Human Intracranial Recordings

Updated 13 October 2025

Human intracranial recordings are direct measurements of brain electrical activity using implanted electrodes, offering millisecond precision and localized spatial resolution.
They employ modalities like ECoG, sEEG, and microelectrode arrays to precisely localize neural events for both clinical applications, such as epilepsy surgery, and cognitive research.
Advances in electrode materials, recording density, and signal processing techniques enable scalable data integration and improved neuromodulation and neuroprosthetic control.

Human intracranial recordings refer to the direct measurement of electrical activity from within the human brain using electrodes implanted on the cortical surface or within brain tissue. Such recordings offer unrivaled spatiotemporal resolution, signal-to-noise characteristics, and the capacity to interrogate the dynamics of local neuronal populations and single neurons during cognition, behavior, and disease. These measurements—encompassing modalities such as electrocorticography (ECoG), stereo-electroencephalography (sEEG), and microelectrode arrays—represent one of the most informative methodologies in contemporary neuroscience and neuroengineering.

1. Methodological Principles and Recording Modalities

Intracranial recordings are acquired through the surgical implantation of electrodes, mainly for clinical indications such as presurgical epilepsy localization. ECoG electrodes are placed on the cortical surface (subdural grids or strips), providing high spatial (often millimeter-scale) and temporal (sub-millisecond) resolution. sEEG involves depth electrodes advanced into deep cortical and subcortical structures. Microelectrodes (single or multi-contact arrays) can resolve single-unit or multi-unit activity, allowing single-neuron and population-level analyses (Stockart et al., 9 Oct 2025).

Key engineering and anatomical considerations include:

Material properties: Innovations in electrode materials (e.g., Pt-Au-PEDOT:PSS, platinum-silicone, hydrogels) enable sub-100 μm thick, flexible, high-density arrays (densities >1,000 sites/cm²) that reduce mechanical mismatch (e.g., Young’s modulus: tungsten ~250 GPa vs. CNS tissue ~1–10 kPa, with stress σ = E ⋅ ε) and chronic inflammatory response (Bourdillon et al., 5 Feb 2024).
Spatial/temporal coverage: Contemporary systems can record from hundreds (to over 1,000) sites at kilohertz sampling rates, capturing both local field potentials and action potentials at multiple scales and cortical depths (surface, columnar, laminar/matrix) (Bourdillon et al., 5 Feb 2024).
Signal processing: Signal quality is maintained through analog conditioning (e.g., low-noise amplifiers, filter banks), referencing strategies (common average, Laplacian), and computational techniques for artifact rejection and feature extraction (e.g., current source density CSD = –σ∇²V).

2. Neuroscientific Applications

Intracranial recordings have transformed our understanding of brain function by enabling:

Cognitive and naturalistic task studies: High-fidelity investigation of perception, language, memory, motor planning, decision-making, and sleep in real-world and controlled paradigms (Wang et al., 2017, Ashmaig et al., 2021, Wang et al., 13 Nov 2024). For instance, large multilineage datasets such as Brain Treebank (10 subjects, 1,688 electrodes, >38,000 sentences) combine millisecond-aligned iEEG with richly annotated linguistic input to reveal fine-grained neural encoding of speech and contextual features (Wang et al., 13 Nov 2024).
Neural correlates of consciousness: ECoG, sEEG, and microelectrode studies have resolved population-level and single-neuron signatures of perceptual awareness, delineated by high gamma activity (HGA) and firing rates in occipitotemporal and prefrontal regions, and informed ongoing debates between cognitive (global workspace) and sensory-localist theories (Stockart et al., 9 Oct 2025).
Self-generated thought and memory: Direct stimulation of the medial temporal lobe (hippocampus, parahippocampal cortex) reliably induces vivid recollection and mental imagery (>50% frequency in case studies), whereas hubs like posterior cingulate and inferior parietal cortices—although active in fMRI—do not elicit such experiences on stimulation, underscoring the causal specificity afforded by invasive methods (Fox, 2017).

3. Clinical and Translational Domains

Intracranial recordings inform a spectrum of diagnostic and therapeutic applications:

Epilepsy surgery and seizure localization: The fine spatial resolution of modern arrays (e.g., Neuropixel, Neural-Matrix) enables precise identification of epileptogenic zones, distinguishing seizure initiation within specific neocortical layers via current source density (CSD) and independent component methods (Bourdillon et al., 5 Feb 2024). Automated onset localization remains an active area, where algorithms’ methodological choices (baseline definitions, frequency band inclusion, consideration of electrodecrement) exert a material effect on identified seizure onset and, consequently, surgical outcomes (Gascoigne et al., 17 Oct 2024).
Neuromodulation and stimulation: iEEG-guided brain stimulation allows both local and network-level control, with prospective prediction of band power effects (delta/theta, 2–8 Hz) based on long-term functional connectivity variability (FC_L) as modeled via linear mixed-effects frameworks (Papasavvas et al., 2021). Adaptive stimulation protocols for epilepsy, depression, and movement disorders increasingly rely on such individualized neural markers.
Real-time neuroprosthetic interfaces: Low-power neuromorphic systems now integrate headstage signal acquisition, event-driven spike conversion, and specialized spiking neural network chips for continuous online detection of biomarkers such as high-frequency oscillations (HFOs, 80–500 Hz) with clinically actionable accuracy (e.g., specificity 100%) and power consumption orders of magnitude below software alternatives (Sharifshazileh et al., 2020).

4. Large-Scale Datasets, Normative Mapping, and Data Integration

The field has witnessed a progression toward multi-center, population-scale analyses:

Normative mapping: Robust population atlases (e.g., 502 subjects, 15 centers) model how age, sex, and institutional factors (e.g., sampling rate, electrode hardware) affect regional band power trajectories; linear mixed-effects models capture subject and site variance, confirming that age significantly modulates frequency-specific power (negative for δ, positive for α/β) across the lifespan, while sex effects are negligible (Woodhouse et al., 27 Apr 2024).

| Predictor | Effect | Frequency specificity | |-----------------|------------------------|--------------------------------------| | Age | Decrease in δ, increase in α/β | Uniform across regions | | Sex | Minimal/negligible | All bands | | Hospital site | Significant variance | All bands (up to 30% variability) |

Handling data heterogeneity: Modern frameworks—such as Homogeneity-Heterogeneity Disentangled Learning (H2DiLR) and Neural Mixture of Brain Regional Experts (Neuro‑MoBRE)—explicitly decompose multi-subject intracranial signals into shared (cross-subject) and idiosyncratic (subject/task-specific) representations to enable generalizable decoding (e.g., across Mandarin lexical tones or language, seizure, and motor tasks) and robust zero-shot performance on unseen individuals (Wu et al., 13 Oct 2024, Wu et al., 6 Aug 2025).

5. Computational Approaches and Foundation Models

Data richness and inherent complexity in intracranial signals have fueled a methodological shift:

Deep learning and self-supervised methods: Transformer models (e.g., BrainBERT) and multimodal CNN-LSTM architectures (e.g., AJILE movement prediction) provide efficient representation learning; self-supervised masked modeling on large, continuous iEEG corpora supports downstream decoding with minimal supervision, even with variable electrode coverage (Wang et al., 2023, Wang et al., 2017).
Imputation and data quality: Deep Neural Imputation (DNI) leverages autoencoder and nearest-neighbor strategies to recover missing electrode data, supporting both time-series and spectral fidelity and preserving downstream decoding accuracy even with >50% missing channels (Talukder et al., 2022).
Benchmarking and task suites: Resources such as Neuroprobe aggregate standardized, multi-task decoding benchmarks for language, vision, and auditory features using datasets like Brain Treebank, providing a rigorous basis to evaluate linear and foundation models (e.g., BrainBERT, PopT), and revealing that appropriately tuned linear models can outperform more complex approaches in some regimes (Zahorodnii et al., 25 Sep 2025).

6. Limitations, Challenges, and Future Directions

Despite their scientific and translational power, intracranial recordings present nontrivial constraints and continually evolving frontiers:

Sampling bias and generalizability: Recordings are limited to clinical populations (mainly epilepsy patients), with electrode placement dictated by patient-specific medical needs, leading to highly variable cortical and subcortical coverage (Stockart et al., 9 Oct 2025). Multi-site normative mapping and imputation approaches are required to mitigate sampling bias and interpret findings.
Technical and analytical heterogeneity: Varying hardware, referencing standards, filtering, and artifact rejection strategies across institutions contribute to site effects that can explain a substantial fraction of the variance in normative maps—and must be controlled through cross-site harmonization or random-effects modeling (Woodhouse et al., 27 Apr 2024).
Ethical and practical constraints: The invasive nature restricts participation, session length, and repeatability. Emerging chronic implants and miniaturized, biocompatible, wireless systems may expand access and duration of recording (Bourdillon et al., 5 Feb 2024).
Dynamic, distributed, and multiscale modeling: Future paradigms aim to synthesize macro- and micro-electrode data, develop causally informed experimental designs (e.g., fusing contrastive and supraliminal paradigms for isolating neural correlates of consciousness), and construct large-scale foundation models tuned for the unique statistical structure of invasive neurophysiological data (Stockart et al., 9 Oct 2025, Wang et al., 2023).

7. Significance and Outlook

Human intracranial recordings have become indispensable for unraveling the neural basis of higher cognition, perception, language, memory, action, and clinical disorders. They offer millisecond temporal and millimeter spatial resolution, access to high gamma and single-unit activity, and the ability to causally interrogate brain function via stimulation. Technological advances in electrode materials, recording density, and analytic frameworks are extending their reach—enabling normative mapping, unbiased decoding across individuals, improved neuroprosthetic control, and rigorous tests of neurobiological theory. Persistent challenges of invasiveness, sampling bias, institutional heterogeneity, and the need for scalable analytic solutions remain central areas of research and development.