Stereotactic EEG Applications

• Stereotactic EEG Applications is a technique using 3D-precise implantation of intracranial electrodes to record both cortical and subcortical activity.
• It employs high-resolution imaging protocols and advanced electrode localization workflows to ensure accurate mapping for presurgical epilepsy evaluation and neural decoding.
• The method supports computational modeling and biophysical head simulations, enhancing source localization, surgical planning, and translational neuroscience research.

Stereotactic electroencephalography (SEEG) refers to the 3D-precise implantation of intracranial electrodes via stereotactic guidance to record local field potentials (LFPs) from both cortical and subcortical regions of the human brain. SEEG is foundational to the presurgical evaluation of refractory epilepsy, serves as a ground truth for non-invasive source localization, and supports a wide array of computational, clinical, and translational neuroscience applications. Its utilization encompasses the domains of epileptogenic zone localization, patient-specific modeling of brain conductivity, neural decoding, and the development of advanced visualization and source imaging pipelines.

1. Neuroimaging and Electrode Localization Workflows

SEEG applications are underpinned by high-precision image-based protocols for surgical targeting and electrophysiological analysis. Pre-implant imaging typically entails high-resolution T1-weighted MRI (1 mm³ isotropic), followed by full cortical reconstruction (e.g., FreeSurfer v6.0+) to demarcate the pial and white matter surfaces and provide sub-millimeter-precision anatomical meshes (Adamek et al., 2023). After lead implantation, a volume CT is acquired and rigidly co-registered to the pre-operative MRI via mutual information optimization, ensuring electrodes are mapped in patient-specific MRI space. Automated tools such as the Versatile Electrode Localization Framework (VERA) segment the postoperative CT, extract electrode centroids, and associate contact points with their anatomical counterparts.

A critical step is the selection of “cortical” contacts: electrode centroids $e_j$ are retained if within 4 mm of any surface mesh vertex $p_{ci}$ (i.e., $d_j = \min_i\|e_j - p_{ci}\| < 4\ \mathrm{mm}$), thereby enforcing anatomical correspondence to cortex and strengthening the accuracy for subsequent functional overlay and network modeling (Adamek et al., 2023).

2. Advanced Visualization and Surface Morphing

A longstanding bottleneck in SEEG data interpretation is the obscuration of sulcal contacts by the folded cerebral cortex. This is addressed via cortical inflation—a mesh-based, energy-minimizing morph (implemented in FreeSurfer), parameterized by a blend between “folded” ($p_{ci}$) and “inflated” ($p_{ii}$) geometry. The energy functional minimized is

$$E_\text{total}(q) = \lambda E_\text{smooth}(q) + (1-\lambda) E_\text{convex}(q)$$

where $E_\text{smooth}$ enforces local Euclidean proximity, and $E_\text{convex}$ penalizes mean curvature. A linear blend, $p_{mi}(\sigma) = (1-\sigma)p_{ci} + \sigma p_{ii}$, allows investigators to animate between fully folded and inflated presentations, preserving topological adjacency and geodesic distances (distortions $<$5%) (Adamek et al., 2023). Integration into MATLAB and Python (MNE-Python) toolchains supports interactive, real-time morphing, assisting with the localization of both sulcal and gyral contacts during clinical conferences or data exploration. This representation facilitates direct overlay with functional atlases and circuit-level analyses.

3. Functional and Computational SEEG Applications

SEEG enables direct recording from epileptogenic networks and supports multidimensional computational analyses:

A. Epileptogenic Zone (SOZ) Localization

Interictal biomarker extraction—for high-frequency oscillations (HFOs), interictal epileptiform discharges (IEDs), and phase-amplitude coupling (PAC)—is performed on 2 h nocturnal epochs parsed into 3 s windows. Biomarker rates per channel form feature vectors, enabling SVM-RBF classification (AUC up to 0.79, sensitivity 70.4%, specificity 75.1%) for SOZ vs. non-SOZ channels (Varatharajah et al., 2018). Notably, composite multi-biomarker approaches outperform single-marker models. Recording durations of 90–120 min suffice for near-maximal discriminability, and millisecond-scale per-channel feature computation feeds real-time, intraoperative pipelines for rapid surgical planning.

B. Multi-Subject Neural Decoding

Integration of SEEG data across variable electrode configurations requires architectures that tokenize neural activity per electrode (via 1D convolutions) and encode 3D electrode position using Gaussian RBF-based spatial embeddings. Alternating transformer-based self-attention blocks in the time and electrode dimensions capture spatiotemporal dependencies despite heterogeneous implant patterns. The "seegnificant" approach yields unified, multi-subject decoders with $R^2$ values up to 0.54 (overall), provides substantial performance improvements over single-subject models (Δ$R^2\sim$0.09), and supports few-shot behavioral decoding transfer to new subjects (Mentzelopoulos et al., 2024). These advances pave the way for foundational SEEG BCI development and clinical cognitive monitoring.

4. Conductivity Estimation and Biophysical Head Modeling

SEEG, in conjunction with simultaneous scalp EEG and controlled intracerebral electrical stimulation (IES), can probe compartment-specific head tissue conductivities. A five-compartment finite-element head model (scalp, skull, CSF, GM, WM) is constructed from co-registered MRI and CT. Contact potentials during IES are modeled via the quasi-static PDE $\nabla\cdot(\sigma\nabla V)=0$ with Dirichlet and Neumann boundary conditions.

Conductivity estimation is formulated as an inverse problem minimizing a relative difference measure (RDM) between simulated ($x_s$) and measured ($x_d$) potentials:

$$J(\sigma) = \mathrm{RDM}(x_s(\sigma), x_d) = \sqrt{\sum_{k=1}^M \left(\frac{x_{d,k}}{\|x_d\|_2} - \frac{x_{s,k}(\sigma)}{\|x_s(\sigma)\|_2}\right)^2}$$

Optimization employs Nelder–Mead simplex (multi-start), yielding RDM$\sim$0.17–0.75 in-vivo (Altakroury et al., 2022). SEEG provides highest sensitivity for deep compartments (CSF, GM, WM), while scalp EEG constrains superficial layers (scalp, skull). Optimal precision is achieved with measurement electrodes <5 cm from the stimulation and broad spatial sampling. This methodology yields patient-specific conductivities, improving both source localization and non-invasive neuromodulatory targeting.

5. Source Imaging and Complete Electrode Modeling

Finite element (FE) modeling with the complete electrode model (CEM) enables the simultaneous simulation of scalp and depth electrodes, accounting for contact impedance, geometry, and conductivity discontinuities (Prieto et al., 31 Jan 2026). The BST-to-ZI plugin, leveraging Brainstorm and Zeffiro Interface, supports multi-compartment tetrahedral meshes (e.g., WM=0.14, GM=0.33, CSF=1.79, skull=0.0064 S/m), with element sizes refined to 0.3–0.4 mm near invasive electrodes.

Forward modeling solves

$$\nabla\cdot(\sigma(\mathbf x) \nabla u(\mathbf x)) = S(\mathbf x)$$

with CEM constraints:

• Zero-flux (Neumann) on scalp
• Current conservation per electrode
• Interface potential jumps via electrode impedance ($z_\ell$)

Concatenation of scalp and probe lead fields ($L_\mathrm{full} = \begin{pmatrix}L_\mathrm{scalp}\L_\mathrm{probe}\end{pmatrix}$) sharpens sensitivity near depth contacts. Simulations with SEEG probes (e.g., DBS-like arrays) yield localization errors $<$3–5 mm and marked gains in spatial focality, particularly for sources parallel to the probe. This pipeline is suited for deep source characterization in hybrid SEEG-EEG experiments and supports clinical translation, assuming careful model parameterization and high-computational resources (Prieto et al., 31 Jan 2026).

6. Non-Invasive Planning and Integration with SEEG

The spatial limitations of SEEG coverage necessitate hypothesis-driven implantation strategies. ICA-based source separation of ictal scalp EEG, followed by equivalent current dipole localization using a boundary-element head model, offers a non-invasive means to guide SEEG electrode targeting (Barborica et al., 2021). Validation with simultaneous scalp-SEEG demonstrates localization errors as low as 10 mm for superficial neocortical sources, with accuracy declining for deep or mesial generators (mean dipole-to-SOZ centroid distance 47 mm). High scalp-ICA coherence suggests neocortical focus, while its absence points to deeper or mesial sources. This integrative approach refines stereotactic implantation hypotheses and may inform sub-lobar targeting, especially when combined with MRI, PET, and semiological data.

7. Limitations and Future Directions

Key challenges include computational demands of multi-million element FE meshes, uncertainty in real electrode impedance, and inter-individual anatomical variability. Further, spatial positional encoding in neural decoding yields limited gains, suggesting a role for more sophisticated atlas priors or functional-connectivity-based approaches (Mentzelopoulos et al., 2024). Future developments may include self-supervised SEEG representation learning on large-scale datasets, multi-task models for simultaneous behavioral and clinical decoding, and real-time, adaptive visualization integrated within surgical workflows. Combining SEEG with non-invasive EEG for hybrid modeling and employing realistic tissue conductivities remains critical for both source localization accuracy and neuromodulatory applications.

In summary, SEEG provides unparalleled access to deep and superficial human brain electrophysiology. Its applications span clinical, computational, and biophysical domains, integrating advanced imaging, deep learning, inverse modeling, and real-time visualization to support diagnostic, prognostic, and neuroscientific discovery (Adamek et al., 2023, Mentzelopoulos et al., 2024, Altakroury et al., 2022, Varatharajah et al., 2018, Barborica et al., 2021, Prieto et al., 31 Jan 2026).