Abstract
Seizure onset detection in electroencephalography (EEG) signals is a challenging task due to the non-stereotyped seizure activities as well as their stochastic and non-stationary characteristics in nature. Joint spectral-temporal features are believed to contain sufficient and powerful feature information for absence seizure detection. However, the resulting high-dimensional features involve redundant information and require heavy computational load. Here, we discover significant low-dimensional spectral-temporal features in terms of mean-standard deviation of wavelet transform coefficient (MS-WTC), based on which a novel absence seizure detection framework is developed. The EEG signals are transformed into the spectral-temporal domain, with their low-dimensional features fed into a convolutional neural network. Superior detection performance is achieved on the widely-used benchmark dataset as well as a clinical dataset from the Chinese 301 Hospital. For the former, seven classification tasks were evaluated with the accuracy from 99.8% to 100.0%, while for the latter, the method achieved a mean accuracy of 94.7%, overwhelming other methods with low-dimensional temporal and spectral features. Experimental results on two seizure datasets demonstrate reliability, efficiency and stability of our proposed MS-WTC method, validating the significance of the extracted low-dimensional spectral-temporal features.
Introduction
Detecting absence seizures—brief lapses in consciousness often mistaken for daydreaming—is a critical clinical challenge. While Electroencephalography (EEG) is the gold standard for diagnosis, the signals are inherently non-stationary and stochastic, varying significantly between patients and even within the same individual over time.
Current automated detection methods typically fall into two categories. Low-dimensional methods utilize simple features like Fast Fourier Transform (FFT) or time-domain statistics; while computationally efficient, they often fail to capture the complex "spike-wave" patterns of a seizure. Conversely, high-dimensional methods use joint spectral-temporal representations, such as those from a Continuous Wavelet Transform (CWT). While detailed, these are computationally intensive, prone to overfitting, and contain significant redundancy. This research introduces the MS-WTC framework as a "Goldilocks" solution: a feature set that is low-dimensional for efficiency yet rich enough to capture essential rhythmic seizure characteristics.
The MS-WTC Framework
The core innovation of this work is the Mean-Standard Deviation of Wavelet Transform Coefficient (MS-WTC) method. The pipeline begins with preprocessing, where EEG signals are segmented into 4-second windows and bandpass filtered (0.1–35 Hz).
The feature extraction process involves applying a CWT using Morse wavelets, which are particularly effective at analyzing the "spike-and-slow-wave" patterns of absence seizures. Instead of using the full 2D time-frequency image, the researchers calculate the mean ($\mu_a$) and standard deviation ($\sigma_a$) of the coefficients across the time axis for every frequency scale. This transforms a massive 2D matrix into two compact 1D vectors, capturing both the average energy at a frequency and its fluctuations. These features are then classified using a lightweight 2-layer 1D Convolutional Neural Network (CNN), which maintains stability and prevents overfitting.
Experimental Datasets
The model was validated on two distinct datasets to ensure robustness:
- Bonn University Dataset: A standard benchmark consisting of five subsets (A-E) representing healthy states, seizure-free intervals, and ictal (seizure) activity.
- C301 Dataset: A clinical dataset from the Chinese 301 Hospital involving 19 subjects. This dataset presents real-world challenges, including artifacts from eye movements, muscle activity (EMG), and electrode displacements.
Results and Discussion
On the Bonn benchmark dataset, the MS-WTC method achieved near-perfect results, with accuracies ranging from 99.8% to 100.0% across seven classification scenarios. In the more challenging C301 clinical environment, the method achieved a mean accuracy of 94.7%, significantly outperforming traditional methods like Empirical Mode Decomposition (EMD) and standard FFT-based features.
The computational efficiency of the model is noteworthy, with prediction times between 1.27 ms and 1.48 ms for a 4-second sample. This speed, combined with the low dimensionality of the features, suggests high potential for real-time applications. Biological insights further justify this approach, as first and second-order moments (mean and variance) can capture over 90% of the information in neuronal spiking signals.
Conclusion
The MS-WTC method demonstrates that low-dimensional spectral-temporal features can provide a highly reliable and efficient means of detecting absence seizures. By focusing on the first and second-order moments of wavelet coefficients, the framework captures the essential dynamics of neuronal spiking while discarding redundant information. Future work will focus on addressing boundary effects in the wavelet transform, testing the framework on larger multi-type seizure databases like CHB-MIT, and integrating the system into wearable EEG hardware for at-home monitoring.
Summary updated