A Brain–Computer Interface Framework Based on Compressive Sensing and Deep Learning

Abstract

Compression of brain–computer interface (BCI) signals is significant to reduce transmission bandwidth to cloud/remote servers and to minimize storage cost. Precise reconstruction of the compressed signal is also crucial as these data are further used for spike detection and/or classification. The conventional compressive sensing (CS) techniques to reconstruct the compressed BCI signals are computationally expensive. There are several existing techniques for CS reconstruction, including block-sparse Bayesian learning and block-based CS, which also work to replace a reconstruction methodology of CS in medical imaging with deep learning (DL) techniques. DL can be helpful in reconstructing compressed BCI signals, including Electroencephalography (EEG) and Electrocorticography (ECoG). Pertinent to that, in this work, a convolutional neural network (CNN) based reconstruction framework has been proposed to reconstruct spike signals that has been highly compressed using the CS technique. An accuracy of 91.62% has been achieved over signals compressed at 90% compression rate, when compared with original signals using a cross-correlation technique.