Spike sorting is a fundamental signal processing step in intracortical neuroscience and BCI systems. When an electrode sits among hundreds of neurons, it picks up overlapping electrical signals from multiple cells. Spike sorting algorithms disentangle this mixture, attributing each detected action potential (spike) to a specific neuron based on its characteristic waveform shape, amplitude, and timing. This transforms raw voltage traces into the activity patterns of identified individual neurons.
Traditional spike sorting methods relied on manual waveform clustering by expert researchers — a tedious and subjective process. Modern approaches use automated algorithms such as Kilosort, MountainSort, and SpykingCircus, which employ template matching, principal component analysis, and machine learning to sort spikes from thousands of channels in near real time. As electrode arrays scale from 96 channels (Utah Array) to over 1,000 (Neuralink's N1) and beyond, scalable automated spike sorting becomes essential.
For BCI applications, spike sorting quality directly impacts decoding performance. Clean isolation of individual neurons provides the richest possible input to decoding algorithms. However, some modern BCI decoders have shown strong performance using threshold crossings or multi-unit activity without explicit spike sorting, suggesting that perfect neuron isolation may not always be necessary. The trade-off between sorting accuracy and computational latency remains an active area of optimization for real-time BCI systems. For deeper coverage, see BCIIntel.