Normalized deep learning algorithms based information aggregation functions to classify motor imagery EEG signal

Recently, the discipline of Brain-Computer-Interface (BCI) has attracted attention to exploiting Electroencephalograph (EEG) mental activities such as Motor Imagery (MI). Neurons in the human brain are activated during these MI tasks and generate an electrical potential of small magnitude reached to the scalp as a signal. Classification of MI data is a primary problem in BCI systems. Classification accuracy of these biomedical signals emerges as a significant task in the scientific community. This work proposes two main ideas: a new preprocessing technique based on four EEG frequency bands and a new stacking method for three deep-learning architectures used to decode three classes of MI signals. The preprocessing stage was introduced using Fast Fourier Transform to perform frequency analysis and data aggregation functions to enhance the data view. Performance was evaluated using well-defined metrics: accuracy, precision, recall, and f1-score for multiple batch sizes, optimizers, and epochs. Experimental results were evaluated using a publicly available dataset (BCI Competition IV dataset 2a) and local data collected from four subjects using the EMOTIV EPOC headset. The highest f1-scores achieved with the R-CNN model were 94% and 84% using the aforementioned datasets. Our proposed models also outperform many related models studied in the literature.