Frequency Band Personalization for Seizure Network Analysis in Multifocal Patients

Stereo-electroencephalography (SEEG) is commonly used for pre-surgical evaluation in patients with multifocal epilepsy undergoing responsive neurostimulation (RNS). Seizure network modeling, a data-driven approach to assist RNS implantation and adjustment, benefits from individualized ictal signatures learned from SEEG data, including repetitive sharp waves, rhythmic discharges, and prominent frequency bands in seizure activities. This is all in pursuit of the earliest possible detection so that the RNS therapy can be delivered as soon as a potential seizure is on the horizon. To improve individualized treatment, this paper proposes a data-driven methodology for personalizing frequency-band selection using a SEEG-based seizure network in patients with multifocal epilepsy. A directed seizure network is first proposed using SEEG data, where network nodes are selected SEEG points, and spectral edges are characterized by a directed transfer function. To further enhance the robustness of network modeling, surrogate data analysis is applied to ensure the significance of network estimation. Then, the density differences between the seizure onset zone and remaining regions are extracted using subgraph density as a heuristic to select the discriminative frequency range with the highest difference as the optimal band. The selection result is further validated using a machine learning classifier for seizure detection. Across ten multifocal patients, our personalized selection achieved an Area-Under score of 0.94 out of 1 in the seizure classification task, outperforming standard bands in distinguishing ictal from pre-ictal states, and exhibiting type-specific spectral patterns with consistency through different frequency-domain network metrics.