Automated annotation of low-frequency stimulation-induced seizures uncovers seizure generating networks

Evaluating epilepsy patients for invasive treatment currently requires recording spontaneous seizures—a process that is costly, time-consuming, and may falsely localize surgical targets. Electrical stimulation can be used to induce seizures, offering a potentially faster solution, but it is unclear if induced seizures localize seizure generators. The lack of reliable, quantitative seizure mapping tools has hampered answering this question at sufficient scale and precision to advance this approach.

We retrospectively analyzed 441 seizures from 104 patients across two epilepsy centers who underwent intracranial EEG and low-frequency electrical stimulation. Using a novel deep-learning algorithm, we quantified spatial spread and electrographic similarity between spontaneous and stimulation-induced seizures (stim seizures) and their relationship to clinical semiology and anatomical boundaries.

The unsupervised algorithm annotated seizures with expert-level precision. Stim seizures that reproduced patients’ habitual clinical symptoms began in the same regions as spontaneous seizures and were associated with seizure freedom after surgery. In contrast, stim seizures with atypical semiology began in regions rapidly recruited during spontaneous seizures—an established biomarker of secondary generators—and were associated with post-operative seizures. Low-frequency stimulation most often induced seizures in mesial temporal structures, particularly in adult-onset epilepsy.

This study supports using low-frequency electrical stimulation routinely during invasive evaluation for epilepsy surgery to target destructive and neurostimulation therapy. Stim seizures can rapidly map seizure generating tissue, detect mesial temporal involvement, and establish the extent of secondary seizure generators that may be missed by studying spontaneous seizure onsets alone. This study supports a shift in epilepsy surgery from passive intracranial recording to stimulation-induced mapping.