CrossSeizure: Learning Generalizable Representations for Cross-Dataset Epileptic Seizure Classification

A CNN-Transformer hybrid framework for generalizable epileptic seizure detection across diverse EEG datasets.

Overview

Traditional machine learning models for epileptic seizure classification perform well within a single dataset but fail when tested on EEG data from different clinical settings. This project develops a lightweight CNN-Transformer hybrid architecture that learns generalizable seizure representations directly from minimally preprocessed raw EEG signals, achieving robust cross-dataset performance without dataset-specific tuning.

Key Contributions

  • Novel Architecture: Combined EEGNet-style convolutions with Transformer attention for temporal-spatial feature learning
  • Cross-Dataset Generalization: Achieved robust performance across BEED, BONN, and CHB-MIT datasets with different sampling rates and channel configurations
  • Minimal Preprocessing: Used only 0.5 Hz high-pass filtering, preserving native EEG characteristics
  • Domain Adaptation: Implemented DANN and CDAN strategies for domain-invariant feature learning

Problem Statement

Current seizure detection models exhibit poor cross-dataset generalization due to:

  • Dataset-specific frequency distributions and manifold geometries
  • Dependence on hand-crafted features (FFT, UMAP)
  • Overfitting to recording-specific artifacts

Our framework addresses these limitations by learning universal seizure dynamics rather than dataset-specific patterns.

Methodology

Datasets

  • BEED: 8,000 samples, 16 channels, 256 Hz (4 classes)
  • BONN: 500 samples, 1 channel, 173.61 Hz (5 classes)
  • CHB-MIT: Clinical data, 23 channels, 256 Hz (2 classes)

Model Architecture

  1. CNN Component: Depthwise separable convolutions for local temporal-spatial features
  2. Transformer Component: Multi-head self-attention for long-range dependencies
  3. Training Strategy: Multi-dataset pretraining with domain adaptation

Data Augmentation

Time-warping (60%), window slicing (30%), and amplitude scaling (10%) to enhance robustness.

Results

On-going Project

Supervision

Conducted under the supervision of Dr. Saptarshi Purkayastha, Department of Biomedical Engineering and Informatics, Indiana University Indianapolis.