Abstract
Automatic classification of epileptic seizure types in electroencephalograms (EEGs) data can enable more precise diagnosis and efficient management of the disease. This task is challenging due to factors such as low signal-to-noise ratios, signal artefacts, high variance in seizure semiology among epileptic patients, and limited availability of clinical data. To overcome these challenges, in this paper, we present SeizureNet, a deep learning framework which learns multi-spectral feature embeddings using an ensemble architecture for cross-patient seizure type classification. We used the recently released TUH EEG Seizure Corpus (V1.4.0 and V1.5.2) to evaluate the performance of SeizureNet. Experiments show that SeizureNet can reach a weighted F1 score of up to 0.94 for seizure-wise cross validation and 0.59 for patient-wise cross validation for scalp EEG based multi-class seizure type classification. We also show that the high-level feature embeddings learnt by SeizureNet considerably improve the accuracy of smaller networks through knowledge distillation for applications with low-memory constraints.
Introduction
While automated systems for seizure detection (binary classification) have matured, the more granular task of seizure type classification remains a significant clinical hurdle. Distinguishing between focal and generalized seizures, for instance, is paramount for determining appropriate pharmacological or surgical interventions. However, EEG signals are characterized by low signal-to-noise ratios, frequent artifacts from muscle movement, and extreme inter-patient variability. Furthermore, the historical lack of large-scale, annotated datasets has hindered the development of robust deep learning models. SeizureNet addresses these challenges by treating EEG analysis through a computer vision lens, utilizing novel feature encoding and ensemble techniques.
Methodology
The SeizureNet framework is built upon three core innovations designed to extract robust features from noisy EEG data.
Saliency-encoded Spectrograms
Rather than processing raw 1D waveforms, SeizureNet transforms EEG signals into 3-channel, RGB-like images. The first channel contains a standard Fast Fourier Transform (FT). The second and third channels incorporate "spectral residual" and "multi-scale saliency" maps, respectively. This encoding highlights anomalous patterns and captures features across multiple temporal resolutions, effectively allowing the model to focus on seizure-related activity amidst background noise.
Multi-Spectral Feature Sampling (MSFS)
To improve generalization and reduce prediction variance, the authors employ an ensemble of three deep Convolutional Neural Networks. Each sub-network is trained on data sampled at different frequencies (ranging from 24Hz to 96Hz). This multi-view approach ensures that the ensemble captures a diverse set of spectral features.
Architecture and Compression
The backbone of SeizureNet utilizes DenseNet-style "Dense Blocks," where intensive skip connections facilitate gradient flow and feature reuse. To enable deployment on resource-constrained wearable devices, the authors utilized knowledge distillation. A large SeizureNet ensemble acts as a "teacher" to train a compact 3-layer ResNet "student," achieving a balance between accuracy and computational efficiency.
Experimental Results
The model was evaluated using the TUH EEG Seizure Corpus (v1.4.0 and v1.5.2), the largest public dataset of its kind. The study focused on seven distinct seizure types, including Focal Non-Specific (FN), Generalized Non-Specific (GN), and Absence Seizures (AB).
Performance Metrics
SeizureNet achieved a state-of-the-art weighted F1 score of 0.94 in seizure-wise cross-validation. In the more challenging patient-wise cross-validation—where the model is tested on patients it has never seen—it achieved an F1 score of 0.59, significantly outperforming baselines like XGBoost (0.56) and kNN (0.46).
Impact of Saliency Encoding
A key finding was the superiority of saliency-encoded spectrograms over standard Fourier transforms. As shown in the confusion matrices, the saliency-encoded approach provides much clearer class separation, particularly for complex seizure types that are often misclassified by traditional methods.
Knowledge Distillation
The "student" ResNet model achieved an F1 score of 0.746. While lower than the full ensemble, this model is 45x faster and requires 1100x fewer FLOPs, demonstrating a viable path for real-time monitoring on mobile or wearable hardware.
Conclusion and Future Work
SeizureNet demonstrates that treating EEG signals as multi-spectral images can significantly improve seizure type classification. The use of saliency encoding and multi-spectral sampling provides a robust framework for handling the inherent noise and variability of brain activity. However, the performance gap between seizure-wise and patient-wise validation remains a critical area for improvement, highlighting the need for models that can better adapt to individual patient semiology. Future research will explore multi-modal fusion—incorporating video and other physiological sensors—and real-time clinical validation in epilepsy monitoring units.
Summary updated