Neural Signal Biophysics and BCI Signal Taxonomy

这种跨尺度的信号

Volume conductor theory is the cornerstone of BCI signal modeling, as it describes how neural current sources generate measurable potentials at recording sites separated by tissue layers with distinct electrical conductivities. The simplest and most widely used model is the homogeneous spherical volume conductor, which assumes the head is a uniform isotropic sphere with constant conductivity $\sigma$. For a point current source $I_0$ located at position $\mathbf{r}_0$ inside the sphere of radius $R$, the extracellular potential $V(\mathbf{r})$ at a recording site $\mathbf{r}$ on the surface is given by the following closed-form solution: $$V(\mathbf{r}) = \frac{I_0}{4\pi\sigma} \cdot \frac{R^2 - r^2}{|\mathbf{r} - \mathbf{r}_0|^3} \cdot (\mathbf{r} - \mathbf{r}_0) \cdot \hat{\mathbf{r}}$$ where $\hat{\mathbf{r}}$ is the unit vector pointing from the center of the sphere to the recording site. This equation reveals two critical biophysical properties of neural signal propagation: first, potential amplitude decays with the square of the distance between source and recording site (for sources near the center of the sphere, $|\mathbf{r} - \mathbf{r}_0| \approx R$, so $V \propto 1/R^2$); second, the spatial distribution of the potential is dipolar for intracellular current sources (e.g., action potentials) and monopolar for extracellular current injections (e.g., stimulation currents).

For real-world BCI design, the homogeneous sphere model is insufficient due to the layered structure of the head, with each layer having distinct conductivity: scalp ($\sigma_{scalp} = 0.33 \, \text{S/m}$), skull ($\sigma_{skull} = 0.01 \, \text{S/m}$, 30× lower than soft tissue), cerebrospinal fluid (CSF, $\sigma_{csf} = 1.79 \, \text{S/m}$, 5× higher than soft tissue), gray matter ($\sigma_{gm} = 0.33 \, \text{S/m}$), and white matter ($\sigma_{wm} = 0.14 \, \text{S/m}$, anisotropic). The forward model for this layered structure is typically solved via boundary element method (BEM) or finite element

For real-world BCI design, the homogeneous sphere model is insufficient due to the layered structure of the head, with each layer having distinct conductivity: scalp ($\sigma_{scalp} = 0.33 \, \text{S/m}$), skull ($\sigma_{skull} = 0.01 \, \text{S/m}$, 30× lower than soft tissue) [^1], cerebrospinal fluid (CSF, $\sigma_{csf} = 1.79 \, \text{S/m}$, 5× higher than soft tissue), gray matter ($\sigma_{gm} = 0.33 \, \text{S/m}$), and white matter ($\sigma_{wm} = 0.14 \, \text{S/m}$, anisotropic). The forward model for this layered structure is typically solved via boundary element method (BEM) or finite element method (FEM), each with distinct tradeoffs for BCI system design. BEM discretizes only the interfaces between tissue layers, offering computational efficiency (a 19-channel 10-20 system EEG forward model can be solved in <10 s on a standard laptop) but limited accuracy for anisotropic tissues like white matter, with errors up to 15% when skull thickness varies by >10% across the scalp. FEM discretizes the entire head volume, enabling precise modeling of anisotropic conductivity and complex geometries (e.g., CSF cavities, skull foramina) but requires 10-100× longer computation time, making it more suitable for pre-surgical planning than real-time BCI operation. Validation studies using MRI-derived head models show FEM can reduce forward model error to <5% for most scalp locations, making it the gold standard for high-accuracy source localization in clinical BCI applications.

A critical biophysical property of the layered volume conductor is frequency-dependent signal attenuation, driven by the capacitive nature of biological tissues. The complex impedance of skull tissue at 10 Hz is ~300 Ω·m, dropping to ~80 Ω·m at 100 Hz, while the capacitive reactance increases with frequency, creating a low-pass filter effect for signals propagating through the skull. This explains why scalp EEG has negligible power above 50 Hz under normal conditions: a 100 Hz γ-band oscillation with 100 μV amplitude at the cortical surface is attenuated to <1 μV at the scalp, below the noise floor of standard EEG systems. In contrast, ECoG electrodes placed directly on the cortical surface avoid skull attenuation entirely, enabling measurement of high-frequency oscillations (HFOs, 100-500 Hz) with amplitudes of 50-200 μV, which are strongly correlated with motor execution and speech production.

> 📌 Design Tip: When selecting recording sites for frequency-specific BCI paradigms, prioritize subdural ECoG for γ-band (>30 Hz) applications (e.g., high-dexterity motor control) to avoid skull attenuation, while scalp EEG remains sufficient for low-frequency paradigms (e.g., P300, motor imagery μ/β desynchronization).

The second foundational component of BCI signal taxonomy is rigorous characterization of noise sources, which dominate signal quality for all non-invasive and semi-invasive modalities. Noise in neural recordings is classified into three broad categories, each with distinct spectral and spatial properties: 1. Biological noise: Generated by physiological processes unrelated