Multiscale space-time-frequency feature-guided multitask learning CNN for motor imagery EEG classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) electroencephalography (EEG) classification is crucial for brain-computer interface (BCI) systems.
• Accurate decoding of neural activity is challenging due to low signal-to-noise ratios and non-stationarity.
• Existing methods often overlook vital frequency domain information present in raw EEG signals.

Purpose of the Study:

• To develop a novel convolutional neural network (CNN) architecture for enhanced MI-EEG classification.
• To integrate complementary features from space-time and time-frequency domains.
• To leverage multitask learning for improved generalization and accuracy in BCI tasks.

Main Methods:

• A multiscale space-time-frequency feature-guided multitask learning CNN was developed.
• The architecture comprises four modules: space-time feature representation, time-frequency feature representation, multimodal fused feature generation, and classification.
• The framework utilizes multitask learning, training four modules simultaneously across three tasks for joint optimization.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art techniques on three public EEG datasets.
• Quantitative analysis confirmed the robustness and effectiveness of the developed approach.
• The method was successfully applied to control a robot using EEG signals, validating its real-time application feasibility.

Conclusions:

• The novel deep CNN architecture effectively fuses complementary input features for BCI tasks.
• Multitask learning within the architecture improves both subject-dependent and subject-independent classification accuracy, even with limited data.
• This approach offers a significant advancement in EEG-based BCI system performance and generalization.