Arithmetic-Free Personalized Compressed Sensing Based On Deep Neural Networks for Wireless Transmission From Brain–Computer Interfaces

State-of-the-art brain–computer interfaces can carry out neural recording from hundreds of channels with high resolution. Such massive data makes it easy to study the brain better than ever before, but on the flip side, it leads to increased chip size, power consumption, heat dissipation and risk for patient safety. As such, compression of the data prior to transmission from the implant could be key to improving reliability and usability of such microsystems. In recent years, starting from sparsifying transforms all the way to compressive autoencoders (AEs), this compression has been offered by substantial arithmetic on the implant side, which in turn incurs its own inevitable costs. In this work, we analyze spike waveforms to prioritize subintervals by their importance. Thereupon, we design a temporal undersampling pattern matching the importance of each subinterval for compressive sensing of spikes. Following such sensing, we reconstruct spikes using a deep neural network (DNN) trained to capture spike representation from the undersampled measurements, with possible adaptation to individual subjects. This approach offers what we believe is the first spike compression-reconstruction framework that imposes no arithmetic on the compressing side, yet on the restoration side, performs at least on par with most on-chip arithmetic-heavy techniques. For instance, given a spike length of $N=64$ at eightfold compression, the famed symmlet-4 method yields a mean signal-to-noise-and-distortion ratio (SNDR) of 7.14 dB at a total compression arithmetic cost of $16N$ sums and products per spike, while for the proposed method, the figure is 8.38 dB at 0 sums and products.