CAAST-Net: Causality-Aware Spiking Transformer with Flask API for EEG-Based Seizure Detection

Abstract

Epileptic seizure detection from electroencephalogram (EEG) signals is an essential task for real-time neurological monitoring. Traditional models face challenges with interpretability, energy efficacy, and capturing temporal causality in neural data. To address these drawbacks, this manuscript proposes a Causality-Aware Attention with Spiking Transformer Network (CAAST-Net). The frequency-domain features are extracted by fast Fourier transform (FFT) to acquire band power across five canonical EEG bands. Subsequently, the features are normalized using a Standard Scalar and transformed into spike trains by rate coding, which enables biologically inspired processing. The CAAST-Net model includes a linear projection layer, leaky integrate-and-fire (LIF) neurons and causality-aware attention module that ensures temporal consistency by facilitating signals from past to present. This model learns discriminative and temporally predictive EEG patterns with minimized computational overhead. The whole model is deployed through a Flask API for real-time seizure detection. The proposed CAAST-Net obtains a higher accuracy of 98.95% with a CHB-MIT dataset and 99.85% accuracy with a BONN dataset when compared with traditional models.