Propagation mapping using iterative independent component analysis for seizure onset zone localization in temporal...

Epilepsy affects approximately 70 million people worldwide, with a third of them being drug-resistant and requiring surgical intervention. Accurate localization of the seizure onset zone (SOZ) is crucial for effective surgery but remains challenging.

We proposed a method using iterative independent component analysis (ICA) to map seizure propagation of drug-resistant temporal lobe epilepsy (TLE). For each assumed seizure origin, ICA was applied to the remaining contacts to identify propagation components, with the highest correlating component being the propagating signal. Iterative removal of nearby contacts revealed spatial propagation profiles. Machine learning models were applied to the propagation profiles to distinguish the SOZ.

Seizure propagation features differed significantly between SOZ and non-SOZ contacts in seizure free patients (both local cohort N = 21, and independent dataset N = 13), but not in non-seizure free patients (N = 11). Propagation-based classifiers achieved robust performance (AUC = 0.85), outperforming iEEG source imaging (AUC = 0.73). In mesial TLE, propagation maintained high accuracy (AUC = 0.84) while iEEG source imaging dropped markedly (AUC = 0.64).

Traditional methods for SOZ localization, such as visual inspection of SEEG and source imaging techniques, rely heavily on expert interpretation. iEEG source imaging assumes linear forward models and can be susceptible to inaccuracies due to electrode placement and noise. In contrast, our proposed iterative ICA approach is purely data-driven and adaptively identifies the dominant propagation pathways across individual seizures.

This work introduces a data-driven strategy to characterize seizure propagation, potentially improving SOZ localization with deep brain origins.