A Spatiotemporal Posterior Graph Convolutional Neural Network Based on Multihead Attention With Squeeze-and-Excitation Module for Patient-Specific Epileptic Seizure Prediction

Abstract

The activity of epileptic lesions is closely related to the functional connectivity patterns between channel locations of the brain. Previous studies have used feature extraction methods such as Pearson correlation coefficient (PCC), mutual information, and phase locking value (PLV) to calculate the functional connectivity for seizure prediction. However, these methods can only capture partial dynamical features of functional connections. Therefore, a novel spatiotemporal posterior graph convolutional neural network (WCE-MASE-PGCN) is proposed to comprehensively explore the spatiotemporal topological information of epileptic signals. The content includes: 1) the wavelet convolutional encoder (WCE) is used to extract the time-frequency characteristics of epileptic signals, thereby obtaining a low-dimensional, robust embedding representation; 2) the MASE are employed to construct the posterior graph of brain functional connections, which are adaptively adjusted through backpropagation algorithms; 3) the posterior graph convolutional neural network (PGCN) is utilized to fully extract the features of brain functional connections; and 4) experimental results conducted on publicly available CHB-MIT and Siena datasets demonstrate that posterior-based graph networks exhibit superior capabilities in feature extraction and interpretability. On the CHB-MIT dataset, the area under the curve (AUC), sensitivity (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${S}_{n}$ </tex-math></inline-formula>), and false prediction rate (FPR) were measured as 0.966, 97.9%, and 0.010/h, respectively. On the Siena dataset, the AUC, sensitivity, and FPR were reported as 0.961, 96.5%, and 0.032/h, respectively. Compared to state-of-the-art (SOTA) methods, our approach achieved a 1.9% improvement in AUC, a 5.6% increase in <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${S}_{n}$ </tex-math></inline-formula>, and a 7.2% reduction in FPR on the CHB-MIT dataset. On the Siena dataset, it improved AUC by 3.8%, increased <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${S}_{n}$ </tex-math></inline-formula> by 6.3%, and reduced FPR by 9.7%.

Topics & Concepts