A Multiscale Decoding Approach of Subject-Independent Motor Imagery EEG Signal Combined with Data Alignment Strategy

Brain-computer interface (BCI) leverages artificial intelligence (AI) and wearable electroencephalography (EEG) sensors to decode brain signals, significantly enhancing quality of life. EEG-based motor imagery (MI) brain signals are widely used in various BCI applications, including smart healthcare, intelligent vehicle, smart homes, and robotics control. However, the substantial individual variability in EEG data distribution poses a challenge in applying constructed models to new subjects.

To address this issue, this study proposes a novel deep learning framework, namely Multiscale Spatiotemporal Convolutional Neural Network (MSTCNN) based on Euclidean Space Data Alignment (ESDA), which can effectively decode MI tasks in subject-independent scenarios. First, to address the data distribution variability for different subjects, the proposed ESDA method is utilized to align all MI task data, reducing the discrepancies caused by physiological differences. Subsequently, an end-to-end multiscale model is constructed to extract multi-level MI task-related feature information from EEG signals. Additionally, the Squeeze-and-Excitation (SE) attention mechanism is incorporated to highlight critical local features, further improving MI task-decoding performance.

Experiments conducted on the publicly available BCI Competition IV dataset 2a demonstrate that the proposed multiscale model achieves an average accuracy of 67.5% in subject-independent scenarios, with a relatively small number of parameters (4.4 × 10³). Moreover, the highest decoding accuracy for individual subjects reaches 81.4%, surpassing many other state-of-the-art (SOTA) methods. The ablation experiments further validate the model's performance and robustness.

This study may provide a solid theoretical foundation for advancing BCI system applications in Internet of Things units, clinical, and other areas.