BCI result

Augmented Covariance Approaches

Augmented Covariance Method (ACM)

Augmented Covariance Method in a reduced dataset setting

Augmented SPDNet (AugSPDNet)

The application of Deep Learning (DL) algorithms has gained considerable attention across diverse domains, including EEG motor imagery (MI) classification. Despite DL’s success in fields like Computer Vision, its adoption in MI classification has been limited by several challenges such as limited data availability, high signal-to-noise ratio in EEG signals, and subject/session variability. The emergence of SPDNet, integrating Convolutional Neural Networks with Riemannian geometry, signifies a promising advancement.

This study explores advancements in EEG motor imagery classification by combining the Augmented Covariance Method (ACM) with SPDNet and introducing a novel approach, AugSPDNet applied in a reduced dataset setting.

This approach can be carried out using different SPD matrices as input, in particular we test the Covariance, the Imaginary and Instantaneous Coherence.

We test this algorithm in a reduced dataset setting, using only C4, Cz and C5 electrodes, over several datasets extracted from MOABB library. We focus in particular on right hand vs left hand in a Within-Session evaluation procedure.

Preliminary results find AugSPDNet, with covariance as estimator, with better performances respect to the state-of-the-art DL pipelines. Moreover this algorithm create less carbon footprint, require less trainable parameter and is explainable using GradCam++.