The authors assembled a large routine-EEG corpus to address a practical presurgical problem: localising the seizure onset zone without invasive monitoring. They curated 262 sessions—more than 6,000 hours of recording—from 100 patients with unilateral focal epilepsy. An internal 28-patient dataset and an external 9-patient dataset served as additional evaluation cohorts. The study team reported that sessions included both interictal and ictal segments and that clinical labels were expert-annotated; the preprint did not report further demographic breakdowns in the abstract.
The architecture integrates four explicit components. First, an online augmentation pipeline applied Gaussian noise, amplitude scaling, temporal cropping, and hemispheric channel swap to reduce overfitting and increase variation in routine data. Second, the network contains a multi-scale temporal block tuned to discharge morphologies spanning roughly 20–200 ms. Third, investigators implemented a factorised spatiotemporal Transformer that decouples temporal from spatial attention—an explicit design choice meant to mitigate the effects of volume conduction on scalp potentials. Fourth, the loss function included a hemispheric mutual exclusivity penalty encoding the clinical prior of unilateral seizure focus.
The study team trained a heterogeneous ensemble of five models and combined them via a model-soup fusion. They trained on the curated sessions with clinical labels assigned at session level; the authors framed the output as lobar localisation of the SOZ, intended to augment clinician review rather than replace it. The online augmentation methods were applied during training to widen the effective sample distribution and to emulate plausible recording variability seen in routine scalp EEG.
Investigators emphasised multi-scale temporal processing to capture brief discharge morphologies as well as longer ictal patterns. The factorised spatiotemporal Transformer separated attention over time and attention across electrodes, an approach chosen specifically to counteract spatial smearing attributable to volume conduction. The hemispheric mutual exclusivity penalty enforced unilateral localisation priors during optimisation.
On the internal held-out test set the ensemble yielded a session-level exact match of 88.46% according to the authors' report. By contrast, EEGNet achieved 50.00% on the same internal test data. Performance generalised to a seizure-free holdout cohort with 82.76% session-level exact match, and to the external cross-institutional dataset at 84.62%.
The authors emphasised that classification held up even when ictal events were not present, indicating that interictal features in routine scalp EEG carry useful localisation information. They positioned the method as a clinical triage tool: the model proposes localising windows and lobar-level predictions for targeted expert review rather than automated clinical decision-making.
Investigators reported gains for rare classes. Recall for the rare occipital class rose from 0.500 under baseline comparison to 1.000 with the ensemble. The model produced zero false positives when applied to a healthy control set (n=108), as stated in the abstract. For unlabelled, randomly sampled EEG segments the system achieved 66.67% accuracy; the authors contrasted that with a 14% chance baseline, indicating nonrandom localisation signal in routine segments.
The preprint did not specify in the abstract the detailed confusion matrices, full class distributions, or per-patient variability metrics; those elements may be reported in the full text but are not provided in the abstract summary.
Operational performance was a clear focus. Processing a full 24-hour recording required 4.26 minutes in the reported pipeline, and single-segment inference ran in under 8 seconds. These timings suggest the tool can rapidly triage long recordings to surface candidate localising windows for neurophysiologist review. The authors proposed that this efficiency could shorten the presurgical pathway by identifying informative periods for targeted expert inspection and potentially reducing the need for prolonged inpatient monitoring or invasive implantation.
The preprint explicitly framed the system as an adjunct: data indicated the model supplies spatially localising information to support, not supplant, clinical judgement in delineating the epileptogenic zone proxy represented by the SOZ.
The manuscript is a preprint and has not been peer reviewed; the authors declared no competing interests. The abstract documents cohort sizes and performance metrics but does not report detailed demographic or electrode montage variability in the summary. It also does not include, in the abstract, specifics on training hyperparameters, cross-validation folds, patient-level versus session-level confusion, or how clinical labels were adjudicated beyond expert annotation.
Key unanswered questions remain explicit in the source: how will the model perform across broader, more heterogeneous clinical populations; how sensitive are results to different EEG montages or recording systems; and how do per-patient variability and lesion heterogeneity affect localisation reliability? The preprint does not report those items in the abstract, so further details will require consultation of the full text and peer review. The next documented step in the source is external validation and integration into targeted clinical review workflows rather than immediate clinical deployment.