Optimizing Neural Recording Front-Ends Toward Enhanced Spike Sorting Accuracy in High-Channel-Count Systems

Spike sorting is a pivotal signal-processing technique used to extract information from raw extracellular recordings. Its performance is influenced by the characteristics of the neural recording front-end. This study explores how design choices in amplifiers, filters, and analog-to-digital converters (ADCs) affect the accuracy of well-established spike sorting algorithms. Our primary objective is to identify the minimal requirements that ensure high sorting accuracy while facilitating power- and area-efficient analog front-ends, which is especially needed for multi-channel recording-only applications. To achieve this, we use both synthetic and real datasets, serving as ground truth, processed through a generic MATLAB model of a neural recording front-end that simulates key electrical parameters impacting the signal integrity. These include the filter order and cutoff frequency, ADC resolution, ADC sampling frequency, and nonlinearity. Our findings indicate that optimal spike-sorting results are obtained with a 1st-order bandpass Butterworth filter ranging from 700 Hz to 7.5 kHz, coupled with an ADC that offers a 15-kHz sampling frequency at 8-bit resolution and no missing codes. These insights are crucial for designing high-channel-count neural interfaces where CMOS circuits must efficiently be optimized.