FFT-based deep feature learning method for EEG classification

This study introduces a new method for electroencephalogram (EEG) signal classification based on deep learning model, by which relevant features are automatically learned in a supervised learning framework. The fast Fourier transform (FFT) has been applied in a novel way to generate the EEG matrix. And a PCA neural network (PCANet) is designed to learn the hidden information from the frequency matrix of EEG signals. And these deep features are then given as inputs to train a support vector machine (SVM) for recognition of epileptic seizures. The experiments are carried out with two authoritative databases provided by the Bonn University (Database A) and Children's Hospital in Boston (Database B), relatively. Additionally, we have evaluated the influence of all parameters for the proposed scheme to obtain the optimal model with better generalization and expansibility. The proposed feature learning method concerned in this work is proved very useful to distinguish seizure events from both short and long EEG recordings. Experimental results obtained by analyzing Database A are not less than 99% accuracy in seven problems. The effectiveness is also verified on Database B with an average accuracy of 98.47% across 23 patients. Our FFT-based PCANet not only achieves the satisfied results, but also exhibits better stability across different classification cases or patients, which indicates the worth in practical applications for diagnostic reference in clinics.