Signal-to-noise ratio (SNR) is the fundamental metric of neural recording quality. In BCI systems, high SNR means individual neuron spikes are clearly distinguishable from background neural noise, enabling reliable spike detection and sorting. Low SNR means spikes are buried in noise — reducing the number of usable channels and degrading decoder performance.

Sources of Noise in Neural Recording

Neural recording noise has multiple sources:

  1. Thermal (Johnson-Nyquist) noise: Random voltage fluctuations from electrode impedance; minimized by low-impedance electrode materials (Pt, IrOx) and careful electrode area sizing
  2. Biological noise: Background activity from thousands of neurons not of interest; manifests as "hash" or MUA (multi-unit activity) that sets the noise floor for single-unit isolation
  3. Electronic noise: Amplifier input-referred noise; minimized by low-noise ASIC design (Neuralink's custom chip achieves ~5 µV RMS input noise)
  4. Electromagnetic interference (EMI): From power lines (60 Hz in US, 50 Hz in EU), nearby electronics, etc.; managed via differential recording and physical shielding
  5. Movement artifacts: Mechanical motion of electrodes relative to tissue; particularly problematic in ambulatory subjects

Measuring SNR

Neural recording SNR is typically defined as:

SNR (dB) = 20 × log10(peak-to-peak spike amplitude / 2 × RMS noise)

A SNR of:

  • >6: Excellent — single unit is cleanly isolatable
  • 3-6: Good — workable for most sorting algorithms
  • 1-3: Marginal — multiunit only
  • <1: Electrode is unusable

SNR Degradation Over Time

One of the central challenges of chronic intracortical BCI is SNR degradation over months to years:

  • Glial scar tissue encapsulates the electrode, increasing electrode impedance and increasing the physical distance between the recording tip and target neurons
  • Mean SNR decreases monotonically over 12-24 months in most silicon array implants
  • Neuralink's flexible polymer threads are hypothesized to reduce glial scar formation (and thus preserve SNR) by reducing mechanical micromotion — this is a key scientific question in the PRIME study

SNR Trade-offs in BCI Design

Higher electrode count (more channels) can compensate for lower per-channel SNR — if SNR per channel is marginal but you have 1,024 channels instead of 96, you may still decode more neurons in total. This is part of the rationale for Neuralink's move to 1,024 electrodes despite each flexible thread having lower SNR than a rigid Utah Array shank.