Title: EEG-FuseFormer: A Transformer-Driven Feature Fusion Framework for Seizure Onset Prediction

Abstract:

Epilepsy stands as one of the most prevalent neurological conditions worldwide, marked by recurrent seizures that severely diminish patients' quality of life. Although diagnostic methodologies have advanced, managing the risks associated with epilepsy remains difficult because seizure events are inherently unpredictable. Accurately forecasting the onset of seizures offers a critical pathway to mitigating these dangers. To address this challenge, this study introduces EEG-FuseFormer, a novel framework for predicting seizure onset that leverages a transformer-based approach to fuse features. This model integrates intermediate features derived from two distinct architectures: a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) network and a ResNet-18 network. The CNN-LSTM component is designed to capture both spatial and temporal characteristics directly from raw EEG signals, while ResNet-18 extracts features from the Short-Time Fourier Transform (STFT) representation of the same data. These disparate features are combined via a transformer encoder, with the ultimate prediction produced through fully connected dense layers.

The proposed model was validated using the CHB-MIT dataset, yielding a mean recall rate of 98.85% and demonstrating superior performance compared to many current state-of-the-art methods. Furthermore, the research investigates the model’s generalization capabilities in cross-patient testing environments. Within this framework, fine-tuning pre-trained models on limited data from target patients—a strategy known as target adaptation—yields improved recall, precision, and F1-score metrics when compared to standard cross-patient validation techniques. Lastly, the study assesses the model’s runtime computational complexity across various hardware platforms to illustrate the balance between performance and computational efficiency.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC