Evaluation of spike sorting and compression for digitally reconfigurable non-uniform quantization

A non-uniform quantization approach for neural signals is presented and evaluated in this paper. Depending on the local properties of the signal, the number of bits is varied dynamically: a high number of bits is used for the detected Action Potentials (APs) or Spikes, otherwise a lower number of bits is used. Therefore a simple threshold detector enables the window for high precision mode. The algorithm is tested on simulated and recorded data sets. As opposed to standard thresholding methods, it is shown that keeping information of the background activity permits to achieve better signal reconstruction during spike sorting. At the same time a non-uniform quantization allows significant compression rates, decreasing the power consumption for transmission. A digital implementation is proposed, offering area reduction and more flexibility as opposed to analog thresholding implementations. The technique is also compared to delta compression and found to achieve better results on noisy recorded data.