# BRI-001: A Hybrid Brain-Computer Interface Using Motor Imagery and SSVEP Based on Convolutional Neural Network > 本页由本地论文卡片生成。请不要直接编辑本页；修改源卡片后运行 `vp run docs:generate`。 > 中文支持说明：本页是中文站点镜像，保留论文标题、证据状态和来源字段的原始表述，避免未经核实的翻译改变证据边界。 > 内部 wiki 边界：本页提供论文卡片证据和用户提供的 PDF 下载入口，仅用于私有阅读；全文抽取文本不发布。 > 来源: `library/paper_cards/BRI-001.md` ## 论文访问 * 内部 PDF: 下载论文 PDF * DOI / 官方页面: [needs confirmation](https://arxiv.org/abs/2212.05289) * 开放访问页面: [Open access page](https://arxiv.org/abs/2212.05289) * 部署边界: 这些 PDF 链接只适合私有/受保护的内部 wiki。 ## BRI-001: A Hybrid Brain-Computer Interface Using Motor Imagery and SSVEP Based on Convolutional Neural Network ## Metadata * ID: BRI-001 * Title: A Hybrid Brain-Computer Interface Using Motor Imagery and SSVEP Based on Convolutional Neural Network * Year: 2022 * DOI / URL: https://arxiv.org/abs/2212.05289 * Local PDF: 见上方论文访问区块 * Text artifact: local-only path withheld from docs site * Review status: `extracted` ## Study Type * Track: Hybrid BCI / Brain-Robot Interface * Task: offline neural decoding for MI, SSVEP, and hybrid MI-SSVEP binary classification * Participants or dataset: Korea University EEG dataset with 54 subjects performing binary left/right MI and four-class SSVEP tasks; 40 subjects used for training and 14 for testing in this study * Hardware: BrainAmp amplifier with 62 Ag/AgCl electrodes sampled at 1,000 Hz; the text later refers to "64-channel EEGs," so channel count consistency needs confirmation * Channels or sensors: EEG only; MI stream uses 20 motor-region electrodes and SSVEP stream uses 10 occipital/posterior electrodes ## Methods * Paradigm: MI trials cue imagined left- or right-hand grasping; SSVEP trials present four flicker frequencies, with 6.67 Hz and 8.57 Hz used for left/right binary classification; hybrid mode combines corresponding MI and SSVEP activity * Signal processing or model: 8-30 Hz 5th-order Butterworth filtering for MI, 0-4 s EEG segments, task-specific channel selection, and a two-stream CNN with MI and SSVEP blocks fused for sigmoid binary classification * Training/calibration: TSCNN1 trained on hybrid-mode EEG only; TSCNN2 trained on both single-mode and hybrid-mode EEG; 10-fold cross-validation is performed in training, then 14 held-out subjects are used for testing * Online/offline: offline supervised decoding using the training phase of session 1; no online robotic control or closed-loop BCI experiment is reported ## Results * Metrics: decoding accuracy, sensitivity, specificity, and mean square error * Main findings: TSCNN2 reports 70.2% MI accuracy, 93.0% SSVEP accuracy, and 95.6% hybrid accuracy on held-out subjects; the authors report hybrid-mode improvement over MI and SSVEP modes * Reported limitations: 4-second EEG epochs may be too long for effective BCI temporal resolution; the model is fully supervised and needs substantial labeled data; future work is needed on shorter epochs, additional expert-annotated datasets, and stronger fusion architectures ## Relevance To This Project * Supports: hybrid SSVEP-MI decoding can outperform single-mode decoding in an offline EEG benchmark, and task-specific motor/occipital channel selection is a reasonable baseline * Conflicts with: the study is binary left/right classification, not scene-object selection, YOLO-driven command generation, shared-control grasping, or an online robot experiment * Design implication: TSCNN-style two-stream fusion is a useful decoder baseline for SAH-BRI-Grasp, but project experiments must validate shorter windows, online false activations, dynamic SSVEP targets, and robot task outcomes ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | The proposed model combines MI and SSVEP in a two-stream CNN. | verified | The abstract and method sections describe TSCNN with separate MI and SSVEP blocks followed by fusion. | Abstract; Model Architecture | | The study uses a 54-subject EEG dataset with binary MI and four-class SSVEP tasks. | verified | The dataset description states 54 subjects, binary-class MI, four-class SSVEP, two sessions, and session-1 training data usage. | Materials and Methods: Dataset Description | | The recorded EEG setup is BrainAmp at 1,000 Hz with 62 Ag/AgCl electrodes, though the text also refers to 64-channel EEGs. | verified | The dataset section gives the amplifier/electrode count and sampling rate; the feature-representation section later says selected channels came from 64-channel EEGs. | Dataset Description; Feature Representation | | TSCNN2 achieved 70.2% MI, 93.0% SSVEP, and 95.6% hybrid decoding accuracy. | verified | The results section reports averaged accuracy, sensitivity, specificity, and MSE for TSCNN2 in the three modes. | Results: Decoding Performance | | The paper identifies 4-second epochs and supervised labeled-data dependence as limitations. | verified | The limitations section says shorter epoch lengths are needed and that TSCNN is fully supervised and requires large labeled datasets. | Discussion: Limitations and Future Works | | TSCNN-style MI-SSVEP fusion is relevant as a decoder baseline for SAH-BRI-Grasp. | inferred | The paper supports hybrid EEG decoding, but it does not test scene-aware target boxes or robot grasping. | Full paper | | Online shared-control grasp performance from this model remains unverified. | needs confirmation | The local text reports offline decoding, not a real-time robot manipulation study. | Full paper scope | ## Open Questions * Is the source dataset 62-channel or 64-channel after preprocessing, and how should that be represented in project channel-ablation plans? * Can comparable hybrid accuracy be achieved with shorter windows suitable for responsive target confirmation? * How does TSCNN behave online with asynchronous no-control/rest periods and false activation costs? * Can the binary left/right setup transfer to multi-object dynamic SSVEP target selection plus MI confirmation?