Enhancing the Capability and Accuracy of Motor Imagery Classification: A Deep Neural Network-Powered Multifaceted...

Area of Science:

• Neuroscience and Biomedical Engineering
• Brain-Computer Interface (BCI) Technology
• Machine Learning in Healthcare

Background:

• Motor imagery (MI) is a key noninvasive brain-computer interface (BCI) method.
• Electroencephalogram (EEG) signal nonstationarity and low signal-to-noise ratio limit current MI decoding accuracy.
• Existing frameworks often miss complex spatio-temporal dependencies in EEG data.

Purpose of the Study:

• To introduce DPMS-Net, a deep neural network model designed to improve EEG signal decoding for BCIs.
• To address limitations in feature representation and capture latent discriminative cues in EEG.
• To enhance the accuracy and robustness of motor imagery decoding, particularly for neurorehabilitation applications.

Main Methods:

• Developed DPMS-Net, a deep neural network employing dynamic convolution for multi-dimensional feature extraction (temporal, spatial, frequency).
• Integrated channel and temporal attention mechanisms to capture salient EEG features across diverse spatio-temporal scales.
• Incorporated a spectral-domain analysis component to identify subtle oscillatory signatures within the EEG spectrum.

Main Results:

• DPMS-Net achieved high subject-dependent classification accuracies: 83.93% (BCI Comp IV 2a) and 88.38% (BCI Comp IV 2b).
• Subject-independent accuracies reached 65.88% (BCI Comp IV 2a) and 76.01% (BCI Comp IV 2b).
• On a stroke patient dataset, DPMS-Net achieved 67.67% (subject-dependent) and 57.58% (subject-independent) accuracy.

Conclusions:

• DPMS-Net demonstrates efficient and stable decoding capabilities for motor imagery tasks using EEG signals.
• The model's ability to exploit multi-dimensional features and attention mechanisms overcomes limitations of prior methods.
• DPMS-Net shows significant potential for practical deployment in neurorehabilitation BCI systems.