Epileptic seizure detection based on imbalanced classification and wavelet packet transform

Automatic seizure detection is significant for the diagnosis of epilepsy and the reduction of massive workload for reviewing continuous EEG recordings.

Compared with the long non-seizure periods, the durations of the seizure events are much shorter in the continuous EEG recordings. So the seizure detection task can be regarded as an imbalanced classification problem. In this paper, a novel method based on the weighted extreme learning machine (ELM) is proposed for seizure detection with imbalanced EEG data distribution. Firstly, the wavelet packet transform is employed to analyze the EEG data and obtain the time and frequency domain features, and the pattern match regularity statistic (PMRS) is used as the nonlinear feature to quantify the complexity of the EEG time series. After that, the EEG feature vectors are discriminated by the weighted ELM. It can assign different weights for the EEG feature samples according to the class distribution, so that to effectively moderate the bias in performance caused by imbalanced class distribution.

The metric G-mean which takes into account of both the sensitivity and specificity is used to evaluate the performance of this method. The G-mean of 93.96%, event-based sensitivity of 97.73% and false alarm rate of 0.37/h are yielded on the publicly available EEG dataset.

The comparison with other detection methods shows the superior performance of this method, which indicates its potential for detecting seizure events in clinical practice. Additionally, much larger amounts of true continuous EEG data will be used to test the proposed method further in the future work.