On-chip spike detection

On-chip spike detection identifies action-potential events on the implant itself, so that only the events — timestamps and short snippets — need to be transmitted instead of the full-band waveform of every channel. This is the most direct way to break the data-rate bottleneck: a broadband channel produces far more bits than its firing rate warrants, and most of that stream is inter-spike baseline.

The simplest and most common method is amplitude thresholding, usually against an adaptive threshold set as a multiple of a running noise estimate; a robust noise scale is the median absolute deviation divided by about 0.6745. Detection performance improves by first passing the signal through a nonlinear energy operator (the Teager energy operator) or by using the absolute value, which emphasise the brief high-energy spike over slow drift. More elaborate designs add on-chip feature extraction, template matching, or compressive detection.

The tradeoff is information loss and drift sensitivity. A fixed or slowly-adapted threshold produces false positives from artifacts and false negatives when a unit's amplitude drops, and once only threshold crossings are kept, the raw waveform is gone, so offline re-sorting is impossible. Because electrode impedance and unit amplitudes drift over a chronic implant's life, the detector's threshold must track them or slowly lose units.