Motor Imagery EEG Classification Using Fine-Tuned Deep Convolutional EfficientNetB0 Model

This work proposes a new way to apply the deep convolutional EfficientN etB0 model for the classification to learn various electroencephalogram (EEG) signal properties on BCI competition IV dataset 2b. Several deep convolutional neural networks (DCNN)-based techniques have been applied to enhance the accuracy of motor imagery-based brain-computer interfaces (BCIs). The time-varying nature of various frequency bands makes extracting informative features from EEG signals difficult, causing the loss of DCNN classification accuracy. The EfficientNetB0 baseline model's feature extraction capability is used to overcome these limitations by first transforming EEG 1D signals into 2D images using the feature extraction technique of the Short-Time Fourier Transform (STFT) algorithm to train and evaluate the EfficientNetB0 model. The transfer learning approach is used to expand the initial feature sets for efficient model training in a short period. After learning from the dataset, the entire model is retrained and fine-tuned to work with the proposed layers. Our evaluated results demonstrated that the highest average Accuracy, Precision, Recall, F1-Score and MCC with the STFT method is 86.46%,88.2%, 91.2%,89.78% and 0.815 respectively over 10 epochs. According to the results, the proposed methodology outperforms other state-of-the-art DCNN models for feature extraction and classification of two-class motor imagery, namely right-hand and left-hand movements for the given dataset.