Brain-Computer Interfaces

Derivation 1: Signal-to-Noise and Information Capacity

The information capacity of a neural recording channel is bounded by Shannon's formula:

$$C = B \log_2\left(1 + \text{SNR}\right) \text{ bits/s}$$

where $B$ is bandwidth and SNR is the signal-to-noise ratio. For an intracortical electrode recording $N$ neurons with mean firing rate $r$ and Poisson statistics:

$$\text{SNR}_{\text{pop}} \approx N \cdot r \cdot T \cdot d'^2$$

where $T$ is the integration time and $d'$ is the single-neuron discriminability. The electrode impedance at frequency $f$ determines thermal noise:

$$V_{\text{noise}}^2 = 4 k_B T \cdot \text{Re}[Z(f)] \cdot \Delta f$$

For a typical Utah array electrode ($Z \approx 1$ MOhm at 1 kHz), the thermal noise is ~5 $\mu$V RMS, while spike amplitudes are 50–500 $\mu$V, yielding SNR of 10–100. Neuropixels probes achieve lower impedance (~150 kOhm) with smaller electrode sites, recording from thousands of neurons simultaneously.

Derivation 2: Kalman Filter Decoder

The state (position, velocity) $\mathbf{x}_t$ evolves according to:

$$\mathbf{x}_t = \mathbf{A}\mathbf{x}_{t-1} + \mathbf{w}_t, \quad \mathbf{w}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{W})$$

Neural observations (spike counts) are linearly related to state:

$$\mathbf{z}_t = \mathbf{H}\mathbf{x}_t + \mathbf{v}_t, \quad \mathbf{v}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q})$$

The Kalman filter recursion gives the optimal state estimate:

$$\hat{\mathbf{x}}_t = \mathbf{A}\hat{\mathbf{x}}_{t-1} + \mathbf{K}_t(\mathbf{z}_t - \mathbf{H}\mathbf{A}\hat{\mathbf{x}}_{t-1})$$

where the Kalman gain is:

$$\mathbf{K}_t = \mathbf{P}_t^- \mathbf{H}^T (\mathbf{H}\mathbf{P}_t^- \mathbf{H}^T + \mathbf{Q})^{-1}$$

The Kalman gain balances trust between the dynamical prediction and neural observations. When neural noise is high ($\mathbf{Q}$ large), the decoder relies more on the movement model; when the model is uncertain ($\mathbf{W}$ large), it relies more on neural data. The ReFIT Kalman filter (Gilja et al., 2012) improves performance by re-fitting the observation model using intended rather than observed cursor movements.

Derivation 3: Population Vector Decoder

The population vector (PV) decoder estimates the movement direction from cosine-tuned neurons. For neuron $i$ with preferred direction $\hat{\mathbf{d}}_i$ and firing rate $r_i$:

$$\hat{\mathbf{v}} = \sum_{i=1}^{N} (r_i - b_i) \hat{\mathbf{d}}_i$$

The decoding error variance for $N$ neurons with Poisson noise is:

$$\text{Var}(\hat{\theta}) \approx \frac{2}{N \cdot k^2 T}$$

where $k$ is the modulation depth and $T$ is the integration window. The PV decoder is suboptimal compared to maximum likelihood, but its simplicity makes it robust and interpretable. It was used in the first human BCI trial (BrainGate, 2006).

Derivation 4: Charge-Balanced Stimulation and Safety Limits

Electrical stimulation must be charge-balanced to prevent electrode corrosion and tissue damage. For a biphasic pulse with amplitude $I$ and phase duration $t_p$:

$$Q = I \cdot t_p \quad (\text{charge per phase, in nC})$$

The Shannon safety limit relates charge density and charge per phase:

$$\log\left(\frac{Q}{A}\right) + \log(Q) < k$$

where $A$ is the electrode area and $k \approx 1.85$ (Shannon, 1992). The volume of neural activation is approximately:

$$r_{\text{activation}} \approx \sqrt{\frac{I}{k_r}} \quad (\text{current-distance relation})$$

where $k_r \approx 1292$ $\mu$A/mm$^2$ for cortical neurons. At typical BCI currents (10–100 $\mu$A), activation radii are 100–300 $\mu$m, engaging hundreds of neurons per electrode. Multi-electrode patterns can create more complex percepts through spatial summation.

Derivation 5: Closed-Loop BCI Performance Metrics

BCI performance is quantified by Fitts' law throughput. For a center-out reaching task with target distance $D$ and target width $W$:

$$\text{ID} = \log_2\left(\frac{D}{W} + 1\right) \quad \text{(index of difficulty, bits)}$$

$$\text{Throughput} = \frac{\text{ID}}{MT} \quad \text{(bits/s)}$$

where MT is the movement time. The BCI throughput depends on the number of decoded dimensions $d$ and decoder accuracy:

$$\text{TP} \propto d \cdot \log_2\left(1 + \frac{\text{SNR}_{\text{decode}}}{d}\right)$$

Current intracortical BCIs achieve ~2–4 bits/s for 2D cursor control (comparable to a smartphone touchscreen at ~10 bits/s). Recent advances using recurrent neural network decoders have pushed this to ~5–7 bits/s for handwriting and speech decoding.