The success of resection therapies in patients with drug-refractory epilepsy relies on accurately identifying the epileptogenic zone (EZ), which comprises the specific brain regions contributing to seizure generation. The conventional methodology for EZ identification by clinicians is subjective, time-consuming, and lacks integration of the complex temporospatial dynamic processes involved. Development of an automated network-based EZ detection framework has the potential to not only improve surgery outcomes but also to advance next-generation neurostimulation therapies. Currently, most existing research is constrained to simple linear network analysis or focuses on modeling only one aspect of these complex connections. Our study aims to enhance EZ identification by utilizing various features derived from non-linear, multi-layer graphical models of the epileptogenic network, providing a more holistic understanding of the condition.

We utilize a dataset including patients with intractable mesial temporal lobe epilepsy (MTLE) who have undergone resective or ablative surgeries following stereo-electroencephalography (sEEG) investigation and remained seizure-free for more than 1-year post-surgery (Engel I outcome). We create three distinct graphical representations that capture the nonlinear spectral and temporal connectivity patterns across pre-ictal, ictal, and inter-ictal intervals. Various features extracted from this multi-layer graph structure are then leveraged by machine learning algorithms tasked with detecting the EZ within the modeled graphs. This is formulated as a binary node classification problem, where nodes correspond to the recording channels. Within our framework, these nodes are classified as either EZ or non-EZ. Finally, spatial attributes of the electrodes are incorporated through a graph convolutional filtering process to fine-tune the predicted labels.

 Our results demonstrate the effectiveness of our approach, achieving an F1-score (balance between precision and recall) above 79% and AUC exceeding 94% on the hold-out dataset. Additionally, we observe a significant decrease in performance when excluding the multi-layer based features or the spatial filtering step, highlighting the importance of a multi-perspective representation of the epileptogenic network and the effectiveness of incorporating the electrodes’ spatial properties in detecting the EZ.

Figure 1: A sample multi-layer graph structure generated using recordings from one patient.

Figure 2: Example output of our model. The colormap on the nodes represents the predicted probability of each node being part of the epileptogenic zone (EZ). The red boxes indicate the clinically identified EZ. This visualization highlights the model’s high accuracy in detecting the true EZ (Image created with help of Jasper Wang).