PaperlayerView exact revision on arXiv

Source-linked AI summary

Virtual iEEG from Scalp EEG: Charting the Landscape of Source Imaging, Intracranial Inference and Reconstruction

Dongyi He, Xiangkai Wang, Hongjie Yan, Luping Song, Wai Ting Siok, Nizhuan Wang

arXiv:2608.26998v1cs.CV

TL;DR

iEEG provides precise access to focal and deep activity, but its invasiveness and restricted coverage motivate scalp-to-intracranial inference. This review develops a target-centred framework and evaluates evidence across prediction targets, physical identifiability, validation conditions, and utility. It finds support for selected intracranial events and representations, but not unique recovery of arbitrary contact-level activity.

  • Problem

    Invasiveness and restricted anatomical coverage limit routine iEEG use, motivating inference of intracranial activity from scalp EEG.

  • Method

    The review organizes scalp-to-intracranial inference into event inference, feature translation, and waveform reconstruction, separating predictability from observability, identifiability, fidelity, and utility.

  • Results

    Current evidence supports inference of selected intracranial events, low-frequency components, and task-related representations, but not unique recovery of arbitrary contact-level activity.

  • Takeaways & Limitations

    Virtual iEEG should be evaluated as target- and regime-dependent inference, with value demonstrated beyond scalp EEG and EEG source imaging.

  • Takeaways & Limitations

    Reference conventions and irreducible conditional variance can limit interpretation and prevent recovery of activity absent from scalp data.

Abstract

from arXiv · show

Intracranial electroencephalography (iEEG) provides temporally precise and spatially specific access to neural activity from focal and deep brain regions, but its invasiveness and restricted anatomical coverage limit routine use. These constraints have motivated scalp-to-intracranial inference, termed virtual iEEG when model outputs carry iEEG-defined event, feature, representation, or contact-level waveform semantics. This review presents a target-centred framework distinguishing event inference, feature translation, and waveform reconstruction, while separating predictability from observability, identifiability, fidelity, and utility. Evidence is evaluated according to cohort independence, anatomical and spectral coverage, train--test separation, and target-patient adaptation. Current studies support inference of selected intracranial events, low-frequency components, and task-related representations, but not unique recovery of arbitrary contact-level activity. Stronger validation requires appropriate controls, source-imaging baselines, uncertainty assessment, and incremental-utility testing. Future progress depends on independent paired datasets and prospective evidence that virtual iEEG adds value beyond scalp EEG and EEG source imaging.

1 Introduction

iEEG offers precise access to focal and deep neural activity, but invasiveness and limited coverage motivate scalp-to-intracranial inference. This review distinguishes target types and evidential claims while evaluating whether virtual iEEG can recover clinically meaningful intracranial information.

  • Motivation: iEEG provides spatially specific access to focal and deep activity, whereas scalp EEG is safe and portable but spatially mixed and attenuated.These trade-offs motivate non-invasive inference of intracranial activity.
  • Scope and targets: Virtual iEEG methods span hidden-event inference, intracranial feature or representation inference, and explicit contact-level waveform reconstruction.EEG source imaging remains an anatomy-aware reference framework rather than virtual iEEG itself.
  • Conceptual framework: The review separates observability, predictability, identifiability, fidelity, and utility when assessing scalp-to-intracranial claims.It also distinguishes STIR as the broader task family from virtual iEEG as outputs assigned intracranial semantics.
  • Evidence boundaries: Current evidence supports selected intracranial events, low-frequency components, and task-related representations, but not unique recovery of arbitrary contact-level activity.Recoverability depends on anatomy, frequency, state, population, acquisition context, and target-patient adaptation.
  • Validation and translation: The review audits studies by cohort independence, anatomical and spectral coverage, train–test separation, and target-patient adaptation.It calls for conditioning controls, uncertainty calibration, matched scalp-EEG and ESI baselines, independent paired datasets, and prospective incremental-utility testing.

2 Prediction targets and evidential scope

Scalp-to-intracranial methods differ by target semantics, predictive form, and intended utility. The review separates predictability from identifiability and emphasizes that realistic waveform prediction does not establish unique contact-level recovery.

  • Target semantics: Scalp-to-intracranial inference targets events, intracranial features or representations, and contact-level waveforms.
  • Target semantics: STIR uses scalp EEG to estimate iEEG-defined events, features, waveforms, or representations, whereas ESI estimates generators or source-space activity under a forward model.
  • Predictive form: Deterministic and probabilistic estimation can apply to any target, with probabilistic reconstruction estimating a conditional distribution.
  • Predictability and identifiability: Under squared error, the optimal deterministic predictor is the conditional mean E[Y_i | x, A, r_i, Ω], not a one-to-one inversion of physical measurements.
  • Predictability and identifiability: Waveform similarity measures predictability under an evaluated distribution, while identifiability requires sensitivity to conditioning input, matched alternatives, and uncertainty reflecting broad conditional distributions.
  • Virtual contacts as measurement-level targets: Virtual electrodes and reconstructed channels represent measurement-level targets rather than elemental neural quantities.

3 Biophysical determinants of observability

Observability depends on measurement geometry, source properties, references, frequency, and context rather than a universal depth or frequency hierarchy. Scalp-null source components can change contact waveforms without changing scalp EEG, imposing a physical identifiability limit.

  • Measurement model: Scalp EEG and iEEG are distinct reference-dependent measurements of partially shared latent neural activity.
  • Identifiability limits: A perturbation invisible to the scalp measurement can alter the target-contact waveform, so those components are not identifiable from scalp EEG alone.
  • Source and measurement determinants: Physical observability varies with generator extent, orientation, synchrony, conductivity, sensor geometry, montage, and reference.
  • Frequency: High-frequency activity is generally harder to observe because generators are focal and weakly synchronous, whereas slow rhythms more often reflect extended coherent populations.
  • Empirical determinants: Recognisable scalp correlates occurred for 90% of intracranial spikes engaging more than 10 cm2 of cortex, versus 10% below that area and none below 6 cm2.
  • Empirical determinants: Demonstrated recoverability varies across structure, depth, frequency, event type, state, scalp distance, and reference, with selected-contact performance potentially overstating coverage.

4 EEG source imaging as an anatomical reference frame

ESI provides an anatomy-aware reference and diagnostic baseline for scalp-to-intracranial inference, but its estimates remain non-unique and distribution-dependent. Validation standards support different claims and should not be treated as interchangeable.

  • ESI addresses source localisation, observability, and anatomical plausibility, whereas virtual iEEG additionally requires target-specific predictability, fidelity, and utility evidence.
  • The scalp-to-intracranial inverse problem is non-unique, so regularisation selects one scalp-compatible source estimate shaped by anatomy, geometry, noise, reference, and prior.
  • ESI-derived virtual contacts are forward projections of estimated sources into an intracranial montage, not reconstructions of physical contact measurements.
  • ESI validation spans stimulation, spontaneous intracranial recordings, simulations, and clinical outcomes, but each reference standard supports a different claim.
  • Source-imaging baselines test whether learned inference adds information beyond anatomy-aware non-invasive estimation, while successful localisation alone does not establish contact-level waveform reconstruction.
  • Distribution shift limits both ESI and learned reconstruction across conductivity models, electrode layouts, implantation patterns, pathology, states, and tasks.

5 Hidden-event inference and event-centred representations

Hidden-event inference and event-centred representations show bounded scalp correlates of selected intracranial activity, mainly within constrained cohorts, events, anatomies, and evaluation settings. These findings support target-specific inference but do not establish arbitrary continuous contact-level recovery.

  • 5.1 Scalp correlates of intracranial seizures: Hidden-event inference estimates p(c | x), whereas feature translation tests whether scalp data preserve an iEEG-derived representation.
  • 5.1 Scalp correlates of intracranial seizures: 50 of 68 seizures with more than five seconds of invasive lead were identified as early, yielding 73% sensitivity in an assisted retrospective evaluation.
  • 5.1 Scalp correlates of intracranial seizures: 97%+ epoch-level accuracy was reported for mesial-temporal occult-onset classification across repeated patient-disjoint splits, but epoch accuracy does not establish event-level false-alarm performance.
  • 5.2 Operational event detection: In 8,395 hours from 51 patients, HEAnet reached event-level AUC 0.89 but precision–recall AUC 0.39; at PPV near 0.7, sensitivity was 0.25 ± 0.08.
  • 5.2 Operational event detection: Propagation was associated with scalp observability: 91% of mesial-temporal channels versus 16% of non-propagating events showed significant scalp zero-crossing patterns.
  • 5.3 Event-centred intracranial representations: Representation methods use shared sparse codes or learned encoders to map scalp observations toward intracranial representations, with evaluation restricted to training-represented windows and anatomies.
  • 5.3 Event-centred intracranial representations: Evidence is strongest for selected events and event-centred representations within sampled settings; arbitrary continuous contact-level waveforms across patients, regions, and tasks remain untested.

6 Waveform reconstruction, uncertainty, and generalisation

Waveform reconstruction estimates contact-level iEEG using paired scalp–iEEG data, anatomical and contextual conditioning, and objectives targeting waveform fidelity or distributional realism. Evidence varies substantially by adaptation regime, cohort, and objective, with patient-specific performance and cross-patient limitations requiring careful interpretation.

  • Reconstruction formulation: Contact-conditioned reconstruction combines concurrent scalp EEG with anatomy, contact geometry, and deployment context to estimate target-contact iEEG waveforms.The formalisation includes scalp segments, simultaneous target-contact waveforms, anatomical information, contact location, and acquisition or deployment context.
  • Objectives and estimands: Waveform objectives can combine time-domain, spectral, and correlation terms, while event- and representation-oriented objectives quantify different estimands.Performance under one objective does not establish fidelity for another target.
  • Empirical reconstruction studies: E2SGAN improved transformed dynamic-time-warping and spectral-distance measures over raw EEG in seven-patient pre-ictal scalp–SEEG data.The model used paired signals, 64-Hz downsampling, overlapping windows, and channel-pair selection based on distance and spectral similarity.
  • Empirical reconstruction studies: In an 18-patient foramen-ovale cohort, a deterministic U-Net achieved 68% leave-patient-out IED classification accuracy, while a variational extension reported within-patient MSE of 0.014 and Pearson correlation of 0.35.The studies used event-centred segments, and the U-Net lacked a quantitative waveform-fidelity metric.
  • Generalisation and adaptation: CAST used 20% of simultaneous scalp–iEEG data from each test patient for adaptation and retained contacts meeting r ≥0.15, illustrating adapted cross-patient evaluation.The study pooled 1,282 contacts from working-memory and visual-task patients under leave-one-patient-out training.
  • Uncertainty and evidence boundaries: Probabilistic models represent conditional target distributions, but likelihood, adversarial realism, or latent variables do not by themselves establish scalp dependence or calibrated uncertainty.NeuroFlowNet reported approximately 0.50 random cross-trial and 0.54 held-out-session temporal correlation, with regional cross-session correlations from approximately 0.05 to 0.84; its nominal 90% intervals were not fully calibrated.

7 Validation and translational evidence

Validation must distinguish agreement with invasive targets, information added by concurrent scalp EEG, and persistence under the intended deployment regime. Translational claims therefore require matched baselines, patient-level separation, calibrated metrics, and prospective tests of incremental utility.

  • Validity requirements: Validation asks whether outputs agree with invasive targets, depend on concurrent scalp EEG, and persist under intended deployment conditions.These correspond to measurement, conditional, and deployment validity.
  • Data partitioning: Cross-patient claims require patient-level separation, while within-patient transfer can use separated sessions, seizures, or recording blocks; overlapping windows are not independent observations.Data-dependent preprocessing must also be confined to the training partition.
  • Conditioning controls: Comparing p(y | B) with p(y | x, B), alongside shuffled-pair, scalp-ablation, anatomy-mismatch, and reference controls, tests whether concurrent scalp EEG contributes information.A reproducible improvement from the marginal to scalp-conditioned predictor supports an incremental scalp contribution.
  • Metric validity: Waveform evaluation should combine correlation or concordance with amplitude, bias, morphology, and frequency-resolved measures because Pearson correlation alone is insufficient.Metrics should follow the estimand rather than the model architecture.
  • Utility interpretation: Improved decoder performance is a utility claim rather than evidence of contact-level reconstruction, and representations with limited waveform fidelity may still preserve task-relevant information.Synthetic representations can improve downstream models through regularisation rather than waveform recovery.
  • Translational evidence: Current reconstruction studies have not demonstrated prospective effects on clinical decisions, so translation should progress through external or silent-mode evaluation, blinded workflow studies, and outcome assessment.These stages test incremental value without treating model outputs as direct measurements.

8 Outstanding questions and research priorities

Research priorities centre on identifying where concurrent scalp EEG adds information, where recoverability ends, whether uncertainty detects unsupported outputs, and whether benefits generalise without target-patient invasive data. Progress requires provenance-rich paired datasets, locked benchmarks, independent validation, and utility comparisons beyond scalp EEG and ESI.

  • Information contribution: Future studies should quantify the scalp-conditioned increment by comparing p(y | B) with p(y | x, B), while measuring target reliability and reference sensitivity.The aim is to determine how much information the concurrent scalp segment contributes.
  • Recoverability limits: Recoverability should be stratified by anatomy, depth, frequency, state, and event type, using stimulation and spontaneous recordings as complementary evidence.These factors define where scalp-to-intracranial inference may be bounded.
  • Uncertainty: Calibration and risk–coverage analysis should test whether predictive uncertainty and abstention identify high-error contacts, states, or anatomies.This directly evaluates whether uncertainty can flag unsupported reconstructions.
  • Generalisation: Zero-shot claims require patient- and site-level evaluation without target-patient iEEG for fitting, calibration, contact selection, or hyperparameter choice.Scalp-only test input alone does not establish target-iEEG-free generalisation.
  • Incremental utility: Clinical, BCI, and cognitive-neuroscience claims should compare against tuned non-invasive baselines and use endpoints matched to the intended application.Incremental utility must be assessed beyond scalp EEG and ESI where applicable.
  • Negative evidence: Well-powered failures under reliable targets and adequate controls can define anatomical, spectral, or state-dependent reconstruction boundaries.Boundary results complement pooled positive reconstruction scores.
  • Infrastructure and benchmarks: Paired datasets need provenance for anatomy, montage, synchronisation, hardware, state, event origin, eligibility, and exclusions, with locked splits and patient-level reporting.Benchmarks should separate temporal, patient, and site transfer, adaptation, and target-iEEG-free evaluation.

9 Conclusion

The review frames virtual iEEG as prediction of iEEG-defined events, features, representations, or contact-level waveforms from scalp EEG, distinct from source-space ESI. Current evidence is target- and regime-dependent, supporting selected inferences but not unique arbitrary-contact recovery or established target-iEEG-free deployment.

  • Framework and scope: STIR and virtual iEEG predict iEEG-defined events, features, representations, or contact-level waveforms from scalp EEG, whereas ESI estimates neural generators in source space.The framework separates observability, predictability, identifiability, fidelity, and utility.
  • Supported evidence: Existing studies support selected intracranial events and representations, plus waveform reconstruction in paired, patient-specific, or adapted cross-patient settings.The evidence remains dependent on target and evaluation regime.
  • Boundaries and next steps: Current evidence does not establish unique recovery of arbitrary contact-level activity or general target-iEEG-free deployment.The review calls for conditioning controls, matched scalp-EEG and ESI baselines, calibrated uncertainty, independent validation, and prospective utility testing.
Loading 2608.26998v1…