Organizations: Department of Language Science and Technology, The Hong Kong Polytechnic University, Hung Hom, Hong Kong SAR, China · School of Artificial Intelligence, Chongqing University of Technology, Chongqing, China · Affiliated Lianyungang Hospital of Xuzhou Medical University, Lianyungang, China · School of Artificial Intelligence, Southern University of Science and Technology Hospital, Shenzhen, China
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.
Figures & tables
Figure 1: Target-centred framework for scalp-to-intracranial EEG inference. (a) Available information comprises scalp EEG x , anatomy A , target-contact geometry ri , deployment context Ω , and, where permitted, target-patient iEEG adaptation. (b) Prediction targets comprise event inference, feature translation, and waveform reconstruction, with waveform outputs represented as point estimates or conditional distributions. (c) Evidential claims are distinguished by observability, predictability, identifiability, fidelity, and utility.
Scalp EEG and head model; optional paired-iEEG validation
Paired or aligned scalp–iEEG; supervised target learning
Table 1: Comparison of EEG source imaging and scalp-to-intracranial reconstruction across targets, modelling assumptions, validation criteria, and evidential limits
Figure 2: Measurement model and identifiability limits in scalp-to-intracranial inference. (a) Scalp EEG x(t) and contact waveform yi(t) are distinct, reference-dependent measurements of partially shared neural activity q(t) . (b) A scalp-null perturbation can alter the intracranial waveform without changing the scalp observation, limiting contact-level identifiability. (b) Reconstruction is therefore formulated as conditional estimation p(yi∣x,A,ri,Ω) , constrained by anatomy, context, and optional adaptation; priors cannot recover information absent from scalp measurements.
Determinant
Mechanism
Favourable conditions
Limiting conditions
Source extent and synchrony
Net scalp amplitude depends on effective generator extent and field summation; synchrony increases summation but is not a binary prerequisite
Extended recruitment; coherent or otherwise constructively summating generators; high event-to-background SNR
Small effective extent; destructive cancellation; low SNR; empirical area thresholds are not universal [ 51 , 52 , 53 ]
Depth and orientation
Depth, cortical geometry, source orientation, and sensor coverage shape lead-field amplitude and SNR; none alone fixes visibility
Strong scalp projection; prominent deep rhythms under intracranial-informed analysis
Weak projection; low-SNR pure mesial-temporal single events; current subcortical evidence is limited to prominent 8–10-Hz rhythms [ 54 , 55 , 56 ]
Frequency content
Quasi-static propagation is approximately frequency independent; detectability instead depends on generator extent and frequency-specific background SNR
Sufficient extent and SNR; repeated delta signatures or prominent 8–10-Hz target rhythms
Neural state changes the recorded activity and can alter scalp–intracranial comparison performance
Pre-specified and matched state; state-stratified evaluation
Pooling wake, NREM sleep, or anaesthesia without stratification [ 50 , 57 ]
Event structure and analysis
Known event timing permits aggregation; averaging tests a repeatable scalp contribution, not single-trial recovery
Repeated, consistently timed events; event-locked or trigger-averaged analysis
Irregularly timed events; low-SNR single-trial targets; extrapolating averaged results to single events [ 55 ]
Reference and montage
Re-referencing is spatial processing and can alter amplitude, polarity, spectral power, and coherence
Explicitly reported scalp and iEEG montages; sensitivity analysis across reasonable references
Unreported or mismatched references; treating montage-specific results as reference-invariant [ 58 , 59 ]
Table 2: Biophysical and empirical determinants of scalp observability, with conditions favouring or limiting intracranial-signal detectability
Figure 3: Generalisation regimes defined by held-out units and target-patient iEEG availability. (a) Held-out unit and prior target-patient iEEG jointly distinguish within-patient, adapted cross-patient, and target-iEEG-free zero-shot evaluation; scalp-only test input does not itself establish zero-shot generalisation. (b) NeuroFlowNet [ 31 ] , CAST [ 32 ] , and Dong et al. [ 33 ] illustrate patient-specific, adapted cross-patient, and scalp-only inference after paired supervision, respectively. (c) Deployment ranges from retrospective settings permitting target-patient iEEG to pre-implant zero-shot use without invasive adaptation; anatomical conditioning cannot restore information absent from scalp measurements.
Study / Model
Cohort and estimand
Principal result
Evidence boundary
Lam et al. [ 25 ]
10 patients contributing 25 scalp-negative mesial-temporal seizures plus 13 patients contributing control records; leave-one-patient-out network detection
Eight of the 25 prespecified seizures plus one previously overlooked seizure detected; three false alarms; 0.31 false alarms/day; PPV 75%
PCA fitted to all seizure and control records before cross-validation; feature set selected on cross-validation performance; 8/9 detections occurred during sleep; local waveform recovery not evaluated
Primary training and expert-labelled testing used sleep; external diagnosis/lateralisation cohorts were scalp-only; awake sensitivity =0.06 at the sleep-derived PPV ≃0.9 threshold using FOnet labels
IED accuracy: 62% linear mapping; 65% direct time–frequency scalp features; 68% asymmetric-symmetric-autoencoder representation
The test subject’s scalp data selected the most-correlated of 17 subject-specific mappers; waveform comparison was qualitative; absolute correlation and waveform error were not reported
E2SGAN [ 30 ]
n=7 ; simultaneous pre-ictal EEG–SEEG; 15.875-s windows with 75% overlap; leave-one-patient-out evaluation; preselected channel pairs
log2 distance ratios versus raw EEG: DTW −0.414 , PSD −1.480 , Hellinger −0.221 ; best PSD and Hellinger results among compared models
Pair selection used distance intervals and signal similarity; contact-wise correlation and event endpoints were not reported; downsampling to 64 Hz imposed a 32-Hz Nyquist limit; separation relative to overlapping-window generation was not reported
Intra-subject mean-squared error 0.014, Pearson r=0.35 , cosine 0.34; inter-subject IED accuracy 69% versus 68% for ASAE and cGAN
Waveform metrics were intra-subject only; inter-subject models were restricted to training subjects with intra-subject accuracy >70% ; cross-subject waveform fidelity, diversity, interval coverage, and uncertainty calibration were not evaluated
NeuroFlowNet [ 31 ]
Three selected patients from a nine-patient simultaneous dataset; subject-specific medial-temporal reconstruction; random-trial 90:10 and session-held-out tests
Mean r : random-trial 0.50; session-held-out 0.54; regional session-held-out 0.05–0.84
No cross-patient test; spectral evaluation was mainly 0.5–50 Hz and excluded high-gamma/HFO claims; validation-set coverage of nominal 90% intervals was 38.1–43.2%
Table 3: Representative evidence for scalp-to-intracranial inference, including cohort design, principal results, and boundaries of the supported claims
Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.
Intracranial EEG (iEEG) provides high-fidelity neural recordings essential for clinical and brain-computer interface applications, but acquiring these signals requires invasive surgery. While recent studies have attempted to estimate iEEG from non-invasive scalp EEG, most rely on patient-specific models, creating a circular dependency: if surgery is required to collect training data, the non-invasive model offers limited practical benefit. In this study, we address the challenge of cross-subject iEEG reconstruction by predicting intracranial signals for unseen patients using models trained on other individuals. We propose CAST (Cross-Attention Spatial-Temporal Transformer), a machine learning framework that translates scalp EEG into multi-channel iEEG waveforms through a two-stage transfer learning strategy. First, a temporal encoder extracts multi-scale neural representations at three different resolutions. Then, because electrode placements vary substantially across patients, a channel-aware decoder is calibrated using only a few minutes of data from the target subject. We evaluated the proposed method using leave-one-subject-out cross-validation on two public datasets comprising 1,282 iEEG channels. Experimental results demonstrate that CAST reconstructs cortical signals located near the scalp surface substantially better than deep subcortical activity. In highly observable sensorimotor regions, the model achieved peak correlations of up to r=0.864 in the precentral gyrus. Furthermore, with a channel selection strategy, CAST obtained a mean correlation of r=0.545 on viable subjects, outperforming previous within-subject baselines. These findings indicate that cortical iEEG signals can be reconstructed for unseen subjects from scalp EEG without extensive patient-specific training, and that only a brief calibration phase is sufficient to adapt the model to new hardware configurations.
Tien-Dat Pham, Xuan-The Tran
HAI-Smartlink Research Lab, Anchi STE Company, Haiphong, Vietnam · School of Mechanical Engineering, Vietnam Maritime University, Haiphong, Vietnam
Low-channel wearable electroencephalography (EEG) is attractive for long-term monitoring, but four frontal electrodes provide only a sparse and spatially biased sampling of the scalp potential field. Virtual-channel methods should therefore be framed not as recovery of independent unmeasured brain activity, but as prior-guided conditional inference of posterior predictive scalp-potential representations at target electrode locations. We present FAVC-Net, a compact frequency-calibrated virtual-channel inference network that estimates 13 target channels from Fp1, Fp2, F7, and F8. The model combines shared multi-scale source encoding, source-state embeddings, target-conditioned signed source-block mixing, GATv2-based attention refinement, attention-consistent skip fusion, and weak Welch power spectral density calibration. The generator is trained as a task-agnostic reconstruction module, without class-label, classification, or CSP-like discriminative constraints, so that the virtual montage remains tied to conditional scalp-potential estimation rather than to a specific downstream decision. On the PRED+CT dataset, FAVC-Net achieved the best joint waveform-spectral operating point among neural and interpolation baselines. Its time-domain gains were modest, whereas log-spectral distance and PSD KL divergence were reduced by 30.50% and 38.94% relative to the strongest non-FAVC comparator. Under wearable-like source perturbations, the model preserved spectral fidelity and channel-frequency texture, with anti-collapse benefits most evident under EMG-like bursts and mixed stress. These results support virtual EEG channels as montage-compatible, frequency-calibrated posterior predictive representations derived from sparse frontal measurements, not as independent substitutes for physically recorded electrodes.
Minghao Xiao
School of Biomedical Science and Engineering, South China University of Technology, Guangzhou 511442, China · School of Automation Science and Engineering, South China University of Technology, Guangzhou 510640, China
Although obtaining deep brain activity from non-invasive scalp electroencephalography (sEEG) is crucial for neuroscience and clinical diagnosis, directly generating high-fidelity intracranial electroencephalography (iEEG) signals remains a largely unexplored field, limiting our understanding of deep brain dynamics. Current research primarily focuses on traditional signal processing or source localization methods, which struggle to capture the complex waveforms and random characteristics of iEEG. To address this critical challenge, this paper introduces NeuroFlowNet, a novel cross-modal generative framework whose core contribution lies in the first-ever reconstruction of iEEG signals from the entire deep temporal lobe region using sEEG signals. NeuroFlowNet is built on Conditional Normalizing Flow (CNF), which directly models complex conditional probability distributions through reversible transformations, thereby explicitly capturing the randomness of brain signals and fundamentally avoiding the pattern collapse issues common in existing generative models. Additionally, the model integrates a multi-scale architecture and self-attention mechanisms to robustly capture fine-grained temporal details and long-range dependencies. Validation results on a publicly available synchronized sEEG-iEEG dataset demonstrate NeuroFlowNet's effectiveness in terms of temporal waveform fidelity, spectral feature reproduction, and functional connectivity restoration. This study establishes a more reliable and scalable new paradigm for non-invasive analysis of deep brain dynamics. The code of this study is available in https://github.com/hdy6438/NeuroFlowNet
Dongyi He, Bin Jiang, Kecheng Feng +5
School of Artificial Intelligence, Chongqing University of Technology, Chongqing, 400054, China · School of Smart Health, Chongqing Polytechnic University of Electronic Technology, Chongqing, 401331, China