Evaluation study of compressed sensing for neural spike recordings

In this paper, an evaluation study of compressed sensing (CS) for neural action potential (spike) signals in MATLAB is presented. State-of-the-art neural recorders use 100 or more parallel channels to measure neural activity resulting in a data rate of 16 - 20 Mbit/s. Since a low-power design is required for an implanted neural recorder, it seems advantageous to compress the neural data prior to the wireless transmission. The continuous neural spike signals are compressed and transmitted to facilitate the possibility of an unrestricted data analysis at the receiver. Synthesized and recorded neural data sets are used to test the performance of CS. The 6-level Daubechies-8 wavelet decomposition matrix and two learned dictionary matrices are utilized as dictionaries for CS. The compression results are evaluated with the spike sorting program OSort. CS is shown to work for the compression of low-noise synthesized neural spike signals with a compression rate of 2.05, but cannot be recommended for the compression of neural spike signals in general.