Intracranial EEG: Methods & Applications

• Intracranial EEG (iEEG) is an invasive neurophysiological technique that records brain activity directly with millisecond precision and millimeter spatial specificity.
• It employs advanced signal processing methods like time–frequency analysis and deep learning to extract biomarkers for seizure detection and cognitive decoding.
• Clinical applications include precise mapping of epileptogenic zones, enhanced surgical planning, and the development of robust brain–computer interface systems.

Intracranial electroencephalography (iEEG) is an invasive neurophysiological technique that enables direct measurement of electrical activity from the brain parenchyma or cortical surface. By bypassing the signal distortion and spatial smearing characteristic of scalp EEG, iEEG achieves superior spatial and temporal resolution, making it the gold standard for mapping epileptogenic zones, guiding surgical intervention, decoding cognitive processes, and powering advanced brain–computer interface applications.

1. Principles of iEEG Acquisition and Signal Properties

iEEG records voltage fluctuations via electrodes implanted on the cortical surface (electrocorticography, ECoG) or inserted into the brain parenchyma (depth electrodes). Its distinguishing features include high temporal resolution (sub-millisecond), spatial specificity governed by electrode geometry and spacing (millimeter scale), and signal-to-noise ratios (SNR) orders of magnitude better than scalp EEG due to minimal contamination from extracranial tissues.

Electrode configuration is dictated by clinical hypothesis and neuroanatomy. Implants range from grid/strip arrays for subdural coverage of gyri to stereotactic depth electrodes sampling deep or combined regions. Sampling rates often exceed 1000 Hz to capture high-frequency oscillations and epileptic biomarkers.

Signal anatomy encompasses canonical frequency bands (delta through gamma), pathognomonic events such as interictal epileptiform discharges, high-frequency oscillations (HFOs), seizure patterns, and event-related spectral changes. iEEG channels are highly informative and non-redundant; data-driven analyses emphasize retaining full channel sets to preserve spatial information for robust clinical and decoding applications (Hussein et al., 2019).

2. Quantitative Analysis and Statistical Feature Extraction

iEEG analytics begins with extraction of time–frequency features and basic statistical descriptors. Band power in defined frequencies ($\delta$, $\theta$, $\alpha$, $\beta$, $\gamma$) is quantified using methods such as Welch’s periodogram or short-time Fourier transforms (STFT), producing matrices $X \in \mathbb{R}^{(B \times C) \times T}$ where $B$ is number of bands, $C$ is number of channels, and $T$ is number of epochs.

Dimensionality reduction is often required to manage channel count and avoid redundancy in multivariate modeling. However, principal component analysis (PCA) is generally ineffective for iEEG in high-stakes applications such as seizure prediction: for 16 channels, at least 14 principal components are required to preserve 95% of the variance, evidencing diffuse spatial information and justifying retention of all channels in predictive pipelines (Hussein et al., 2019).

Signals are often resampled to match the dominant spectral content (e.g., 100–200 Hz if most power lies below 50 Hz). Conversion of iEEG time series to 2D time–frequency spectrograms, e.g., by STFT or wavelet transform, is essential for leveraging spatial feature extractors such as convolutional neural networks (Hussein et al., 2019, Al-Radhi et al., 5 Aug 2025). Z-scoring and power ratios are used to obtain standardized, objective descriptors of per-channel dynamic changes, which have been validated to align closely with expert visual analysis (Panagiotopoulou et al., 24 Apr 2025).

3. Clinical and Research Applications

Epilepsy Diagnostics and Surgery Planning

iEEG is central to identifying epileptogenic zones (EZ) in drug-resistant focal epilepsy. Detection of key biomarkers such as interictal spikes or HFOs (80–500 Hz) is enabled by the technique’s bandwidth and SNR advantages (Sharifshazileh et al., 2020, Sadek et al., 22 Dec 2024). HFO detection informs surgical boundary definition, with real-time detection now feasible through neuromorphic SNN-based systems operating at $<1$ mW power with high specificity, supporting intraoperative decisions (Sharifshazileh et al., 2020).

Spatial quantification of interictal iEEG abnormalities, via z-score deviation from normative atlases, provides unbiased assessment of network focality and enhances prediction of surgical outcomes relative to clinical scoring systems alone (Gallagher et al., 2023). Simultaneous MEG-iEEG integration offers objective abnormality mapping for guiding electrode placement and resection planning, with logistic regression models incorporating both modalities (MEG and iEEG abnormality, plus their overlap metrics) achieving AUCs up to 0.80 for seizure-freedom prediction (Owen et al., 2023).

Seizure Prediction

Recent pipelines employ cooperative multi-scale CNNs for automatic feature extraction from iEEG segments reformatted as spectrograms. Networks with convolutional filters of different receptive fields (e.g., $1\times1$, $3\times3$, $5\times5$) and parallel max-pooling paths capture spatial and spectral features underlying preictal dynamics. State-of-the-art models achieve mean sensitivity of 87.85% and AUC 0.84 for seizure prediction, outperforming traditional handcrafted feature and PCA-based approaches (Hussein et al., 2019).

Band Power Fluctuations and Network Chronodes

Long-term iEEG recordings have revealed subject-specific modulations in band power at circadian, ultradian, and multidien timescales. Nonnegative matrix factorization (NMF) and multivariate empirical mode decomposition (MEMD) extract intrinsic mode functions (IMFs) corresponding to these cycles. Quantitative models demonstrate that instantaneous seizure dynamics—including duration and propagation—are significantly associated with the phase and amplitude of these iEEG cycles, with adjusted $R^2$ values exceeding 0.6 in many subjects (Panagiotopoulou et al., 2020, Panagiotopoulou et al., 24 Apr 2025). These findings suggest that subject-specific fluctuations in interictal spectral features serve as biomarkers of seizure-modulating processes and potential targets for chronotherapeutic interventions.

Functional Connectivity and Network Coupling

iEEG enables direct construction of functional and structural brain networks—by quantifying pairwise correlations and tractography-derived anatomical connections, respectively—at the scale of electrode placement. The structure–function coupling, defined as the Spearman correlation between connectivity matrices, robustly predicts surgical outcomes in focal epilepsy: higher pre-surgical coupling correlates with seizure freedom (e.g., coupling $\rho \sim 0.36$ seizure-free vs $\rho \sim 0.20$ not seizure-free) (Sinha et al., 2022). Virtual resection analyses at the node level allow complementary localization of epileptogenic zones by simulating changes in coupling upon electrode removal.

In neuromodulation studies, long-term functional connectivity variability (FC$_L$), computed as the cross-session standard deviation of site-specific correlations, emerges as a critical predictor of delta/theta band power changes following targeted stimulation. Inclusion of FC$_L$ in mixed-effects models improves out-of-sample prediction accuracy over spatial features alone (Papasavvas et al., 2021), suggesting its importance for personalized therapy design.

4. Advances in Signal Processing, Modeling, and BCI

iEEG’s high SNR is instrumental in emerging AI-driven neurotechnology. Deep learning models now leverage direct time–frequency (e.g., STFT, CWT) or time-domain features for automatic classification tasks, including seizure detection, HFO subtyping, and cognitive state decoding (Sadek et al., 22 Dec 2024). Hybrid frameworks combining pre-trained CNNs (e.g., GoogLeNet) with SVM classifiers achieve accuracies above 94% (GoogLeNet-SVM on real iEEG HFOs) and sensitivities above 94% (Sadek et al., 22 Dec 2024).

Compressed sensing and deep quantized autoencoder frameworks (e.g., BrainCodec) exploit iEEG’s high SNR to learn latent spaces transferable to noisier modalities such as scalp EEG. Models trained on iEEG maintain exceptional compression ratios (up to $64\times$) and reconstruction fidelity on both iEEG and EEG, with no loss in downstream seizure detection or motor imagery classification performance (Carzaniga et al., 10 Feb 2025).

Decoder generalization across patients with heterogeneous electrode configurations remains a core challenge. Solutions like RISE-iEEG deploy patient-specific learned projection layers that map individual electrode arrays to a common latent space preceding a shared CNN-based decoder, achieving F1-scores up to 0.83 and outperforming alternatives that require MNI-based electrode localization (Memar et al., 12 Aug 2024).

Single-subject and combined channel ensemble models integrating features from spatially distributed electrodes (e.g., gamma band descriptors) increase decoding robustness and reveal physiologically appropriate weighting (e.g., superior temporal lobe for auditory tasks, precentral gyrus for motor tasks) (Memar et al., 9 Dec 2024).

State-of-the-art BCI pipelines facilitate fully closed-loop interventions, autonomous operation, and accurate speech synthesis. Deep learning-based frameworks can reconstruct natural speech or decode continuous language content from high-gamma iEEG through two-phase transfer learning, with strong semantic and surface-level accuracy even from $<$30 min of training data (Shams et al., 31 May 2025, Al-Radhi et al., 5 Aug 2025).

5. Modeling, Forward Problem, and Source Localization

For precise source localization, the iEEG forward problem requires accurate representation of post-implant brain geometry and tissue conductivities. Patient-specific approaches include image warping driven by biomechanics models to deform preoperative MRI/DTI according to electrode placement, followed by finite element method (FEM) solutions on the deformed, segmented geometry. This image-as-a-model paradigm, which preserves DTI-derived anisotropic conductivity in white matter, significantly improves the fidelity of predicted potentials and lead field matrices, and mitigates source localization errors that can arise from unmodeled brain shift (Zwick et al., 2021, Safdar et al., 2022).

Automated frameworks integrating these steps within standardized visualization tools (e.g., 3D Slicer) permit rapid construction and validation (<10 min per patient) of bioelectrical field models, supporting integration with clinical workflows and surgical planning (Safdar et al., 2022).

6. Standardization, Benchmarks, and AI for iEEG

Standardized, open-source frameworks such as Neuroprobe provide benchmarks for decoding models applied to iEEG in naturalistic settings, including multimodal movie watching with aligned linguistic and sensory annotations. Key contributions include a taxonomy of decoding tasks (visual, auditory, language), cross-validation strategies (within-session, cross-session, cross-subject), and AUROC-based evaluation, demonstrating the comparative power of linear and deep neural models (Zahorodnii et al., 25 Sep 2025). Baseline performances indicate that linear models using Laplacian-referenced spectrograms are nontrivially competitive with contemporary neural foundation models, setting benchmarks for methodology comparison and neuroscientific inference.

Recent reviews highlight the proliferation of deep learning paradigms (CNN, RNN, GNN, transformers, SNN) for iEEG-based diagnostics across broad neurological conditions, and advocate for standardized benchmarks (e.g., BrainBenchmark), self-supervised learning, and multi-task pre-training to promote robustness and generalizability in clinical analysis pipelines (Li et al., 24 Feb 2025).

7. Future Directions and Open Challenges

Future work in iEEG research and applications will center on the following directions:

• Integration of multi-modal data streams (e.g., MRI, DTI, MEG, iEEG) for holistic mapping of structural–functional relationships and dynamic network coupling.
• Development of patient-adaptive, closed-loop neuromodulation platforms informed by longitudinal iEEG network dynamics and chronodes.
• Decoding unconstrained, natural language or motor intent from sparse electrode coverage using transfer learning and data-efficient architectures.
• Bridging iEEG-derived insights to scalpel EEG and noninvasive settings via high-fidelity neural compression and transfer learning (Carzaniga et al., 10 Feb 2025).
• Continued advancement in automated modeling, spatial feature integration, and interpretable neural decoders to ensure clinical reliability and neuroscientific insight.

In sum, iEEG remains a uniquely informative and versatile modality for probing human brain function, driving progress in epilepsy care, cognitive neuroscience, and next-generation brain–machine interfaces, underpinned by ongoing innovations in data analysis, modeling, and artificial intelligence.