EEG source imaging works backward from scalp voltages to estimate where on the cortex the electrical activity came from. Clinicians increasingly use it to decide which piece of brain to remove in epilepsy surgery, yet it has never been checked against what that tissue looks like under a microscope. Current validation compares the source map with the outline of the surgical cavity and asks whether the patient became seizure-free. That tells us whether the map pointed roughly at the right lobe. It cannot tell us whether a "hub" in the reconstructed network sits on diseased tissue (malformed cortical layers, scarring by support cells, activated immune cells) or is a blur produced by currents spreading through the skull. This proposal does the missing check. Resected tissue is cut into spatially registered blocks, the cellular pathology of each block is quantified, and we test whether source-space metrics predict that pathology block by block, beyond what the resection outline, PET imaging, or plain scalp EEG already predict.
The second ingredient addresses a specific blind spot. Source methods are tuned and judged on focal lesions, but inflammatory pathology is often patchy and multifocal, so we do not know whether it is localized as well. The relaxation of pandemic restrictions created an unusual opportunity. Accumulated population susceptibility, called "immunity debt", produced synchronized surges of respiratory infection in 2023, and infection can precede autoimmune encephalitis and later epilepsy. We use this surge only as a way to enrich the cohort for infection-preceded, inflammation-dominant cases, and as a pre-registered time-series test. It is not used to infer anything about individual patients.
Core Claim
Source-space network metrics computed from routine scalp EEG carry information about cellular pathology at the scale of millimetre tissue blocks, and that information exceeds what follows from distance to the resection centre or from PET hypometabolism. We expect this information to be weaker for inflammatory pathology than for focal dysplasia, and we expect geometry-constrained inverse methods to narrow that gap. If the first part fails, source imaging is a lobe-level pointer and should be validated and reported that way.
What Exists and What Does Not
The closest work is A network-based EEG source imaging framework for noninvasive localization of epileptogenic zones in MRI-negative focal drug-resistant epilepsy. It reconstructs sources with sLORETA, builds directed delta-band networks from the ten minutes before seizures, and scores nodes with centrality measures and a Seizure Index. In 15 patients it validates these scores against resection overlap, Engel Class I outcome, stereotactic EEG (SEEG, electrodes implanted in the brain), and FDG-PET. It never examines tissue. That study supplies our strongest method and our cohort template, and we replace its ground truth.
Advancing EEG/MEG Source Imaging with Geometric-Informed Basis Functions supplies a second inverse method. It constrains solutions to the patient's own cortical folding by using shape modes of the cortical surface. It beats MNE, dSPM and sLORETA on synthetic data down to −20 dB signal-to-noise ratio, and on visual and somatosensory tasks. Its real-data validation is "activity lands in the expected functional area", which is again not tissue.
The design for tissue validation comes from Nuclear Overhauser Enhancement Imaging of Glioblastoma at 7 Tesla. That study co-registered 12 stereotactic biopsies per patient to MRI and quantified cell density automatically. It also found that a widely trusted proxy, diffusion MRI, correlated inconsistently with cell counts while the alternative contrast reached r = 0.69 and 0.87. This both motivates our study and warns us: a clinically accepted map can fail outright once it is checked against cells. That study had only 3 biopsied patients, which tells us sample size will be the binding constraint.
The infection-to-brain link is motivated by three papers. Specific viral antibodies associate with anti-NMDAR encephalitis after herpes simplex encephalitis ties viral infection to later autoimmune encephalitis. Acute Symptomatic Seizures and Risk of Epilepsy in Autoimmune Encephalitis ties that encephalitis to later epilepsy. Characterising seizures in anti-NMDA-receptor encephalitis with dynamic causal modelling models seizure circuitry in encephalitis with lumped models rather than cortical source maps. The surge itself is documented by Sun et al. in The circulating characteristics of common respiratory pathogens in Ningbo, China, both before and following the cessation of COVID-19 containment measures, which reports pathogen prevalence up more than 50% and coinfections tripling after zero-COVID ended. It is also documented in the Beijing 2023 pediatric surge, the Zhejiang RSV and influenza burden study, and the EV-D68 severe pneumonia cohort. None of these papers connects surge timing to neurological tissue.
Research Program
Cohort. We will enroll patients undergoing resective epilepsy surgery or diagnostic brain biopsy at centres in the regions with documented surges. Each patient will have pre-surgical high-density scalp EEG, 3T MRI, and FDG-PET, with SEEG where it is clinically indicated, following the Soochow protocol. We record history of preceding infection, pathogen PCR where available, neural autoantibodies, and antiviral serology. Our target is about 40 patients. That is roughly ten times the biopsied subset in Paech et al. and still small, so all inference uses mixed-effects models with patient as a random effect.
Tissue registration. Surgeons will mark resected specimens using neuronavigation, so that each specimen is cut into roughly 8–15 oriented blocks with recorded coordinates. Each block's position is then reconciled with the postoperative MRI cavity and mapped onto the patient's cortical mesh. We stain each block for four markers, with automated density quantification as in Paech et al.:
- NeuN, to assess the neuronal layering that is disrupted in cortical dysplasia;
- GFAP, for astrocytic scarring (gliosis);
- Iba1, for microglia;
- CD8, for cytotoxic T cells.
Because brain shift is the dominant error, we estimate registration error per patient. All correlations are recomputed after smoothing the pathology maps at that error scale, so that a result cannot rest on sub-registration precision.
Source metrics. For each mesh vertex we compute the Soochow pipeline's outputs on the same pre-seizure delta window: sLORETA power, the network centralities (degree, out-degree, in-degree), and the Seizure Index. We compute the same metrics with geometry-constrained basis functions. Each method is run with an individual boundary-element head model and with a template head. The template comparison matters because Longitudinal multitask wireless electrocorticography data from two fully implanted nonhuman primates shows that template anatomy shifts localization systematically. For each block, the metric is averaged over the vertices it covers.
Analysis. The primary model predicts each block's pathology score from its source metric. It includes covariates for distance to the resection centroid, the block's PET hypometabolism, and the patient. The primary statistic is the coefficient on the source metric and the gain in out-of-patient AUC (area under the ROC curve) for classifying blocks as pathological versus histologically normal. A secondary interaction tests whether that gain differs between dysplasia-dominant and inflammation-dominant patients. Patients are labelled on histology alone, before anyone looks at the EEG.
Epidemiological arm. We run an interrupted time series of monthly referrals with new-onset focal seizures or encephalitis, together with the fraction of those referrals preceded by documented infection. These are modelled against regional pathogen surveillance curves at pre-registered lags. This arm answers a narrower question: did the surge measurably enrich the inflammatory stratum we need?
Baselines
The naive reference assigns every block inside the resection the same score, which is how resection-overlap validation implicitly treats tissue. A second naive reference scores blocks only by distance from the resection centroid. Any source metric must beat both on block-level AUC under leave-one-patient-out evaluation.
The sensor-level baseline is computed on the same EEG window and requires no inverse solution. It takes delta power and spike rate at the scalp electrode nearest each block and projects them radially onto the cortex. If this matches the source metrics, the inverse step adds nothing at tissue scale.
The imaging baseline is FDG-PET hypometabolism per block, taken from scans the cohort already has. It is the strongest non-electrophysiological competitor and enters every model as a covariate.
The strongest existing method is the sLORETA plus directed-network Seizure Index from the Soochow study. Against resection and outcome it reported sensitivity 60%, specificity 80%, and negative likelihood ratio 0.5. No tissue-level result exists for it. We rerun it exactly as published, with the same frequency band and window, as the reference implementation. Geometry-constrained basis functions are the second strong method. No reported result found on Lacuna for them on clinical epilepsy data, and their real-data results on VEPCON and OSE are qualitative.
Beating these methods means one of two things. Either a source metric adds a significant positive coefficient once distance and PET are in the model, or it improves block AUC by a pre-registered margin of 0.05 over the Seizure Index computed with a template head.
We include three ablations. Template versus individual head model isolates the contribution of anatomy. sLORETA versus geometry-constrained bases isolates the inverse prior. Shuffling block labels within each patient gives a null that destroys spatial information but preserves the mix of pathology per patient.
SEEG serves as an invasive upper reference rather than a competitor. It is available only in a subset of patients.
Why not just…?
Why not just keep Engel outcome and resection overlap as ground truth? These measures cannot distinguish a map that sits on diseased tissue from one that falls somewhere inside a generous resection. The Soochow study's 60% sensitivity is measured against that coarse target. The Paech et al. result shows that an accepted imaging proxy can look adequate at coarse scale and fail at the cellular scale. The "uniform score inside resection" baseline directly measures what outcome-based validation would have credited.
Why not just read scalp EEG and skip the inverse problem? Volume conduction blurs scalp signals. Cortical Classification with Rhythm Entropy for Error Processing in Cocktail Party Environment shows that this blurring distorts sensor-level metrics. This is a real empirical question at block scale, though, and the radial-projection baseline settles it. If that baseline ties the source metrics, we report it.
Why not just use PET or MRI? The target population is often MRI-negative, which is the premise of the Soochow cohort. PET is a serious competitor and is included as a covariate, so the claim is only made if source imaging adds information beyond it.
Why not just measure each patient's infection history with serology instead of modelling immunity debt? This is correct, and it is why the proposal changed. Individual exposure is measured directly with serology, PCR, and autoantibodies, and population susceptibility models such as Analyzing Social Distancing and Seasonality of COVID-19 with Mean Field Evolutionary Dynamics or Modeling COVID-19 Dynamics in German States Using Physics-Informed Neural Networks are not used to label patients. Immunity debt enters only as a source of cohort enrichment and as a time-series test of that enrichment.
Decisive Checks
The central test comes first. Under leave-one-patient-out evaluation, and with distance and PET in the model, the best source metric must predict block-level gliosis, dyslamination, or microglial density significantly, and the result must survive smoothing at the registration-error scale. If it fails, we conclude that source imaging is reliable only at lobe scale.
Second, we compare accuracy between inflammatory and dysplastic patients. We predict lower accuracy for inflammatory patients with sLORETA's smooth solutions and a smaller gap with geometry-constrained bases. A finding of no gap would also be informative.
Third, the time series must show that infection-preceded referrals rose after the 2023 surge, at a lag consistent with the encephalitis literature. If not, the inflammatory stratum is accrued without enrichment and the cohort timeline lengthens.
A known risk is that drug-resistant epilepsy surgery usually comes years after onset. Inflammatory cases within the study window will therefore come mostly from rapidly progressing pediatric cases and from diagnostic biopsies. We will report power for the inflammatory interaction separately and treat it as exploratory if fewer than ten such patients accrue.
References
- A network-based EEG source imaging framework for noninvasive localization of epileptogenic zones in MRI-negative focal drug-resistant epilepsy
- Advancing EEG/MEG Source Imaging with Geometric-Informed Basis Functions
- Nuclear Overhauser Enhancement Imaging of Glioblastoma at 7 Tesla: Region Specific Correlation with Apparent Diffusion Coefficient and Histology
- Specific viral antibodies associate with anti-NMDAR encephalitis after herpes simplex encephalitis
- Acute Symptomatic Seizures and Risk of Epilepsy in Autoimmune Encephalitis: A Retrospective Cohort Study
- Characterising seizures in anti-NMDA-receptor encephalitis with dynamic causal modelling
- The circulating characteristics of common respiratory pathogens in Ningbo, China, both before and following the cessation of COVID-19 containment measures
- The circulating characteristics of common respiratory pathogens in Ningbo, China, both before and following the cessation of COVID-19 containment measures
- Increase of respiratory illnesses among children in Beijing, China, during the autumn and winter of 2023
- The epidemiology and burden of respiratory syncytial virus and influenza infections in hospitalized children under 5 years old in Zhejiang, China (2018–2023)
- Subclade B3 and D3 enterovirus D68 among children with severe community-acquired pneumonia in China, 2021 to 2024
- Longitudinal multitask wireless electrocorticography data from two fully implanted nonhuman primates
- Cortical Classification with Rhythm Entropy for Error Processing in Cocktail Party Environment Based on Scalp EEG Recording
- Analyzing Social Distancing and Seasonality of COVID-19 with Mean Field Evolutionary Dynamics
- Modeling COVID-19 Dynamics in German States Using Physics-Informed Neural Networks