Intracranial EEG (iEEG): Techniques & Applications

• Intracranial EEG (iEEG) is a method using electrodes placed directly on or within the brain to record electrical activity with millimeter precision and sub-millisecond resolution.
• Advanced signal processing and machine learning techniques, including CNNs and transfer models, improve the analysis of iEEG data for seizure detection and cognitive studies.
• iEEG is pivotal for clinical applications such as epilepsy surgery planning, brain-computer interfaces, and biomarker discovery by offering superior signal-to-noise ratio and detailed neural insights.

Intracranial electroencephalography (iEEG) is the measurement of electrical activity of the brain recorded from electrodes placed directly on the cortical surface or within the brain parenchyma. iEEG encompasses both electrocorticography (ECoG; subdural or epicortical arrays) and stereoelectroencephalography (sEEG; penetrating depth electrodes) and provides millimeter-scale spatial and sub-millisecond temporal resolution unobtainable with noninvasive methods. Its superior signal-to-noise ratio (SNR), substantial bandwidth (routinely 0.1–2,000 Hz), and insulation from scalp/muscle artifacts have established iEEG as the gold-standard technique for localizing epileptogenic zones in drug-resistant epilepsy, performing high-precision functional mapping, and conducting basic neuroscientific studies of human cognition (Carzaniga et al., 10 Feb 2025, Burrello et al., 2018). Recent expansions of large, annotated iEEG datasets, high-fidelity compression tools, and cross-modal machine learning have vastly enlarged its translational and computational landscape.

1. Principles of iEEG Acquisition and Signal Properties

iEEG recording utilizes clinical subdural grids/strips or depth electrodes, with patient-specific channel counts (typically 20–250) and a range of sampling rates (500–2,000 Hz) to interrogate diverse cortical and subcortical regions (Duan et al., 17 Feb 2026). Electrode localization, referencing schema (common average, bipolar), amplifier chain, and artifact rejection (automatic and expert-visual inspection) underlie data quality. iEEG displays characteristic broadband power-law 1/f spectra, physiologic oscillations (delta–high-gamma: 1–250 Hz), and pathological events such as epileptiform spikes and high-frequency oscillations (HFOs, 80–500 Hz) (Carzaniga et al., 10 Feb 2025, Sadek et al., 2024). SNR routinely exceeds 20 dB, an order of magnitude greater than scalp EEG (Carzaniga et al., 10 Feb 2025).

iEEG signal heterogeneity reflects implantation strategy, clinical protocol, and biological context. Multi-day continuous recordings, often with complex nonstationarity (weeks-to-months meta-state switching), enable the study of slow disease dynamics and necessitate adaptive algorithmic strategies (Yang et al., 2022). The recording paradigm is integral to downstream analytics (epilepsy surgery, BCI, cognitive studies).

2. Computational Methods and Analytical Frameworks

2.1 Preprocessing and Feature Extraction

Standard iEEG workflows begin with band-pass and notch filtering, artifact rejection (e.g., ICA, threshold-based exclusion), referencing, and (where appropriate) downsampling (Li et al., 24 Feb 2025). Feature extraction encompasses time-domain metrics (mean, variance, entropy, nonlinear energy operator), spectral (band power, coastline, spectral edge), and time–frequency representations (wavelet scalograms, Hilbert-envelope for high-gamma) (Memar et al., 2024, 2207.13190). For event-related work, epochs are segmented to synchronize with task cues or clinical events.

2.2 Machine Learning and Deep Learning

iEEG analysis leverages an array of “shallow” and deep learning models:

• Shallow models: SVMs (including multiclass ECOC), random forests, logistic regression, Gaussian Naive Bayes; salient in small-N or low-dimensional feature contexts (Memar et al., 2024, Sadek et al., 2024).
• Deep neural networks: CNNs (1D/2D, e.g., ResNet, GoogLeNet, VGG), RNNs (LSTM/GRU), Transformers, graph neural networks (Li et al., 24 Feb 2025). High-resolution 2D scalograms (wavelet domain) as CNN inputs yield high performance for HFO and seizure detection (Sadek et al., 2024).
• Hybrid and transfer models: Hybrid CNN–SVM, cross-modal transfer learning (e.g., iEEG-trained compressors for downstream EEG applications), patient-specific adaptation layers (RISE-iEEG) to address variable electrode geometry (Memar et al., 2024, Carzaniga et al., 10 Feb 2025).
• State-space and dynamical models: Linear dynamical systems (Kalman filtering); nonlinear extensions (DFINE) combine deep autoencoders with latent LSSM backbones, offering superior forecasting and missing-data robustness (Vaziri et al., 28 Dec 2025).

2.3 End-to-End Pipelines

Processing and analytics pipelines are modular, spanning:

• Preprocessing → Feature extraction (temporal, spectral, spatio-temporal) → Dimensionality reduction (e.g., PCA, patient-specific projection) → Model fitting (supervised, unsupervised, transfer-learning) → Evaluation/visualization.
• Unified toolchains, e.g. within 3D Slicer (Safdar et al., 2022), support FEM-based modeling for biophysical forward/inverse problems.

3. Clinical and Translational Applications

3.1 Epilepsy Surgery Planning

iEEG is the gold standard for epileptogenic zone (EZ) localization in drug-resistant epilepsy (Owen et al., 2023, Duan et al., 17 Feb 2026). Surgical workflows integrate:

• Preoperative planning: Multimodal imaging (MRI, MEG) and quantitative band-power abnormality mapping guide electrode placement; MEG–iEEG spatial overlap and the extent of resection of “most abnormal” tissue predict surgical outcome (combined AUC 0.80) (Owen et al., 2023).
• Pathological zone annotation: iEEG enables direct recording of ictal/interictal events, high-frequency oscillations, and spike–HFOs (spkHFO), now annotated at >36,000 events in curated datasets (Duan et al., 17 Feb 2026).
• Automated and real-time analytics: On-device and neuromorphic architectures, such as event-based SNNs and sparse hyperdimensional computing accelerators, deliver sub-10 s latency seizure or HFO detection with sub-μW power budgets, suitable for implantable systems (Cuyckens et al., 10 Oct 2025, Sharifshazileh et al., 2020).

3.2 Brain-Computer Interfaces (BCI)

iEEG enables BCIs for communication, motor restoration, and adaptive closed-loop stimulation (2207.13190). Key points include:

• Decoding paradigms: Spatio-temporal feature extraction and neural network decoders achieve high binary classification (e.g., F1 = 0.81–0.84 with ensemble methods (Memar et al., 2024)) and robust movement or speech decoding (real-time click or word rates, WER ≈ 26%) (2207.13190).
• Spatial information integration: Greedy channel combination strategies and spatially aware projection layers (RISE-iEEG) enhance decoding accuracy and generalizability across heterogeneous electrode montages (Memar et al., 2024).
• Cross-modal virtual sensing: Multi-scale Transformers (CAST) and conditional normalizing flows (NeuroFlowNet) reconstruct virtual iEEG from scalp EEG, achieving high waveform fidelity in superficial and some deep structures (peak r = 0.864 in precentral gyrus; whole-cortex r = 0.545 with CAST) (Pham et al., 17 May 2026, He et al., 27 Feb 2026). Diffusion-based, geometric-constraint-infused frameworks further exploit anatomical priors to bridge non-invasive and invasive recordings (Dong et al., 3 Apr 2026).

3.3 Biomarker Discovery and Sleep/Cognitive State Tracking

Automated classification of sleep stages via complexity–entropy mapping (using Bandt–Pompe weighted permutation entropy and statistical complexity) in large-scale iEEG (N = 106) achieves >92% accuracy; whole-brain mean-field AdEx models reproduce empirical trajectories (Lucas et al., 12 Nov 2025). Information-theoretic, symbolic, and spectral features underpin iEEG-based biomarkers for epilepsy, cognition, psychiatric disorders, and BCI state-monitoring.

4. Physical and Biophysical Modeling

High-precision solution of the iEEG forward and inverse problems is essential for localization of sources and integration with multimodal imaging. Methodological advances include:

• Finite element modeling (FEM): Patient-specific inhomogeneous and anisotropic conductivity models, image-based mesh generation (“image-as-model”), and biomechanics-based brain warping are implemented within 3D Slicer or equivalent platforms (Safdar et al., 2022, Zwick et al., 2021). Accounting for electrode-induced brain deformation (up to 22 mm shift) reduces lead field and potential error by >20% compared to models ignoring post-implant shift.
• Automated workflow: MRI/CT-based segmentation, DTI-driven conductivity assignment, and electrode localization are streamlined into modular CLI/Slicer pipelines (Safdar et al., 2022).

5. Data, Benchmarking, and Generalizability

The field has transitioned from small, single-center collections to large, harmonized, multi-center datasets:

Dataset	N (Patients)	Channels	Tasks/Evaluation	Public Metadata / Annotations
Omni-iEEG	302	20–>150/pat	EZ localization, HFO, sleep	Standardized BIDS, 36K+ event annotations (Duan et al., 17 Feb 2026)
SWEC-ETHZ	16/18	36–128	Online seizure (iEEG)	Multiple cohorts, preprocessed metadata
Music Recon.	29	~100	Speech/music BCI	Open feature/label structure, MNI alignment
AJILE12	12	Varies	Move/rest BCI	Extensive event segmentation

Standardized evaluation metrics (macro-F1, ROC-AUC, channel/patient-level outcome AUC) and clinical labels (SOZ, outcome, resection) now enable quantitative, reproducible benchmarking (Li et al., 24 Feb 2025, Duan et al., 17 Feb 2026). Transfer learning (e.g., from audio models or clean iEEG to noisy EEG) increases cross-dataset performance and robustness (Carzaniga et al., 10 Feb 2025).

6. Technical Limitations and Future Directions

Current challenges and promising avenues include:

• Heterogeneity and generalizability: Inter-patient variation in electrode coverage and brain anatomy necessitates models robust to montage variation (as in RISE-iEEG), but performance often degrades by 5–20% in cross-subject/cross-dataset generalization (Memar et al., 2024, Memar et al., 2024).
• Adaptive and interpretable modeling: Nonstationarity and “meta-state” transitions in long-term iEEG recordings require adaptive retraining or meta-state-driven transfer learning (Yang et al., 2022).
• Data fidelity and compression: Lossy neural compressors (BrainCodec) demonstrate up to 64× compression of iEEG without loss in automated or expert performance, enabling real-time telemetered or wearable applications (Carzaniga et al., 10 Feb 2025).
• Real-time and ultra-low-power operations: Neuromorphic, event-driven iEEG and sparse hyperdimensional computing accelerators achieve actionable detection in sub-10 s with <1 μW power for implantable or intraoperative use (Cuyckens et al., 10 Oct 2025, Sharifshazileh et al., 2020).
• Cross-modal and non-invasive iEEG: Geometric-physics-guided and data-driven approaches for reconstructing iEEG from EEG continue to advance toward clinical-grade “virtual” iEEG (Pham et al., 17 May 2026, Dong et al., 3 Apr 2026, He et al., 27 Feb 2026).
• Unified benchmarks and open-source infrastructure: Distributed, multimodal annotation tools and standardized task definitions are establishing a replicable foundation for biomarker discovery and clinical translation (Duan et al., 17 Feb 2026, Li et al., 24 Feb 2025).

iEEG remains foundational for both neuroscience discovery and clinical neurotechnology. Its rapidly expanding computational, data, and translational landscape is paralleled by the shift to open, scalable, and harmonized research practices.