Time-Series Anomaly Detection Based on Dynamic Temporal Graph Convolutional Network for Epilepsy Diagnosis

Abstract

Electroencephalography (EEG) is typical time-series data. Designing an automatic detection model for EEG is of great significance for disease diagnosis. For example, EEG stands as one of the most potent diagnostic tools for epilepsy detection. A myriad of studies have employed EEG to detect and classify epilepsy, yet these investigations harbor certain limitations. Firstly, most existing research concentrates on the labels of sliced EEG signals, neglecting epilepsy labels associated with each time step in the original EEG signal-what we term fine-grained labels. Secondly, a majority of these studies utilize static graphs to depict EEG's spatial characteristics, thereby disregarding the dynamic interplay among EEG channels. Consequently, the efficient nature of EEG structures may not be captured. In response to these challenges, we propose a novel seizure detection and classification framework-the dynamic temporal graph convolutional network (DTGCN). This method is specifically designed to model the interdependencies in temporal and spatial dimensions within EEG signals. The proposed DTGCN model includes a unique seizure attention layer conceived to capture the distribution and diffusion patterns of epilepsy. Additionally, the model incorporates a graph structure learning layer to represent the dynamically evolving graph structure inherent in the data. We rigorously evaluated the proposed DTGCN model using a substantial publicly available dataset, TUSZ, consisting of 5499 EEGs. The subsequent experimental results convincingly demonstrated that the DTGCN model outperformed the existing state-of-the-art methods in terms of efficiency and accuracy for both seizure detection and classification tasks.

Keywords: electroencephalography; graph convolutional network; seizure detection and classification; time-series anomaly detection.

Figure 1

Changes in EEG signals at different periods of seizure. The EEG signals in the blue box indicate the preictal period, the EEG signals in the red box indicate the ictal period, the EEG signals in between the two boxes indicate the postictal period, and the EEG signals elsewhere indicate the interictal period.

Figure 2

Framework of the DTGCN Model. X and $X^{\prime }$ respectively denote the input and reconstructed output. $V_t$ is a sequence of windows sliced from $X^{\prime }$ . $H_{t-1}$ represents the output of TGCN at the $(t-1)-th$ time step. $E_t$ represents the adjacency matrix of the graph structure generated at the $t-th$ time step. Gconv represents graph convolution. The upper right and lower right two different branches represent seizure detection and classification respectively.

Figure 3

Fine-grained and coarse-grained labels.

Figure 4

Confusion matrices of the baselines and DTGCN for 60 s seizure classification. The intensity of the color increases with the improvement of prediction accuracy, with darker hues corresponding to higher levels of accuracy.

Figure 5

The impact of parameter changes on model performance: (A) the window size K, (B) the time step n that needs to be averaged, and (C) the weight factor $\lambda$ of the KL divergence loss item.

Figure 6

Mean adjacency matrix for EEG in the test set for (A,F) non-seizure EEG clips, (B,G) generalized seizures, (C,H) focal seizures, (D,I) difference between focal seizure and non-seizure adjacency matrices, and (E,J) difference between generalized seizure and non-seizure adjacency matrices.