Robust dimensionality-reduced epilepsy detection system using EEG wavelet packets and machine learning

In response to the challenges in epileptic seizure detection arising from the high dimensionality of EEG signals and the localization of the epileptogenic focus, this paper presents a robust EEG-type-independent system designed to tackle these issues by means of wavelet packet features (WPF) and both supervised and unsupervised machine learning algorithms. Moreover, our study underscores the significance of mother wavelet selection to enhance accuracy and effectiveness in epilepsy detection.

The developed system consists of two complementary subsystems, namely the Representative EEG Channel Creator (RECC) and the Seizure Affected EEG Channel Detector (SAECD). The RECC utilizes WPF to generate representative channels from EEG signals, subsequently feeding them into an LDA classifier for epilepsy detection. If the RECC returns a positive result, the SAECD is then launched to track the seizure’s diffusion path using the Energy-to-Shannon-Entropy ratio and k-means clustering.

Using db6 mother wavelet, the RECC achieved a high sensitivity of 98.46% for epileptic seizure identification with a significant reduction in EEG dimensionality with up to 93.75%. On the other hand, through the developed SAECD, we were able to circumscribe the epileptogenic area with an average silhouette range of [51.21 \(-\) $-$ 88.18]%.

The outcomes of the developed system are promising, confirming that epileptic seizures may indeed be detected within a few amount of data, providing evidence in terms of processing time and efficiency.