Multiregional iEEG Data Analysis

• Multiregional iEEG data are high-resolution recordings obtained via implanted electrodes across diverse cortical and subcortical regions, capturing precise spatiotemporal neural dynamics.
• Advanced preprocessing and feature extraction pipelines standardize power spectral density and band power metrics to enable robust cross-subject and region-level comparisons.
• Spatially informed decoding models, including transformer architectures, enhance neural classification accuracy for applications in disease localization and functional mapping.

Multiregional intracranial electrophysiology (iEEG) data comprise high-resolution neural recordings obtained via electrodes implanted in distributed cortical and subcortical regions in human subjects, most commonly epilepsy patients undergoing presurgical assessment. These datasets enable the simultaneous measurement of spatiotemporally precise neural signals across functionally distinct networks. Rigorous acquisition, standardized annotation, and advanced analytic pipelines have enabled the construction of normative atlases, decoding benchmarks, and the development of foundation models for brain state classification and neural representation learning. Multiregional iEEG provides a uniquely direct window into human neurophysiology, supporting quantitative investigation into functional organization, interregional connectivity, disease localization, and cognitive processes.

1. Data Acquisition and Anatomical Coverage

Multiregional iEEG datasets are collected via stereotactic or subdural electrode implantation spanning diverse cortical and subcortical target regions. Clinical protocols dictate anatomical coverage, resulting in variable channel counts and heterogeneity across subjects. In large-scale normative mapping studies, channels are localized via post-implant CT coregistered to preoperative T1 MRI, then assigned to segmented regions using parcellations such as the Desikan–Killiany or Lausanne 36/250 atlases; only contacts within 5 mm of grey matter parcels are retained (Wang et al., 2023, Woodhouse et al., 27 Apr 2024, Wang et al., 13 Nov 2024). Typical datasets exhibit broad spatial sampling, e.g., hundreds of contacts per subject across temporal, frontal, parietal, insular, mesial structures, and deep nuclei (GPi, STN, thalamus) (Javadzadeh et al., 31 Oct 2025).

Recording durations span minutes to multiple days (mean up to 4.58 days, >4,800 hours aggregate per cohort (Wang et al., 2023)), with sampling rates ranging from 200 Hz (after downsampling) up to 2,048 Hz (native acquisition). Anatomical metadata include subject-space coordinates, standardized atlas assignments, and cross-modality co-registration files in formats such as BIDS-iEEG, HDF5, or MATLAB structure (Delorme et al., 2022).

2. Preprocessing, Segmentation, and Feature Extraction

Preprocessing pipelines apply robust referencing (common-average, Laplacian, or bipolar), bandpass filtering (0.5–80 Hz), and line-noise rejection (notch filters at 50/60 Hz and harmonics). Channels are excluded based on extreme amplitude, flatlines, or manual artifact review (Wang et al., 2023, Zahorodnii et al., 25 Sep 2025, Oganesian et al., 13 Dec 2025).

Epoching segments continuous traces into windows tailored for downstream analysis: 30 s (abnormality mapping (Wang et al., 2023)), 70 s (normative mapping (Woodhouse et al., 27 Apr 2024)), 1–3 s (decoding tasks (Oganesian et al., 13 Dec 2025)), or aligned to events such as word onsets in naturalistic tasks (Wang et al., 13 Nov 2024). Each window is classified into ictal, peri-ictal, or interictal based on clinical seizure annotation, enabling targeted analysis of both baseline and task-induced states (Wang et al., 2023).

Feature extraction includes power spectral density (PSD; Welch’s or multitaper), band power integration over canonical ranges (delta–gamma, often custom exclusions around line frequency), and log transforms. Channel- or region-level features are normalized (sum-to-one, L1 or log normalization) (Woodhouse et al., 27 Apr 2024). Additional functional metrics include high-gamma envelopes, phase-locking values, and entropy-based descriptors (Oganesian et al., 13 Dec 2025, Memar et al., 9 Dec 2024).

3. Normative Mapping, Abnormality Scoring, and Comparative Analysis

Normative multiregional iEEG mapping involves aggregating population-level recordings from non-pathological brain tissue (e.g., RAM cohort, n=249 (Wang et al., 2023); multi-centre cohort, n=502 (Woodhouse et al., 27 Apr 2024)). Canonical band powers are computed per region and frequency band, yielding empirical means and standard deviations for each region/band pair ($\mu_{i,f}$, $\sigma_{i,f}$). Statistical models incorporate subject demographics (age, sex), hospital effects (random intercepts), and anatomical normalization (Woodhouse et al., 27 Apr 2024).

Test-cohort abnormality maps are constructed via z-scoring: $$z_{i,f}(t) = \frac{b_{i,f}(t) - \mu_{i,f}}{\sigma_{i,f}}$$ Max absolute z-score across bands produces the regional abnormality score $A_i(t)$ (Wang et al., 2023). Comparative metrics—such as Distinguishability of Resection vs. Spared (DRS/AUC)—quantify the alignment of abnormal regions with surgically resected tissue, providing principled and temporally stable outcome prediction (median DRS AUC=0.69 for seizure-free classification) (Wang et al., 2023).

Normative lifespan atlases adjust for age and centre effects via linear mixed-effects regression: $$y_{i,r,b} = \beta_0 + \beta_{\rm age}\,\mathrm{Age}_i + \beta_{\rm sex}\,\mathrm{Sex}_i + \gamma_{h(i)} + \varepsilon_{i,r,b}$$ Enables baseline normalization in disease localization and cross-paper aggregation (Woodhouse et al., 27 Apr 2024).

4. Multiregional Decoding, Benchmarking, and Functional Mapping

Multiregional iEEG decoding leverages the spatial integration of features derived from distributed electrodes. Combined-channel models fuse spatial information by ensemble voting among classifiers trained on channel-specific (e.g., 18-dimensional (Memar et al., 9 Dec 2024)) feature vectors. Greedy forward selection identifies optimal channel subsets per task, boosting F1 scores compared to best-channel models (up to 0.84 ± 0.08) (Memar et al., 9 Dec 2024). Region-level channel appearance in final ensembles aligns with physiological task specificity, e.g., superior temporal in music, precentral gyrus in movement (Memar et al., 9 Dec 2024).

Advanced benchmarks, such as Neuroprobe, provide standardized task definitions, evaluation splits (within-session, cross-session, cross-subject), and open leaderboards using large datasets like Brain Treebank (>1,500 electrodes, 223k word tokens, fine-grained multimodal annotations) (Zahorodnii et al., 25 Sep 2025, Wang et al., 13 Nov 2024). Decoding protocols utilize both hand-crafted and deep-learned features, reporting metrics such as AUROC, F1, and regression $r$, facilitating rigorous model comparison.

Spatially informed modeling (BaRISTA (Oganesian et al., 13 Dec 2025), MVPFormer (Carzaniga et al., 25 Jun 2025), functional transformer frameworks (Javadzadeh et al., 31 Oct 2025)) enables tokenization at multiple anatomical scales—channel, parcel, lobe—and self-supervised pretraining tasks. Larger-scale spatial encoding (parcel or lobe-level) enhances downstream classification (e.g., sentence onset AUC up to 0.862 ± 0.016), with region-level tokenization aiding in efficient and robust representation learning while allowing accurate channel-level reconstruction (Oganesian et al., 13 Dec 2025).

5. Aggregation, Cross-Subject Modeling, and Foundational Architectures

Cross-subject aggregation is enabled by functional embeddings: CNN-based or wavelet-based encoders produce locality-sensitive vector representations of each electrode or region, inducing clustering consistent with neural identity rather than just anatomical coordinates (Javadzadeh et al., 31 Oct 2025, Carzaniga et al., 25 Jun 2025). Contrastive objectives (pairwise Siamese, supervised contrastive, variance penalties) encourage region-specific grouping (held-out channel discrimination up to 49.18%) and facilitate zero-shot transfer across subject- and session heterogeneity (Javadzadeh et al., 31 Oct 2025).

Transformer architectures leverage multi-variate parallel attention (MVPA), disentangling content, time, and channel attention terms in self-attention computation: $$A_{(c,t),(c',t')} = a^{\mathrm{cont}} + a^{\mathrm{time}} + a^{\mathrm{chan}}$$ Relative encoding supports variability in electrode placements, enabling generative models (MVPFormer) to achieve expert-level seizure detection across unseen subjects (episodic Cohen’s kappa = 0.57; false positives/h = 0.17) (Carzaniga et al., 25 Jun 2025).

Masked latent reconstruction tasks (BaRISTA) pretrain on spatially masked tokens. Joint space-time attention and multiscale CNNs propagate spatial structure, improving decoding and preserving fine-grained temporal content (Oganesian et al., 13 Dec 2025).

6. Data Structure, Standards, and Computational Reproducibility

Multiregional iEEG datasets are organized under standardized frameworks such as BIDS-iEEG (folder hierarchy, TSV/JSON metadata, coordinatesystems), supporting interoperability and FAIR principles (Delorme et al., 2022). Electrode placements are annotated by atlas label and coordinate, referencing to MNI or subject-native space. Event tags adopt HED standards for cross-paper analysis. Data formats include EDF/BDF/HDF5/MATLAB, with sidecar files defining parameters and region identity (Wang et al., 13 Nov 2024).

NEMAR and related platforms model immediate visualization (Time/Frequency-Domain plots, STFT, wavelet transforms), online electrode overlay, and batch computational workflows via supercomputing portals (NSG/XSEDE). Pipelines and environments are containerized (Singularity/Docker); all code, outputs, and metadata are shared for reproducibility (Delorme et al., 2022). Multiregional comparative analysis mandates coordinate system alignment, region definition, and reporting of electrode inter-distance (Delorme et al., 2022, Wang et al., 13 Nov 2024).

7. Applications, Limitations, and Future Directions

Multiregional iEEG data support disease localization (epileptogenic zone identification), functional mapping (language, movement, cognition), cross-subject modeling, and neural decoding in BCIs. Temporal stability of abnormality maps, region-level aggregation, and normative atlas referencing inform presurgical planning and outcome forecasting (Wang et al., 2023, Woodhouse et al., 27 Apr 2024).

Limitations include clinical-driven and sparse regional coverage, reliability of anatomical-to-functional alignment, constraints in cohort demographic range (e.g., lifespan coverage), and variability in acquisition hardware/protocols (Woodhouse et al., 27 Apr 2024, Javadzadeh et al., 31 Oct 2025). Data-driven spatial embedding, functionally defined groupings, and multi-scale transformers offer plausible avenues to address aggregation and generalization challenges (Oganesian et al., 13 Dec 2025, Javadzadeh et al., 31 Oct 2025, Carzaniga et al., 25 Jun 2025).

A plausible implication is that integration of spatial information at anatomically meaningful scales enables robust neural decoding and model pretraining even under heterogeneous, sparse, and subject-specific montages—facilitating scalable population-level studies and next-generation neural foundation models (Carzaniga et al., 25 Jun 2025, Oganesian et al., 13 Dec 2025, Javadzadeh et al., 31 Oct 2025).