Enhancing vector autoregressive transformer with memory convolutional network for explainable end-to-end seizure...

Accurate early prediction of epileptic seizures is critical for improving patient outcomes, yet it remains challenging due to significant patient heterogeneity, which often leads to inadequate feature representation and limited explainability in existing methods. To address these, this study proposes a novel memory convolutional network-enhanced vector autoregressive transformer (MemConvVAT) for end-to-end epileptic seizure prediction. The model primarily comprises two components: a vector autoregressive transformer (VAT) and a memory convolutional network (MemConv). Specifically, the VAT aligns a multi-layer transformer with vector autoregression (VAR) through reconstructed data flows, while incorporating path pruning and shortcut connections to maintain computational efficiency. Furthermore, it enhances the modeling of complex temporal dependencies by introducing temporal influence paths across layers, enabling dynamic aggregation of multi-step information and yielding potentially explainable, decomposable neural pathways. The MemConv, on the other hand, employs an external memory mechanism to dynamically store and retrieve input-relevant prototype features. This design overcomes the limited receptive field of conventional convolutional networks and empowers the model with global pattern matching and feature associative capabilities. Notably, by conducting a comparative analysis of feature evolution between preictal and interictal states, and combined with systematic feature ablation, this study provides a valuable perspective for improving the explainability of deep learning. In intra-patient experiments, MemConvVAT achieves over 90% accuracy on both CHB-MIT and SWEC-ETHZ datasets, with model performance comparable to current state-of-the-art methods across multiple key metrics. Our implementation is openly available at https://github.com/JW-Image/MemConvVAT.