Abstract
A brain-computer interface (BCI) based on motor imagery (MI) from the same limb can provide an intuitive control pathway but has received limited attention. It is still a challenge to classify multiple MI tasks from the same limb. The goal of this study is to propose a novel decoding method to classify the MI tasks of four joints of the same upper limb and the resting state. EEG signals were collected from 20 participants. A time-distributed attention network (TD-Atten) was proposed to adaptively assign different weights to different classes and frequency bands of the input multiband Common Spatial Pattern (CSP) features. The long short-term memory (LSTM) and dense layers were then used to learn sequential information from the reweight features and perform the classification. Our proposed method outperformed other baseline and deep learning-based methods and obtained the accuracies of 46.8% in the 5-class scenario and 53.4% in the 4-class scenario. The visualization results of attention weights indicated that the proposed framework can adaptively pay attention to alpha-band related features in MI tasks, which was consistent with the analysis of brain activation patterns. These results demonstrated the feasibility and interpretability of the attention mechanism in MI decoding and the potential of this fine MI paradigm to be applied for the control of a robotic arm or a neural prosthesis.
Introduction
Most current Brain-Computer Interfaces (BCIs) based on motor imagery (MI) rely on "gross" movements, such as distinguishing between left-hand and right-hand imagery. While effective, this approach creates a cognitive disconnection for users who wish to control a specific prosthetic limb naturally. The "holy grail" of BCI research is same-limb MI, where a user imagines moving specific joints—hand, wrist, elbow, or shoulder—to control a corresponding robotic joint.
However, decoding these signals is exceptionally difficult. Because these movements originate from the same limb, their neural signatures are concentrated in the same small area of the brain's contralateral motor cortex. This leads to significant spatial overlap and low signal-to-noise ratios in EEG data, making traditional classification methods insufficient.
Proposed Method: The TD-Atten Framework
To address these challenges, researchers developed the Time-Distributed Attention Network (TD-Atten). This framework is designed to extract and prioritize the most discriminative information from complex EEG signals through a multi-stage pipeline:
- Feature Extraction: The system uses a sliding window to segment EEG data, applying Filter-Bank Common Spatial Patterns (FBCSP) to extract spatial features across 16 frequency bands (4–38 Hz).
- Adaptive Attention: The core innovation is the TD-Atten module, which employs two types of attention. Class Attention weights features based on their ability to distinguish specific tasks, while Band Attention focuses on the most relevant frequency ranges (such as alpha or beta rhythms) for the current input.
- Temporal Modeling: Because MI is a dynamic process, the reweighted features are processed by a Long Short-Term Memory (LSTM) network to capture how brain activity evolves over the 4-second imagery period.
Experimental Results
The study evaluated the model using a dataset of 20 healthy participants performing 5-class MI tasks (hand, wrist, elbow, shoulder, and rest). The TD-Atten model significantly outperformed traditional baselines like FBCSP+SVM and other deep learning architectures like EEGNet.
- Classification Accuracy: The model achieved 46.8% accuracy in the 5-class scenario and 53.4% in a 4-class scenario. Given that the chance level for the 5-class task is approximately 23%, these results represent a substantial improvement.
- Class Performance: As shown in the confusion matrix, the model was particularly effective at identifying the "rest" state (72% accuracy) and the "hand" imagery (51% accuracy). The most significant confusion occurred between adjacent joints, such as the wrist and elbow, reflecting the spatial overlap in the motor cortex.
Discussion and Neurophysiological Alignment
A key finding of this research is the interpretability of the attention mechanism. When visualizing the learned weights, the model naturally prioritized the alpha (10-14 Hz) and beta (14-30 Hz) bands. This aligns with established neuroscience regarding Event-Related Desynchronization (ERD) during motor tasks, suggesting the model is learning genuine physiological patterns rather than noise.
Furthermore, ablation studies highlighted the necessity of the temporal component. Replacing the LSTM with a standard dense layer caused accuracy to drop from 46.8% to 34.5%, confirming that the sequential evolution of the EEG signal is critical for distinguishing fine-grained movements within the same limb.
Conclusion
The TD-Atten framework demonstrates the feasibility of decoding fine-grained motor imagery from a single limb. While current accuracy levels require further improvement before clinical deployment, the model's ability to adaptively focus on relevant neural features provides a robust foundation for future intuitive BCI systems. Future work will focus on improving decoder stability over long-term use and testing the system in closed-loop, real-time environments with patient populations.
Summary updated