SeizureNet: Multi-Spectral Deep Feature Learning for Seizure Type Classification

Umar Asif, Subhrajit Roy, Jianbin Tang, Stefan Harrer

TL;DR

This paper introduces SeizureNet, a novel deep learning framework that leverages multi-spectral feature learning and an ensemble architecture to improve the accuracy of classifying different seizure types from EEG data. By creating "saliency-encoded spectrograms" and training diverse sub-networks, SeizureNet achieves state-of-the-art performance, particularly in challenging patient-wise cross-validation scenarios.

What Problem Does This Paper Address?

Epilepsy affects a significant portion of the global population, and accurately classifying seizure types is crucial for effective diagnosis and treatment. However, this task is notoriously difficult due to noisy EEG signals, artifacts, patient-specific variations in seizure presentation, and a scarcity of well-annotated clinical data. Existing methods have primarily focused on seizure detection or prediction, leaving seizure type classification as an underdeveloped area. The paper aims to bridge this gap by developing a robust deep learning model capable of distinguishing between various seizure types, even when generalizing to new, unseen patients.

Methods

SeizureNet employs a multi-pronged approach:

  • Saliency-Encoded Spectrograms: Instead of using raw EEG or standard spectrograms, the authors propose a novel representation that captures multi-scale saliency information. This involves combining a Fourier Transform map, a spectral saliency map, and a multi-scale saliency map. This richer representation is designed to highlight more discriminative features within the EEG data.
  • Multi-Spectral Feature Sampling (MSFS): To encourage diversity within the ensemble and combat overfitting, especially with limited data, MSFS is introduced. This technique involves sampling EEG data at different frequency and temporal resolutions to train individual sub-networks. This diversification helps the ensemble learn more robust and generalizable features.
  • Ensemble Architecture: SeizureNet is built upon an ensemble of Deep Convolutional Neural Networks (DCNs). Each DCN utilizes a "Dense Block" architecture, known for its efficient feature reuse. By combining the predictions of multiple diverse sub-networks, the ensemble approach aims to reduce variance and improve overall prediction stability.
  • Knowledge Distillation: To enable deployment on resource-constrained devices, the authors explore knowledge distillation. They demonstrate that the feature embeddings learned by the larger SeizureNet can be effectively transferred to smaller, more efficient networks (like a 3-layer ResNet), significantly improving their accuracy while drastically reducing computational requirements.

Datasets

The primary dataset used for training and evaluation is the TUH EEG Seizure Corpus (TUH-EEGSC), specifically versions V1.4.0 and V1.5.2. This corpus is highlighted as the largest publicly available dataset for seizure type classification. The authors conducted experiments using both seizure-wise and patient-wise cross-validation to assess the model's performance under different generalization scenarios.

Key Insights and Ideas

  • Saliency is Key: The proposed "saliency-encoded spectrograms" are a significant departure from standard time-frequency representations. The paper convincingly shows that these saliency-enhanced features lead to a substantial reduction in classification confusions, particularly when dealing with limited training data.
  • Diversity Breeds Robustness: The Multi-Spectral Feature Sampling (MSFS) technique is crucial for the ensemble's success. By forcing sub-networks to learn from data sampled at different resolutions, it creates a more diverse ensemble, which in turn leads to better generalization and reduced variance in predictions.
  • Ensembling for Generalization: The ensemble approach, combined with MSFS, proves highly effective for patient-wise cross-validation. This is a critical finding, as patient-wise generalization is more representative of real-world clinical scenarios where models need to perform on individuals not seen during training.
  • Knowledge Distillation for Practicality: The ability to distill knowledge from a large, complex model into a smaller, efficient one is a vital contribution for practical deployment. This opens doors for using advanced seizure classification in real-time monitoring systems or on edge devices.

Key Results

  • State-of-the-Art Performance: SeizureNet achieved a weighted F1 score of up to 0.94 for seizure-wise cross-validation and 0.59 for patient-wise cross-validation on the TUH-EEGSC dataset. These results represent significant improvements over existing methods.
  • Improved Patient-Wise Accuracy: The paper highlights that SeizureNet improved patient-wise weighted F1 scores by approximately 3 points compared to existing methods, underscoring its ability to generalize to unseen patients.
  • Saliency Spectrograms Outperform FT: Figure 2 demonstrates a dramatic reduction in confusions when using saliency-encoded spectrograms compared to standard Fourier Transforms, especially with limited training data (10%).
  • MSFS Boosts Performance: Table 3 shows that using MSFS consistently improved seizure-wise F1 scores across different data sizes, with a notable 9-point improvement at 50% data size.
  • Knowledge Distillation Benefits: A 3-layer ResNet model trained with knowledge distillation from SeizureNet (ResNet-KD) achieved a 2% improvement in seizure-wise mean F1 score compared to the same ResNet model without distillation, while being significantly smaller and faster.

Limitations and Future Work

  • Limited Data for Certain Seizure Types: The paper notes that Myoclonic (MC) seizures were excluded due to their extremely low count, and Tonic (TN) and Simple Partial (SP) seizures in v1.4.0 had data from only 2 patients, making patient-wise cross-validation statistically challenging for these specific types.
  • Future Directions: The authors plan to investigate the fusion of data from wearable sensors and videos for multi-modal seizure type classification. This suggests a move towards more comprehensive and context-aware seizure analysis systems.

Summary updated

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 0.95 for seizure-wise cross validation and 0.62 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.
BibTeX
@inproceedings{asif2020seizurenet,
  author = {Umar Asif and Subhrajit Roy and Jianbin Tang and Stefan Harrer},
  title = {{SeizureNet: Multi-Spectral Deep Feature Learning for Seizure Type Classification}},
  booktitle = {MLCN/RNO-AI @ MICCAI},
  year = {2020},
  url = {https://openreview.net/forum?id=VlbrzQTKVJy},
}
Figure