Study of Compressed Sensing and Predictor Techniques for the Compression of Neural Signals under the Influence of Noise

In this paper an analysis of compression schemes based on compressed sensing (CS) and predictor techniques for neural signals is presented. The focus is on how much a compression algorithm can reduce data while not affecting the subsequent signal processing. Since neural signals are processed by means of spike sorting algorithms the evaluation is not trivial and not well defined, since there exists in fact many different ways to detect and cluster the spikes. Evaluating how much a compression scheme affects the result of spike sorting programs is a crucial step before implementing such compression technique. In the analysis two use cases are evaluated: in the first, spikes are detected and extracted and only thereafter compressed. In the second case, no information on the spikes is available and the whole raw signal is compressed. When dealing only with spike frames CS offers great compression at almost no loss, in the case of the whole recording its performances are greatly impaired and delta compression outperforms it in terms of data reduction and spike sorting results. In this case the reduction rates are modest but significant, ≈3 – 4 times data reduction and the whole signal is preserved avoiding big permanent losses of information.