The seizure embedding map: a spatio-temporal transformer for comparing patients by ictal intracranial EEG features at...

Objective.Planning invasive treatment for medication-resistant epilepsy depends on interpreting intracranial EEG (iEEG) recordings to identify seizure onset patterns and locations. Clinicians currently rely on multimodal data, experience and literature to recommend treatments. Referencing a new patient's seizures against past cases remains subjective because implant strategies, electrode placements, and seizure onset zones vary across individuals and centers. This study introduces a transformer-based framework to rigorize this process by embedding spatial and temporal information in iEEG recordings to categorize seizure networks and relate them to clinical outcomes across a large cohort of drug-resistant epilepsy patients. The ultimate goal is to quantitatively compare new patients' seizure characteristics with thousands of historical cases to guide treatment recommendations.Approach.We design a spatiotemporal transformer that uses convolutional tokenization and anatomical positional encoding to extract spatial-temporal features in multi-channel iEEG seizure onset data. The model accommodates variable electrode counts and implant types, including stereotactic EEG and electrocorticography implants. We validated the learned seizure embeddings through unsupervised clustering and a cross-validated multi-class logistic-regression model.Main Results.The model was applied to 882 seizures from 102 subjects with drug-resistant epilepsy. Unsupervised clustering reveals 74 cross-subject clusters, and a multi-class regression model with 10-fold cross-validation reveals significant clustering of onset patterns in embedding space (acc. = 0.8159). Seizures occurring closer in time exhibited greater embedding similarity (p<0.05). Although seizure clusters did not differentiate therapy or surgical outcome, they were significantly associated with the anatomical onset region and seizure classification.Significance.We propose a method for representing iEEG recordings of seizures with embeddings that contain spatial and temporal information. These embeddings enable comparison of common patterns in seizure onset. Future work will refine this framework by incorporating multimodal data, including imaging, semiology and patient history to move toward quantitative, evidence-based decision making for patients with drug-resistant epilepsy.