Tianyou Yu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Tianyou Yu has a total of 102 co-authors. Here are the top co-authors by number of shared publications:
- Yuanqing Li
South China University of Technology
18 shared publications - Zhenghui Gu
School of Automation Science and Engineering, South China University of Technology, Guangzhou, Guangdong, China
16 shared publications - Zhuliang Yu
Shien-Ming Wu School of Intelligent Engineering, South China University of Technology, Guangzhou, China
14 shared publications - Wei Wu
Songjiang Hospital and Songjiang Research Institute, Shanghai Key Laboratory of Emotions and Affective Disorders, Shanghai Jiao Tong University School of Medicine, Shanghai, China
6 shared publications - Jun Xiao
School of Electric Power Engineering, South China University of Technology, Guangzhou, China
5 shared publications - Jun Zhang
Guangdong University of Technology
4 shared publications
Latest Publications
- 31 August 2026
- Published by Elsevier in Neurocomputing
- Vol. 705
Neural network-based dynamic functional connectivity estimation for motor imagery decoding from EEG
- 31 August 2026
- Published by Elsevier in Neurocomputing
- Vol. 705
Neural network-based dynamic functional connectivity estimation for motor imagery decoding from EEG
- Open Access
- 31 August 2026
- Published by Elsevier in SSRN Electronic Journal
A Dynamic Polysomnographic Sleep-State Signature Discriminates Early-Stage Multiple System Atrophy from Parkinson's Disease
Background: Multiple system atrophy (MSA) and Parkinson's disease (PD) are clinically similar yet prognostically distinct synucleinopathies, and their early differential diagnosis remains one of the most challenging problems in movement disorder medicine. Sleep disturbances are common in both disorders, yet the diagnostic value of whole-night polysomnography (PSG) beyond conventional sleep staging remains underexplored. Methods: We developed a generalizable machine learning framework that derives a dynamic PSG signature for MSA--PD discrimination from overnight recordings in a discovery cohort with clinically confirmed diagnoses (MSA = 141, PD = 104), with MSA diagnosed according to the 2022 MDS criteria for MSA and PD diagnosed according to the MDS clinical diagnostic criteria for Parkinson's disease. Multimodal features extracted from 30-s PSG epochs were organized into data-driven dynamic states, enabling state-resolved characterization of sleep microstructure. The signature was evaluated on an independent validation cohort (MSA = 69, PD = 28) acquired over a separate period. Findings: The dynamic PSG signature achieved an area under the receiver operating characteristic curve (AUC) of 0.858 for MSA--PD discrimination in the independent validation cohort. In the clinically critical early-stage setting, the dynamic PSG signature discriminated early-stage MSA from early-stage PD with an AUC of 0.831 on the corresponding validation subset. Decision-curve analysis further demonstrated positive net benefit for the multimodal PSG models across clinically relevant threshold probabilities. Feature stability analyses identified a compact and reproducible multimodal PSG signature, while individual-level state trajectories revealed disease-specific alterations in sleep microstructure not captured by conventional staging or summary metrics. Interpretation: These findings highlight dynamic PSG as a non-invasive, clinically accessible, and interpretable biomarker candidate for early MSA--PD discrimination and emphasize the value of state-resolved sleep dynamics in neurodegenerative disease stratification.
- 31 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Dynamic Polysomnographic Sleep-State Signature Discriminates Early-Stage Multiple System Atrophy from Parkinson's Disease
Background: Multiple system atrophy (MSA) and Parkinson's disease (PD) are clinically similar yet prognostically distinct synucleinopathies, and their early differential diagnosis remains one of the most challenging problems in movement disorder medicine. Sleep disturbances are common in both disorders, yet the diagnostic value of whole-night polysomnography (PSG) beyond conventional sleep staging remains underexplored. Methods: We developed a generalizable machine learning framework that derives a dynamic PSG signature for MSA--PD discrimination from overnight recordings in a discovery cohort with clinically confirmed diagnoses (MSA = 141, PD = 104), with MSA diagnosed according to the 2022 MDS criteria for MSA and PD diagnosed according to the MDS clinical diagnostic criteria for Parkinson's disease. Multimodal features extracted from 30-s PSG epochs were organized into data-driven dynamic states, enabling state-resolved characterization of sleep microstructure. The signature was evaluated on an independent validation cohort (MSA = 69, PD = 28) acquired over a separate period. Findings: The dynamic PSG signature achieved an area under the receiver operating characteristic curve (AUC) of 0.858 for MSA--PD discrimination in the independent validation cohort. In the clinically critical early-stage setting, the dynamic PSG signature discriminated early-stage MSA from early-stage PD with an AUC of 0.831 on the corresponding validation subset. Decision-curve analysis further demonstrated positive net benefit for the multimodal PSG models across clinically relevant threshold probabilities. Feature stability analyses identified a compact and reproducible multimodal PSG signature, while individual-level state trajectories revealed disease-specific alterations in sleep microstructure not captured by conventional staging or summary metrics. Interpretation: These findings highlight dynamic PSG as a non-invasive, clinically accessible, and interpretable biomarker candidate for early MSA--PD discrimination and emphasize the value of state-resolved sleep dynamics in neurodegenerative disease stratification.
- 2 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33  (10709908) , 2465-2469
Band-Specific Graph Learning for EEG-Based Emotion Recognition
Electroencephalogram (EEG)-based emotion recognition has attracted increasing attention due to the high temporal resolution of EEG signals. However, accurate EEG-based emotion recognition remains challenging because EEG signals inherently exhibit low signal-to-noise ratios, non-stationarity and considerable inter-individual variability. In addition, brain functional connectivity presents distinct band-dependent characteristics, whereas many existing graph-based methods employ shared or static graphs across spectral bands, limiting their ability to characterize band-specific inter-channel interactions. To address these issues, we propose BSGNet, a band-specific graph learning framework for EEG-based emotion recognition. Specifically, instead of assuming a unified connectivity structure, BSGNet learns a set of band-specific channel adjacency matrices to capture functional interactions associated with different spectral bands. The learned graphs are further integrated into a STGCN backbone to jointly capture spatial dependencies among EEG channels and temporal dynamics of emotional responses. Experimental results on two public EEG emotion recognition datasets show that BSGNet achieves improved average performance over representative deep learning and graph-based baselines, supporting the effectiveness of band-specific connectivity modeling for EEG representation learning.
- 2 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33  (10709908) , 2465-2469
Band-Specific Graph Learning for EEG-Based Emotion Recognition
Electroencephalogram (EEG)-based emotion recognition has attracted increasing attention due to the high temporal resolution of EEG signals. However, accurate EEG-based emotion recognition remains challenging because EEG signals inherently exhibit low signal-to-noise ratios, non-stationarity and considerable inter-individual variability. In addition, brain functional connectivity presents distinct band-dependent characteristics, whereas many existing graph-based methods employ shared or static graphs across spectral bands, limiting their ability to characterize band-specific inter-channel interactions. To address these issues, we propose BSGNet, a band-specific graph learning framework for EEG-based emotion recognition. Specifically, instead of assuming a unified connectivity structure, BSGNet learns a set of band-specific channel adjacency matrices to capture functional interactions associated with different spectral bands. The learned graphs are further integrated into a STGCN backbone to jointly capture spatial dependencies among EEG channels and temporal dynamics of emotional responses. Experimental results on two public EEG emotion recognition datasets show that BSGNet achieves improved average performance over representative deep learning and graph-based baselines, supporting the effectiveness of band-specific connectivity modeling for EEG representation learning.
- Open Access
- 18 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Engineering
- Vol. PP  (00189294) , 1-14
Assessing Disorders of Consciousness Using Temporal Sleep Dynamics Extracted From Whole-Night PSG
Accurate assessment of patients with disorders of consciousness (DoC) remains a major clinical challenge due to the limitations of behavior-based evaluations and task-dependent neurophysiological paradigms. Whole-night polysomnography (PSG), a passive and noninvasive monitoring tool, offers unique potential for revealing residual brain function during sleep. In this study, we propose a temporal-dynamic feature extraction and aggregation framework for PSG analysis to enable machine learning-based diagnosis and prognosis in DoC patients. Whole-night EEG/EOG signals were segmented into non-overlapping 30-second epochs, from which time-domain, spectral, and nonlinear complexity features were extracted. To obtain a unified and compact representation of variable-length feature sequences, two aggregation strategies were applied: stage- wise averaging based on sleep staging and clustering-based grouping via unsupervised learning. A two-stage feature selection pipeline further reduced dimensionality while preserving discriminative power and interpretability. Classifiers trained on the aggregated features achieved strong performance in distinguishing minimally conscious state (MCS) from vegetative state (VS), with AUC values exceeding 0.84, and demonstrated robust predictive ability for long-term recovery outcomes (AUC=0.79). These findings highlight the diagnostic and prognostic value of whole-night PSG and support the development of fully automated, task-free assessment tools for DoC.
- 18 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Engineering
- Vol. PP  (00189294) , 1-14
Open AccessAssessing Disorders of Consciousness Using Temporal Sleep Dynamics Extracted From Whole-Night PSG
Accurate assessment of patients with disorders of consciousness (DoC) remains a major clinical challenge due to the limitations of behavior-based evaluations and task-dependent neurophysiological paradigms. Whole-night polysomnography (PSG), a passive and noninvasive monitoring tool, offers unique potential for revealing residual brain function during sleep. In this study, we propose a temporal-dynamic feature extraction and aggregation framework for PSG analysis to enable machine learning-based diagnosis and prognosis in DoC patients. Whole-night EEG/EOG signals were segmented into non-overlapping 30-second epochs, from which time-domain, spectral, and nonlinear complexity features were extracted. To obtain a unified and compact representation of variable-length feature sequences, two aggregation strategies were applied: stage- wise averaging based on sleep staging and clustering-based grouping via unsupervised learning. A two-stage feature selection pipeline further reduced dimensionality while preserving discriminative power and interpretability. Classifiers trained on the aggregated features achieved strong performance in distinguishing minimally conscious state (MCS) from vegetative state (VS), with AUC values exceeding 0.84, and demonstrated robust predictive ability for long-term recovery outcomes (AUC=0.79). These findings highlight the diagnostic and prognostic value of whole-night PSG and support the development of fully automated, task-free assessment tools for DoC.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33, 3421-3425
A Hybrid Periodic Decomposition-Based Network for MI-EEG Decoding
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33, 3421-3425
A Hybrid Periodic Decomposition-Based Network for MI-EEG Decoding
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 30  (10) , 1-14
Efficient Sleep Staging With Bayesian Uncertainty-Guided Active Learning
Automated sleep staging is essential for large-scale and home-based sleep monitoring; however, in routine clinical practice, sleep annotation remains largely dependent on experienced experts performing time-consuming and labor-intensive manual scoring. Existing automatic systems often struggle to adapt reliably to new subjects, limiting their clinical adoption and reinforcing the reliance on expert review. This creates a strong demand for adaptive and efficient sleep staging systems that can substantially reduce annotation workload while preserving expert-level accuracy. We propose BayesSleepNet, a novel framework that integrates Bayesian uncertainty quantification with active learning for adaptive sleep staging. BayesSleepNet employs principled Bayesian modeling by placing distributions over network weights and performing Monte Carlo sampling at inference, enabling explicit quantification of model (epistemic) uncertainty. These uncertainty estimates drive a two-stage sample selection strategy that first fine-tunes the model using representative epochs and subsequently prioritizes persistently uncertain samples for expert review. Across four public sleep datasets, BayesSleepNet consistently improves performance-by 7.60% in accuracy, 8.27% in macro-F1, and 0.104 in Cohen's $\kappa$-while requiring manual annotation of only 20% of data from new subjects. Despite its adaptive learning capability, BayesSleepNet remains computationally lightweight, using substantially fewer parameters than representative high-capacity state-of-the-art models. These results demonstrate the clinical promise of uncertainty-aware active learning as a practical and cost-efficient paradigm for semi-automated sleep staging. Code is available at https://github.com/yuty2009/bayesugal.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 30  (10) , 1-14
Efficient Sleep Staging With Bayesian Uncertainty-Guided Active Learning
Automated sleep staging is essential for large-scale and home-based sleep monitoring; however, in routine clinical practice, sleep annotation remains largely dependent on experienced experts performing time-consuming and labor-intensive manual scoring. Existing automatic systems often struggle to adapt reliably to new subjects, limiting their clinical adoption and reinforcing the reliance on expert review. This creates a strong demand for adaptive and efficient sleep staging systems that can substantially reduce annotation workload while preserving expert-level accuracy. We propose BayesSleepNet, a novel framework that integrates Bayesian uncertainty quantification with active learning for adaptive sleep staging. BayesSleepNet employs principled Bayesian modeling by placing distributions over network weights and performing Monte Carlo sampling at inference, enabling explicit quantification of model (epistemic) uncertainty. These uncertainty estimates drive a two-stage sample selection strategy that first fine-tunes the model using representative epochs and subsequently prioritizes persistently uncertain samples for expert review. Across four public sleep datasets, BayesSleepNet consistently improves performance-by 7.60% in accuracy, 8.27% in macro-F1, and 0.104 in Cohen's $\kappa$-while requiring manual annotation of only 20% of data from new subjects. Despite its adaptive learning capability, BayesSleepNet remains computationally lightweight, using substantially fewer parameters than representative high-capacity state-of-the-art models. These results demonstrate the clinical promise of uncertainty-aware active learning as a practical and cost-efficient paradigm for semi-automated sleep staging. Code is available at https://github.com/yuty2009/bayesugal.
- 19 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Affective Computing
- Vol. 16  (4) , 3231-3247
EmotionMIL: An End-to-End Multiple Instance Learning Framework for Emotion Recognition From EEG Signals
Emotion recognition from EEG signals offers significant advantages in affective computing, as EEG more accurately reflects internal emotional states than other modalities, such as facial expressions or peripheral physiological signals. Modeling and capturing subtle affective changes over time is crucial for real-world applications to achieve better human-computer interaction. However, training such models usually requires segment-level emotion labels, which are costly and may not be feasible. Assigning the overall label to all EEG segments within a trial can lead to inaccurate model training and degraded performance, as emotions evolve continuously. This highlights the need for models capable of learning from trial-wise emotion labels while capturing temporal dynamics of emotional responses within each segment because trial-wise post-stimulus labels are more accessible. To this end, we propose EmotionMIL, an end-to-end EEG-based emotion recognition framework that leverages recent advances in deep multiple instance learning (MIL). This framework enables robust emotion recognition from weakly labeled EEG signals and identifies the most prominent emotional responses. EmotionMIL captures the temporal dynamics of emotions using a retentive self-attention mechanism, which adaptively assigns weights to EEG segments based on their relevance in predicting the overall emotion label. A pseudo-bag augmentation strategy is also introduced to enhance the model's generalization ability by generating additional pseudo-bags from the original ones. Evaluated on three benchmark datasets—DEAP, DREAMER, and SEED—EmotionMIL outperforms state-of-the-art non-MIL and MIL models in both subject-dependent and subject-independent tasks, achieving superior accuracy and F1-score. Ablation study further validates the model design, while visualization results demonstrate that EmotionMIL effectively identifies both spatial EEG patterns and temporal emotional dynamics. These findings underscore EmotionMIL's potential for robust, interpretable emotion recognition, paving the way for real-world applications in emotion-aware systems. The code is available at https://github.com/yuty2009/emotionmil.
- 19 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Affective Computing
- Vol. 16  (4) , 3231-3247
EmotionMIL: An End-to-End Multiple Instance Learning Framework for Emotion Recognition From EEG Signals
Emotion recognition from EEG signals offers significant advantages in affective computing, as EEG more accurately reflects internal emotional states than other modalities, such as facial expressions or peripheral physiological signals. Modeling and capturing subtle affective changes over time is crucial for real-world applications to achieve better human-computer interaction. However, training such models usually requires segment-level emotion labels, which are costly and may not be feasible. Assigning the overall label to all EEG segments within a trial can lead to inaccurate model training and degraded performance, as emotions evolve continuously. This highlights the need for models capable of learning from trial-wise emotion labels while capturing temporal dynamics of emotional responses within each segment because trial-wise post-stimulus labels are more accessible. To this end, we propose EmotionMIL, an end-to-end EEG-based emotion recognition framework that leverages recent advances in deep multiple instance learning (MIL). This framework enables robust emotion recognition from weakly labeled EEG signals and identifies the most prominent emotional responses. EmotionMIL captures the temporal dynamics of emotions using a retentive self-attention mechanism, which adaptively assigns weights to EEG segments based on their relevance in predicting the overall emotion label. A pseudo-bag augmentation strategy is also introduced to enhance the model's generalization ability by generating additional pseudo-bags from the original ones. Evaluated on three benchmark datasets—DEAP, DREAMER, and SEED—EmotionMIL outperforms state-of-the-art non-MIL and MIL models in both subject-dependent and subject-independent tasks, achieving superior accuracy and F1-score. Ablation study further validates the model design, while visualization results demonstrate that EmotionMIL effectively identifies both spatial EEG patterns and temporal emotional dynamics. These findings underscore EmotionMIL's potential for robust, interpretable emotion recognition, paving the way for real-world applications in emotion-aware systems. The code is available at https://github.com/yuty2009/emotionmil.
- 15 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Engineering
- Vol. 72  (10) , 3056-3070
Multi-View Self-Supervised Learning Enhances Automatic Sleep Staging From EEG Signals
Deep learning-based methods for automatic sleep staging offer an efficient and objective alternative to costly manual scoring. However, their reliance on extensive labeled datasets and the challenge of generalization to new subjects and datasets limit their widespread adoption. Self-supervised learning (SSL) has emerged as a promising solution to address these issues by learning transferable representations from unlabeled data. This study highlights the effectiveness of SSL in automated sleep staging, utilizing a customized SSL approach to train a multi-view sleep staging model. This model includes a temporal view feature encoder for raw EEG signals and a spectral view feature encoder for time-frequency features. During pretraining, we incorporate a cross-view contrastive loss in addition to a contrastive loss for each view to learn complementary features and ensure consistency between views, enhancing the transferability and robustness of learned features. A dynamic weighting algorithm balances the learning speed of different loss components. Subsequently, these feature encoders, combined with a sequence encoder and a linear classifier, enable sleep staging after finetuning with labeled data. Evaluation on three publicly available datasets demonstrates that finetuning the entire SSL-pretrained model achieves competitive accuracy with state-of-the-art methods—86.4%, 83.8%, and 85.5% on SleepEDF-20, SleepEDF-78, and MASS datasets, respectively. Notably, our framework achieves near-equivalent performance with only 5% of the labeled data compared to full-label supervised training, showcasing SSL's potential to enhance automated sleep staging efficiency.
- 15 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Engineering
- Vol. 72  (10) , 3056-3070
Multi-View Self-Supervised Learning Enhances Automatic Sleep Staging From EEG Signals
Deep learning-based methods for automatic sleep staging offer an efficient and objective alternative to costly manual scoring. However, their reliance on extensive labeled datasets and the challenge of generalization to new subjects and datasets limit their widespread adoption. Self-supervised learning (SSL) has emerged as a promising solution to address these issues by learning transferable representations from unlabeled data. This study highlights the effectiveness of SSL in automated sleep staging, utilizing a customized SSL approach to train a multi-view sleep staging model. This model includes a temporal view feature encoder for raw EEG signals and a spectral view feature encoder for time-frequency features. During pretraining, we incorporate a cross-view contrastive loss in addition to a contrastive loss for each view to learn complementary features and ensure consistency between views, enhancing the transferability and robustness of learned features. A dynamic weighting algorithm balances the learning speed of different loss components. Subsequently, these feature encoders, combined with a sequence encoder and a linear classifier, enable sleep staging after finetuning with labeled data. Evaluation on three publicly available datasets demonstrates that finetuning the entire SSL-pretrained model achieves competitive accuracy with state-of-the-art methods—86.4%, 83.8%, and 85.5% on SleepEDF-20, SleepEDF-78, and MASS datasets, respectively. Notably, our framework achieves near-equivalent performance with only 5% of the labeled data compared to full-label supervised training, showcasing SSL's potential to enhance automated sleep staging efficiency.
- 24 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Engineering
- Vol. 72  (1) , 401-412
Supervised Contrastive Learning-Based Domain Generalization Network for Cross-Subject Motor Decoding
Developing an electroencephalogram (EEG)-based motor imagery and motor execution (MI/ME) decoding system that is both highly accurate and calibration-free for cross-subject applications remains challenging due to domain shift problem inherent in such scenario. Recent research has increasingly embraced transfer learning strategies, especially domain adaptation techniques. However, domain adaptation becomes impractical when the target subject data is either difficult to obtain or unknown. To address this issue, we propose a supervised contrastive learning-based domain generalization network (SCLDGN) for cross-subject MI/ME decoding. Firstly, the feature encoder is purposefully designed to learn the EEG discriminative feature representations. Secondly, the domain alignment based on deep correlation alignment constrains the representations distance across various domains to learn domain-invariant features. In addition, the class regularization block is proposed, where the supervised contrastive learning with domain-agnostic mixup is established to learn the class-relevant features and achieve class-level alignment. Finally, in the latent space, clusters of domain-agnostic representations from the same class are mapped closer together. Consequently, SCLDGN is capable of learning domain-invariant and class-relevant discriminative representations, which are essential for effective cross-subject decoding. Extensive experiments conducted on six MI/ME datasets demonstrate the effectiveness of the proposed method in comparison with other state-of-the-art approaches. Furthermore, ablation study and visualization analyses explain the generalization mechanism of the proposed method and also show neurophysiologically meaningful patterns related to MI/ME.
- 24 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Engineering
- Vol. 72  (1) , 401-412
Supervised Contrastive Learning-Based Domain Generalization Network for Cross-Subject Motor Decoding
Developing an electroencephalogram (EEG)-based motor imagery and motor execution (MI/ME) decoding system that is both highly accurate and calibration-free for cross-subject applications remains challenging due to domain shift problem inherent in such scenario. Recent research has increasingly embraced transfer learning strategies, especially domain adaptation techniques. However, domain adaptation becomes impractical when the target subject data is either difficult to obtain or unknown. To address this issue, we propose a supervised contrastive learning-based domain generalization network (SCLDGN) for cross-subject MI/ME decoding. Firstly, the feature encoder is purposefully designed to learn the EEG discriminative feature representations. Secondly, the domain alignment based on deep correlation alignment constrains the representations distance across various domains to learn domain-invariant features. In addition, the class regularization block is proposed, where the supervised contrastive learning with domain-agnostic mixup is established to learn the class-relevant features and achieve class-level alignment. Finally, in the latent space, clusters of domain-agnostic representations from the same class are mapped closer together. Consequently, SCLDGN is capable of learning domain-invariant and class-relevant discriminative representations, which are essential for effective cross-subject decoding. Extensive experiments conducted on six MI/ME datasets demonstrate the effectiveness of the proposed method in comparison with other state-of-the-art approaches. Furthermore, ablation study and visualization analyses explain the generalization mechanism of the proposed method and also show neurophysiologically meaningful patterns related to MI/ME.
- Open Access
- 10 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Neural Systems and Rehabilitation Engineering
- Vol. 31  (15344320) , 3988-3998
A Multi-Domain Convolutional Neural Network for EEG-Based Motor Imagery Decoding
Motor imagery (MI) decoding plays a crucial role in the advancement of electroencephalography (EEG)-based brain-computer interface (BCI) technology. Currently, most researches focus on complex deep learning structures for MI decoding. The growing complexity of networks may result in overfitting and lead to inaccurate decoding outcomes due to the redundant information. To address this limitation and make full use of the multi-domain EEG features, a multi-domain temporal-spatial-frequency convolutional neural network (TSFCNet) is proposed for MI decoding. The proposed network provides a novel mechanism that utilize the spatial and temporal EEG features combined with frequency and time-frequency characteristics. This network enables powerful feature extraction without complicated network structure. Specifically, the TSFCNet first employs the MixConv-Residual block to extract multiscale temporal features from multi-band filtered EEG data. Next, the temporal-spatial-frequency convolution block implements three shallow, parallel and independent convolutional operations in spatial, frequency and time-frequency domain, and captures high discriminative representations from these domains respectively. Finally, these features are effectively aggregated by average pooling layers and variance layers, and the network is trained with the joint supervision of the cross-entropy and the center loss. Our experimental results show that the TSFCNet outperforms the state-of-the-art models with superior classification accuracy and kappa values (82.72% and 0.7695 for dataset BCI competition IV 2a, 86.39% and 0.7324 for dataset BCI competition IV 2b). These competitive results demonstrate that the proposed network is promising for enhancing the decoding performance of MI BCIs.
- 10 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Neural Systems and Rehabilitation Engineering
- Vol. 31  (15344320) , 3988-3998
Open AccessA Multi-Domain Convolutional Neural Network for EEG-Based Motor Imagery Decoding
Motor imagery (MI) decoding plays a crucial role in the advancement of electroencephalography (EEG)-based brain-computer interface (BCI) technology. Currently, most researches focus on complex deep learning structures for MI decoding. The growing complexity of networks may result in overfitting and lead to inaccurate decoding outcomes due to the redundant information. To address this limitation and make full use of the multi-domain EEG features, a multi-domain temporal-spatial-frequency convolutional neural network (TSFCNet) is proposed for MI decoding. The proposed network provides a novel mechanism that utilize the spatial and temporal EEG features combined with frequency and time-frequency characteristics. This network enables powerful feature extraction without complicated network structure. Specifically, the TSFCNet first employs the MixConv-Residual block to extract multiscale temporal features from multi-band filtered EEG data. Next, the temporal-spatial-frequency convolution block implements three shallow, parallel and independent convolutional operations in spatial, frequency and time-frequency domain, and captures high discriminative representations from these domains respectively. Finally, these features are effectively aggregated by average pooling layers and variance layers, and the network is trained with the joint supervision of the cross-entropy and the center loss. Our experimental results show that the TSFCNet outperforms the state-of-the-art models with superior classification accuracy and kappa values (82.72% and 0.7695 for dataset BCI competition IV 2a, 86.39% and 0.7324 for dataset BCI competition IV 2b). These competitive results demonstrate that the proposed network is promising for enhancing the decoding performance of MI BCIs.