Qi Li

Surface electromyography (sEMG)-based gesture recognition has attracted considerable attention in intelligent prosthesis control, human–computer interaction, and rehabilitation assistance. However, practical deployment remains challenging because new users can usually provide only a few labeled samples for calibration. Under cross-subject and posture-varying conditions, this setting further aggravates distribution shifts and can destabilize target-domain adaptation. To address these issues, this paper proposes a cross-subject sEMG gesture-recognition framework integrating training-time augmentation, two-stage meta-transfer learning, and matched feature regularization. Model-agnostic meta-learning (MAML) is first used on source-subject data to learn an initialization for rapid transfer, after which target-subject fine-tuning is performed with an auxiliary class-consistent feature constraint. Experiments were conducted on a self-collected 11-subject multi-posture dataset using a leave-one-subject-out (LOSO) protocol under FT-1_full, which uses one complete calibration repetition, and FT-2_k10, which uses 10 windows per gesture per posture. The proposed MAML + FT + MATCHED method achieved higher average Accuracy and Macro-F1 than Source-pretrain + FT. Under FT-2_k10, the average improvements were 9.20 and 9.70 percentage points, respectively. Nevertheless, the primary subject-level paired Wilcoxon comparisons did not reach statistical significance (Accuracy: p = 0.2402; Macro-F1: p = 0.2402; Holm-adjusted p = 1.0000 for both). The results therefore indicate favorable average trends and positive effect sizes, while pairwise statistical superiority was not established in the present 11-subject cohort.

Speech-based depression recognition is an important problem in paralinguistic speech analysis and affective computing. Depressive speech may contain heterogeneous acoustic dynamics across temporal scales, including slow-varying prosodic and contextual patterns as well as short-term local spectral-temporal fluctuations. However, many existing neural approaches encode these heterogeneous patterns within a unified continuous representation space, which may weaken the explicit modeling of scale-specific acoustic dynamics. This paper proposes WESGMamba, a wavelet-guided heterogeneous temporal modeling framework for speech-based depression recognition. One-dimensional Haar discrete wavelet transform is applied to Mel-spectrogram sequences to decompose them into low-frequency approximation components and high-frequency detail components. The approximation components are routed to a Mamba branch for continuous long-range acoustic context modeling, whereas the detail components are routed to a spiking neural network branch for event-driven local transient modeling. To improve the optimization of the spiking branch, a wavelet-inspired surrogate gradient, termed WESG, is designed by combining a Gaussian scaling function with a Mexican Hat wavelet. A cross-temporal consistency constraint is further introduced to regularize temporal output states. Experiments on EATD-Corpus and DAIC-WOZ are conducted under an audio-only subject-level evaluation protocol. WESGMamba achieves Macro-F1 scores of 74.80% on EATD-Corpus and 80.11% on the DAIC-WOZ development set, outperforming reproduced audio-only baselines. Ablation studies, noise perturbation analysis, and qualitative visualizations further demonstrate the effectiveness of the proposed cross-scale acoustic temporal modeling strategy.

Head Mounted Displays (HMDs) employ mixed reality (MR) technologies across various fields, affording users immersive visual experiences. However, prolonged use of HMDs devices to watch stereoscopic content may lead to visual fatigue. This study explores stereo-visual fatigue induced by viewing MR visual stimuli, focusing on EEG and peripheral physiological signals. We recorded 24-channel EEG signals from 23 healthy participants and combined them with electrocardiogram (ECG), electrodermal activity (EDA), peripheral oxygen saturation (SpO2), respiratory, and skin temperature data from 13 participants. These physiological signal characteristics were extracted and combined into multimodal datasets. These datasets have shown significant effects in identifying visual fatigue through experimental verification. The study highlights the need for publicly available datasets to further investigate stereo-visual fatigue in an MR environment.

Sleep stage classification is essential for diagnosing sleep disorders and assessing sleep quality. However, achieving accurate classification remains a challenge because of the non-Euclidean spatial distribution of electroencephalography (EEG) electrodes. Existing methods often focus on a single dimension (e.g., the temporal, spectral, or spatial domain) or at most two dimensions of EEG signals, failing to fully capture their temporal-spectral-spatial multidimensional features and dynamic correlations. Additionally, the heterogeneity of and complex interactions between brain regions as well as individual variability further hinder robust classification. To tackle this challenge, this paper presents a new heterogeneous graph adaptive neural network (THSSleepNet), which employs temporal-spectral-spatial multidimensional feature fusion. The network represents EEG signals as a heterogeneous graph sequence structure, incorporating the characteristics of their temporal-spatial and spectral-spatial correlations. THSSleepNet comprises two branches, temporal-spatial streams and spectral-spatial streams, enabling comprehensive feature extraction across the temporal, spectral, and spatial dimensions. The model incorporates local and global temporal-spatial and spectral-spatial heterogeneity of brain regions using the dynamic multiscale path generation module. Additionally, the graph attention module captures intricate interactions among brain regions, while the temporal/spectral adaptive module adaptively accounts for cross-scale dynamic context dependence across temporal-spectral dimensions. Subsequently, a hierarchical feature pyramid fusion module is employed to fuse temporal-spectral-spatial features of EEG signals. In addition, a domain adversarial learning mechanism mitigates the effect of individual variability on classification performance. The experimental results indicate that THSSleepNet surpasses current methods on the publicly available datasets, demonstrating its great potential for use in sensor-based EEG signal analysis and sleep monitoring.

Automated seizure detection based on scalp electroencephalography (EEG) can significantly accelerate the epilepsy diagnosis process. However, most existing deep learning-based epilepsy detection methods are deficient in mining the local features and global time series dependence of EEG signals, limiting the performance enhancement of the models in seizure detection. Our study proposes an epilepsy detection model, CMFViT, based on a Multi-Stream Feature Fusion (MSFF) strategy that fuses a Convolutional Neural Network (CNN) with a Vision Transformer (ViT). The model converts EEG signals into time-frequency domain images using the Tunable Q-factor Wavelet Transform (TQWT), and then utilizes the CNN module and the ViT module to capture local features and global time-series correlations, respectively. It fuses different feature representations through the MSFF strategy to enhance its discriminative ability, and finally completes the classification task through the average pooling layer and the fully connected layer. The effectiveness of the model was validated by experimental evaluations on the publicly available CHB-MIT dataset and the Kaggle 121 people epilepsy dataset. The model achieved 98.85% classification accuracy and other excellent metrics in single-subject experiments on the CHB-MIT dataset, and also demonstrated strong performance in cross-subject experiments on the Kaggle dataset. Ablation experiments demonstrate the complementary roles of the CNN and ViT modules, and their integration significantly improves detection accuracy and generalization. Comparisons with other methods highlight the advantages of the CMFViT model. The CMFViT model provides an efficient, accurate, and innovative solution for complex EEG signal analysis and seizure detection tasks for single and cross-subjects while laying the foundation for developing real-time, accurate seizure detection systems.

Introduction: Extended viewing of 3D content can induce fatigue symptoms. Thus, fatigue assessment is crucial for enhancing the user experience and optimizing the performance of stereoscopic 3D technology. Functional near-infrared spectroscopy (fNIRS) has emerged as a promising tool for evaluating 3D visual fatigue by capturing hemodynamic responses within the cerebral region. However, traditional fNIRS-based methods rely on manual feature extraction and analysis, limiting their effectiveness. To address these limitations, a deep learning model based on fNIRS was constructed for the first time to evaluate 3D visual fatigue, enabling end-to-end automated feature extraction and classification.Methods: Twenty normal subjects participated in this study (mean age: 24.6 ± 0.88 years; range: 23–26 years; 13 males). This paper proposed an fNIRS-based experimental paradigm that acquires data under both comfort and fatigue conditions. Given the time-series nature of fNIRS data and the variability of fatigue responses across different brain regions, a dual-branch convolutional network was constructed to separately extract temporal and spatial features. A transformer was integrated into the convolutional network to enhance long-range feature extraction. Furthermore, to adaptively select fNIRS hemodynamic features, a channel attention mechanism was integrated to provide a weighted representation of multiple features.Results: The constructed model achieved an average accuracy of 93.12% within subjects and 84.65% across subjects, demonstrating its superior performance compared to traditional machine learning models and deep learning models.Discussion: This study successfully constructed a novel deep learning framework for the automatic evaluation of 3D visual fatigue using fNIRS data. The proposed model addresses the limitations of traditional methods by enabling end-to-end automated feature extraction and classification, eliminating the need for manual intervention. The integration of a transformer module and channel attention mechanism significantly enhanced the model’s ability to capture long-range dependencies and adaptively weight hemodynamic features, respectively. The high classification accuracy achieved within and across subjects highlights the model’s effectiveness and generalizability. This framework not only advances the field of fNIRS-based fatigue assessment but also provides a valuable tool for improving user experience in stereoscopic 3D applications. Future work could explore the model’s applicability to other types of fatigue assessment and further optimize its performance for real-world scenarios.