stress detection with a fairly high accuracy using only two-channel EEG recordings. This approach is expected to offer a practical solution for continuous stress monitoring, potentially advancing passive BCI applications outside laboratory settings.

引言
基于脑电图（Electroencephalography, EEG）的心理压力检测具备在多元现实场景中落地的潜力，涵盖职场安全、心理健康监测及人机交互等领域。然而，既往多数被动式脑机接口（brain–computer interface, BCI）研究均采用特定任务执行过程中采集的脑电图数据，致使分类结果易受任务参与度的干扰，而非仅反映单纯的心理压力状态。为解决这一局限，本研究提出“静息-静息”范式：将应激暴露后即刻采集的静息态脑电图，与冥想后采集的静息态脑电图进行对比，从而将心理压力与任务相关的混淆因素分离开来。脑电图采集装置基于双侧耳电（bilateral ear-EEG）的设计思路，该方案具备紧凑、隐蔽的外形优势，适配现实场景的应用需求。此外，本研究还开发了一款全新的受试者无关型深度学习分类器，专门针对半球间神经动力学建模，以提升心理压力检测的性能。

方法
本实验共招募32名成年受试者。为实现受试者无关的心理压力状态分类，本研究提出DeepAttNet模型——一种基于交叉注意力与逐时间点压缩的深度学习模型，其专门设计用于有效捕获左右半球间的交互作用。本研究采用8折受试者水平交叉验证，将该模型的分类性能与EEGNet、ShallowConvNet、DeepConvNet及TSception等经典深度学习模型进行对比。同时通过消融实验，评估交叉注意力模块和/或逐时间点压缩模块对模型性能的影响。

结果
DeepAttNet取得了最高的平均准确率与宏F1值；在消融实验中，移除交叉注意力模块或逐时间点压缩模块均会导致模型性能下降。可解释性分析显示，应激状态下的交叉注意力熵更低，且耳间不对称性更强；时间域遮挡实验证实，中晚期时间窗口对压力分类决策具有关键支撑作用。此外，7个典型头皮脑电图标记中有6个在应激后与放松后静息态的对比中达到错误发现率（False Discovery Rate, FDR）显著性水平。

结论
本研究提出的“静息-静息”范式与DeepAttNet模型，仅通过双通道脑电图采集即可实现鲁棒的、受试者无关的心理压力检测，且准确率较高。该方案有望为持续压力监测提供切实可行的解决方案，有望推动被动式脑机接口在实验室外场景的应用落地。