TFFBN-HDLF: a hybrid deep learning framework based on time-frequency functional brain networks for epileptic seizure detection

TL;DR

This paper introduces a new deep learning framework for detecting epileptic seizures in elderly patients using EEG data, improving accuracy and adaptability.

Contribution

The novel TFFBN-HDLF framework combines time-frequency functional brain networks with a hybrid CNN-Transformer architecture for seizure detection.

Findings

TFFBN-HDLF achieved 98.09% accuracy and 99.45% AUC on the CHB-MIT dataset.

The framework showed 92.49% accuracy and 95.64% AUC on the Siena dataset.

The hybrid model effectively captures multi-scale spatiotemporal features for seizure detection in elderly patients.

Abstract

The detection of epilepsy seizures in the elderly based on electroencephalogram (EEG) is the foundation of an intelligent clinical decision support system. However, due to the often slow background activity and complex non-stationary dynamic characteristics of the brain signals in elderly patients, existing methods often struggle to extract robust discriminative features across different individuals. To address this deficiency, this study proposes a hybrid deep learning framework named TFFBN-HDLF, aiming to enhance the reliability and diagnostic accuracy of artificial intelligence-assisted monitoring of epilepsy seizures in the elderly.

Firstly, this paper presents a time-frequency functional brain network construction method (TFFBNC). By combining the Pearson correlation coefficient (PCC) and phase lag index (PLV), we construct a two-dimensional time-frequency fused functional brain network (TFPPNet). This method can comprehensively simulate the synchronous neural interactions in the time and frequency domains of the elderly brain, converting the complex raw EEG data into high-quality neurophysiological evidence, thereby providing a basis for clinical decision-making. Additionally, we have developed a hybrid deep learning architecture-SeizureTransNet, which combines convolutional neural networks (CNNs) with enhanced Transformer modules. This architecture can dynamically select and integrate multi-scale spatiotemporal features, ensuring accurate inference of the seizure state in the elderly while maintaining high adaptability to the different EEG pattern differences caused by aging.

Extensive evaluations on publicly available CHB-MIT and Siena datasets have confirmed the effectiveness of this framework. The accuracy of TFFBN-HDLF on the CHB-MIT dataset reached 98.09% (AUC of 99.45%), and on the Siena dataset, it was 92.49% (AUC of 95.64%).

These results indicate that the collaborative integration of attention-based time-frequency network fusion and feature learning significantly improves diagnostic performance, demonstrating its potential application in clinical care for epilepsy in the elderly.

Figures