Complex-Valued Multi-Domain Features and Its Application in Motor Imagery Classification
Yabing Li, Zhenbo Sun, Zhenhua Wang, Kun Song · IEEE Access · 2025
The extraction of features from electroencephalography (EEG) signals is a vital step in the classification of motor imagery. However, traditional feature extraction methods for EEG signals either focus either on amplitude or frequency, or the rich phase information. This inability to simultaneously integrate phase information results in the loss of certain implicit details. Hence, effectively merging multiple pieces of features with phase data to establish a complex value feature extraction model is a significant challenge. Based on this, we propose a new hybrid framework combining complex-value multi-domain features (CVF) and complex-valued extreme learning machine network (CVELM), namely, CVF-CVELM to classify the motor imagery tasks. First of all, some multi-domain features are extracted. These features are then encoded into complex domain space based on phase-locked value, which captures the phase information of specific channels and time points. Furthermore, to align with the feature vectors in complex space, we developed a CVELM. Significance testing and separability analysis are employed for real-value features and CVF, the experimental results manifest the effectiveness of the CVF. Meanwhile, we evaluate our proposed method on two publicly available datasets from the representative fields of BCI, notably motor imagery decoding. The proposed model yielded better performance compared with existing state-of-the-art methods in terms of overall classification accuracy and kappa coefficient. The empirical evaluations demonstrate that the features in the complex domain contain richer information, which could be utilized in online BCI systems in the future.