Long-Term EEG Partitioning for Seizure Onset Detection

TL;DR

This paper tackles the lack of explicit seizure onset (SO) localization in EEG-based detection by reframing SO detection as a subsequence clustering problem. The authors propose SODor, a two-stage framework that first learns interpretable, channel-wise logits via a graph-based second-level representation, then enforces long-term temporal consistency through a Toeplitz graphical-lasso clustering model to identify SO transitions. The approach yields state-of-the-art performance on CHB-MIT, HUH, and TUH datasets, with notable improvements over post-processing and classification baselines and provides interpretable insights into channel interactions during seizures. Overall, SODor offers a principled, scalable method for explicit SO detection with potential clinical utility in localization and neuromodulation planning.

Abstract

Deep learning models have recently shown great success in classifying epileptic patients using EEG recordings. Unfortunately, classification-based methods lack a sound mechanism to detect the onset of seizure events. In this work, we propose a two-stage framework, SODor, that explicitly models seizure onset through a novel task formulation of subsequence clustering. Given an EEG sequence, the framework first learns a set of second-level embeddings with label supervision. It then employs model-based clustering to explicitly capture long-term temporal dependencies in EEG sequences and identify meaningful subsequences. Epochs within a subsequence share a common cluster assignment (normal or seizure), with cluster or state transitions representing successful onset detections. Extensive experiments on three datasets demonstrate that our method can correct misclassifications, achieving 5\%-11\% classification improvements over other baselines and accurately detecting seizure onsets.

Figures (6)

• Figure 1: A visualization of 55 two-second epochs using an second-level classification method shows some unexpected abrupt misclassifications. This issue may lead to unexplainable outcomes for clinicians.
• Figure 2: System overview: SODor is a two-stage framework designed to explicitly detect seizure onset, consisting of a classification model ($\mathcal{F}(\cdot)$) and a subsequence clustering model ($\mathcal{H}(\cdot)$). $\mathcal{F}(\cdot)$ learns second-level correlation representations for channel-wise logits in normal and seizure states through supervised learning. $\mathcal{H}(\cdot)$ then clusters a sequence of second-level epochs into subsequences. Epochs in a subsequence are consecutive, and each subsequence is dependent on its neighbors. A transition between subsequence/cluster assignments can be viewed as a seizure onset.
• Figure 3: Comparison of (a) SO detection and (b) SPO (Seizure Preictal Onset) detection. The figure illustrates the performance of the classification model (GNN_ICLR22), w/ post-processing methods (PostProcess2), and w/ our proposed subsequence clustering approach. The mismatch between SODor and ground truth is marked.
• Figure 4: Ablation of pooling methods (i.e., max, mean, and weighted) on channel logits across different training settings. The results show the robustness of max pooling. We visualize logits in max pooling under different parameter settings, and it maintains consistent performance.
• Figure 5: Visualization of learned $\mathbf{\Theta}$. The connections become denser from normal to seizure. The green box highlights consistent connections, while the red circle indicates differences between the two clusters.

Theorems & Definitions (5)

• Definition 1: Multi-channel EEG recordings
• Definition 2: Subsequence clustering
• Proposition 1
• proof
• proof