REST: Efficient and Accelerated EEG Seizure Analysis through Residual State Updates

TL;DR

This paper introduces a novel graph-based residual state update mechanism (REST) for real-time EEG signal analysis in applications such as epileptic seizure detection, and achieves a remarkable 9-fold acceleration in inference speed compared to state-of-the-art models.

Abstract

EEG-based seizure detection models face challenges in terms of inference speed and memory efficiency, limiting their real-time implementation in clinical devices. This paper introduces a novel graph-based residual state update mechanism (REST) for real-time EEG signal analysis in applications such as epileptic seizure detection. By leveraging a combination of graph neural networks and recurrent structures, REST efficiently captures both non-Euclidean geometry and temporal dependencies within EEG data. Our model demonstrates high accuracy in both seizure detection and classification tasks. Notably, REST achieves a remarkable 9-fold acceleration in inference speed compared to state-of-the-art models, while simultaneously demanding substantially less memory than the smallest model employed for this task. These attributes position REST as a promising candidate for real-time implementation in clinical devices, such as Responsive Neurostimulation or seizure alert systems.

Figures (8)

• Figure 1: (a) electrodes placement based on the 10/20 standard and its constructed distanced based graph. Self edges are not shown for better visualization. (b) The Rest framework, where raw signals undergo preprocessing and are structured as a graph before feeding as input to the model. Following multiple (or single) updates, the model provides the detection or classification result. (c) Single update mechanism of the proposed model. Dense represents the fully connected layer and GConv is the graph convolution. See our web page for more visual results at https://arshiaafzal.github.io/REST/.
• Figure 2: AUROC comparison among various models for seizure detection across different clip lengths on TUSZ dataset. A flatter line indicates more consistent performance, with error bars representing variation across five random seeds. Higher values on the y-axis correspond to increased accuracy. Rest(RM) is shown as bold green line to emphasise its stability.
• Figure 3: Performance comparison in seizure analysis across models on TUSZ dataset: a) Seizure detection AUROC vs. Model size. b) Seizure classification weighted F1-score vs. model size. c) Seizure detection AUROC vs. inference. d) Seizure classification weighted F1-score vs. inference. The $\bullet$s represents the accuracy on evaluation set for different train/validation splits and $\star$s represent the mean accuracy across different train/validation splits.
• Figure 4: ROC curves for different clip lengths among Rest and baselines for TUSZ dataset.
• Figure 5: Confusion Matrices for seizure classification task among different models.