Seizure forecasting models read many recording channels and output one alarm: "seizure coming" or "not." Two findings suggest this hides real structure. Distant cortex, 2–3 cm from where seizures start, carries predictive signal up to an hour ahead. Separately, a specific signature appears only in the seizure onset zone, and only minutes before onset: a slow brain rhythm begins to modulate the strength of fast activity. Taken together, these suggest the warning sign may move over time, from a broad network early to a focal hotspot late.
We propose to test this directly by restricting the model's view to anatomically defined rings around the onset zone. The model must then say how far away a seizure is, as an ordinal stage, rather than only whether one is coming. If the best ring changes with the stage, the result fixes a design question that medical imaging answered with masks, the tool that keeps models from relying on irrelevant background. Seizure forecasting has never asked this question, and here the answer would depend on time.
Core Claim
Our hypothesis is that the spatial scope that maximizes preictal information depends on time-to-seizure. Early stages are best read from wide, distal cortex. Late stages are best read from a tight onset-zone region.
The method follows from this. A forecaster pools features through fixed masks that are defined anatomically and independently of the forecast labels. It uses those masks to predict an ordinal preictal stage. We expect it to match or beat channel-attention forecasters on standard binary metrics and to generalize better to unseen patients.
The claim is falsifiable. If the information-versus-radius curve has the same shape at every stage, the hypothesis is wrong. In that case, the masking argument from imaging does not transfer to this problem.
Grounding in Prior Work
Evidence that the signal moves. Three results motivate the hypothesis.
- Intracortical neural activity distal to seizure-onset-areas predicts human focal seizures shows that microelectrode recordings far from the onset area discriminate the hour before a seizure at around 90% ROC-AUC. Early information is therefore not focal.
- Work by Hashimoto and colleagues finds that coupling between the slowest rhythms (<0.1 Hz) and high-frequency oscillations (>80 Hz) emerges specifically in the onset zone minutes before onset. Late information is therefore focal. - A network-based EEG source imaging framework reports two things. It finds delta-band power rising in the ten minutes before onset. It also reconstructs cortical sources from scalp EEG and builds directed networks between cortical regions.
None of these studies asks how predictive information is distributed across space as a function of time.
Where the method comes from. The masking tools come from imaging.
- The Role of Pleura and Adipose in Lung Ultrasound AI found that unmasked models relied on fat thickness. Masking to the pleura raised AUC from 0.6553 to 0.7742.
- Mask-weighted global average pooling for CT triage supplies the exact pooling operation we reuse.
These masks come from anatomy, not from the diagnostic label. This is the property that separates them from the channel attention used in MB-dMGC-CWTFFNet and STAN. Both of those learn channel weights from the forecasting objective, so the weights can latch onto artifacts.
What constrains the design. The lung-ultrasound paper also warns that upstream segmentation errors cannot be recovered downstream. The source-imaging paper shows how unreliable non-invasive localization is, with Seizure Index sensitivity of 60% and specificity of 80%. So our scalp masks must be soft and must be stress-tested.
How granularity enters. We borrow the soft transitional labeling of Shorter latency of real-time epileptic seizure detection. That work applied it to ictal detection, the first seconds of a seizure. We apply it to minutes-to-hours preictal staging instead.
The patient-wise collapse in seizure-type classification is a further warning. The seizure type benchmark reports weighted F1 falling from 0.901 to 0.561, and SeizureNet reports a fall from 0.94 to 0.59. Granular claims must therefore be tested on held-out patients.
Research Program
Defining the onset zone. For each patient, we define an onset region using only ictal onsets from training seizures. Forecast labels never touch the mask.
- In intracranial data, the onset region is the set of electrodes showing earliest ictal change. Concentric rings are then defined by Euclidean distance from those contacts where coordinates exist, or otherwise by clinical focal and extrafocal annotation.
- In scalp data, we follow the source-imaging pipeline above. We use sLORETA on a template head to estimate activity in cortical parcels. The onset region is then soft-weighted by the source-imaging localization of training-seizure onsets, and rings are geodesic shells on the cortex.
Computing ring features. Within each parcel or contact, we compute a fixed set of dynamical features:
- delta and other band powers;
- permutation entropy, following Epileptic Seizure Prediction Based on Permutation Entropy;
- in intracranial data only, phase-amplitude coupling between slow and high-frequency activity.
Features are pooled per ring $r$ as
$$h_r = \frac{\sum_i m_{r,i}\, x_i}{\sum_i m_{r,i}},$$
where $x_i$ is the feature vector of parcel or contact $i$, and $m_{r,i}\in[0,1]$ is its soft membership in ring $r$. We add directed connectivity between rings, using the Directed Transfer Function as in the source-imaging work.
Predicting the stage. An ordinal head with cumulative-link outputs predicts one of five stages:
- interictal;
- 60–30 minutes before onset;
- 30–10 minutes;
- 10–2 minutes;
- under 2 minutes.
Labels are softened at stage boundaries.
The scope sweep. This is the central analysis. Using the same feature extractor, we train a separate stage classifier per ring radius, from focal-only to whole-brain. We record per-stage information, measured as one-vs-rest AUC, as a function of radius. We then test the stage-by-radius interaction with a mixed-effects model across patients. The hypothesis predicts that the optimal radius shrinks as stage advances.
The sweep uses simple features on purpose. The scientific claim should then not depend on a deep architecture.
Scope of the project. Forecasting the type of the coming seizure (absence versus tonic-clonic) is deliberately left out. Section 4 of the research file notes that no study attempts it. TUSZ, the only typed corpus in the file, consists of ictal-centered clinical recordings with little preictal data. Its patient-wise ceiling of about 0.59 F1 suggests the attempt would be underpowered.
Baselines
Naive reference points. These set the floor:
- a predictor that always outputs the stage prior;
- a random alarm generator matched in alarm rate, which keeps sensitivity and false predictions per hour honest;
- the permutation-entropy model with an RBF support vector machine and two-window Firing Power smoothing. On Freiburg intracranial data it reports 94% average sensitivity, 0.111 false predictions per hour, and a 61.93-minute horizon (Epileptic Seizure Prediction Based on Permutation Entropy).
We rerun that pipeline on all channels with our splits. Our model should exceed its sensitivity at equal or lower false-prediction rate. It should also deliver stage resolution, which the baseline cannot.
Strongest existing method. This is MB-dMGC-CWTFFNet (Dynamic Multi-Graph Convolution-Based Channel-Weighted Transformer Feature Fusion Network). It learns evolving channel graphs and channel weights. On CHB-MIT it reports AUC 0.935, 97.8% sensitivity, and 0.059 false predictions per hour under leave-one-out evaluation.
STAN is the second deep competitor. It reports 96.6% sensitivity and 0.011 per hour on CHB-MIT, with an adaptive 15–45-minute warning window.
We retrain both under identical splits, preprocessing, and hyperparameter search budget. We add an ordinal head so they can be scored on staging. Beating them means one of two outcomes:
- higher ordinal agreement (quadratic-weighted kappa) at matched binary sensitivity and false-prediction rate, under patient-specific evaluation; or
- a significant binary-AUC advantage under leave-one-patient-out evaluation, where we expect learned attention to overfit patient-specific artifacts.
Ablations isolating the key ingredient. Each ablation targets one assumption:
- A single global mask removes ring structure.
- A learned attention mask replaces the anatomical rings, with the same pooling and parameter count. This tests whether label-independent anatomy matters.
- Size-matched rings centered on a random non-onset region control for mere dimensionality reduction.
- Hard rings replace soft ones, to measure sensitivity to localization error.
- A binary preictal target and a continuous time-to-seizure regression target replace the ordinal stages.
Why not just…?
Why not just let channel attention learn the scope, as MB-dMGC-CWTFFNet and STAN already do? This is the real threat. Attention conditioned on time could in principle rediscover a radius that shrinks with stage. The learned-attention ablation settles it. If learned masks match anatomical rings within noise under leave-one-patient-out evaluation, we drop the modeling claim and report only the scope finding. What would favor anatomy is evidence that learned weights track artifacts. The MB-dMGC paper itself flags this risk for scalp EEG, and the cross-patient instability in Convolutional Neural Networks for Epileptic Seizure Prediction points the same way: AUC fell to about 0.25 for one human patient.
Why not just use permutation entropy and a support vector machine with good channel selection? It is cheap and strong, and it is our naive baseline. The scope sweep itself is built from such features. If permutation entropy on hand-picked onset channels alone matches the ring model across all stages, distal context adds nothing. That would refute the hypothesis in a useful way. The distal-cortex result suggests otherwise for early stages, but that study had 5 patients and 11 seizures, so the question is open.
Why not just clean the scalp signal with independent component analysis and keep sensor space? Artifact removal addresses noise, not scope. It cannot tell the model that distal cortex matters early and focal cortex late. We include an ICA-cleaned sensor-space version of the strongest baseline. If it closes the gap on scalp data, masking's benefit there is only denoising, and the intracranial results must carry the scope claim.
Why not just predict binary preictal state, or regress time-to-seizure? Binary prediction cannot reveal a time-dependent scope. The hypothesis is about stages, so stage-resolved evaluation is required. Continuous regression is a fair alternative and is included as an ablation. The ordinal head is preferred only if it is better calibrated near stage boundaries, where soft labels help.
Proposed Experiments
Datasets. We use three corpora:
- the CHB-MIT scalp corpus (23 pediatric patients, long continuous recordings), analyzed in source space;
- the Freiburg intracranial dataset (19 patients, 83 seizures);
- SWEC-ETHZ intracranial data (11 patients, 512 Hz), which permits coupling with activity above 80 Hz.
Stage boundaries require at least 60 uninterrupted preictal minutes. Seizures lacking this contribute only to their available stages.
Evaluation regimes. We report both:
- patient-specific evaluation, holding out the latest seizure per patient, to compare against published numbers;
- leave-one-patient-out evaluation, given the documented granular collapse.
Metrics. We report:
- quadratic-weighted kappa and per-stage AUC for staging;
- sensitivity, false predictions per hour, and warning horizon after Firing Power smoothing;
- calibration error near stage boundaries.
Decisive checks. Three checks determine the outcome. First, the stage-by-radius interaction in the scope sweep must be significant and correctly signed on intracranial data, where localization is trustworthy. If it is absent, the project's central hypothesis fails. Second, the anatomical-ring model must beat its learned-attention and random-ring twins under leave-one-patient-out evaluation. Third, on scalp data, the ring model's advantage must survive the localization stress test, in which onset-region weights are perturbed by the observed source-imaging error. Surviving that test is what licenses masked forecasting without implanted electrodes.
References
- Intracortical neural activity distal to seizure-onset-areas predicts human focal seizures
- Coupling between infraslow activities and high‐frequency oscillations precedes seizure onset
- A network-based EEG source imaging framework for noninvasive localization of epileptogenic zones in MRI-negative focal drug-resistant epilepsy
- The Role of Pleura and Adipose in Lung Ultrasound AI
- Computed tomography–based COVID–19 triage through a deep neural network using mask–weighted global average pooling
- Dynamic Multi-Graph Convolution-Based Channel-Weighted Transformer Feature Fusion Network for Epileptic Seizure Prediction
- Spatio-Temporal Attention Network for Epileptic Seizure Prediction
- Shorter latency of real-time epileptic seizure detection via probabilistic prediction
- Machine Learning for Seizure Type Classification: Setting the benchmark
- SeizureNet: Multi-Spectral Deep Feature Learning for Seizure Type Classification
- Epileptic Seizure Prediction Based on Permutation Entropy
- Convolutional Neural Networks for Epileptic Seizure Prediction