Brain Computer Interfact EEG-based Motor Imagery Prediction

Past project (2020-2021) P.I. Dr. Azim Eskandarian

Brain-computer interface (BCI) is a pathway for direct communication between human brains and external devices. EEG-based motor imagery is a no-invasive mental execution for body rehabilitation. The objective of this project is to implement conventional feature extraction methods and deep learning-based methods to estimate and predict human motor imagery resutls

Time-frequency Common Spatial Pattern Feature Extraction Algorithm Paper: A Computationally Efficient Multiclass Time-Frequency Common Spatial Pattern Analysis on EEG Motor Imagery

Time-Frequency Common Spatial Pattern Architecture. The EEG signal is decomposed by a short-time fourier transform algorithm for time-frequency analysis. Then, the optimal freqency and temporal band are processed by the common spatial pattern feature extraction algorithm. Finally, a classifier is applied to classify EEG motor imagery results
"Left hand" class C3 and C4 channel
"Right hand" class C3 and C4 channel

We used an open-source dataset BCI Compeition IV Dataset and the classification results outperform other popular motor imagery classification algorithms

Deep learning-based method for motor imagery classification (EEG-Inception Network) Paper: EEG-Inception: An Accurate and Robust End-to-End Neural Network for EEG-based Motor Imagery Classification

EEG-Inception Network Architecture

To overcome data shortage and improve the algorithm robustness, we have developed a data augmentation method by applying EEG noises to the signal.

Data augmentation method. We extract the high frequency noises from the sample data and randomly applied them to other samples.

The classification results ranked 2nd place on the BCI Competition IV dataset 2a (binary classes) and SOTA and the BCI Competition IV Dataset 2b (multiple classes)

Classification accuracy at BCI Competition IV 2a and 2b dataset