Averaged sparse local representation for the elimination of pseudo-HFOs from intracranial EEG recording in epilepsy

Abstract

Interictal high-frequency oscillation (HFO) is considered a promising biomarker of the epileptogenic zone. The pseudo-HFOs originating from artifacts and noise might escape HFO detectors and mislead the seizure onset zone (SOZ) localization. The purpose of this study is to propose a new sparse representation framework fused with a random forest classifier to detect the real HFOs and eliminate the pseudo-ones. In this scheme, each candidate event that passed a conventional amplitude threshold-based detector was represented locally in a sparse fashion. Specifically, the signal is divided into overlapping windows and using orthogonal matching pursuit, only a few oscillatory atoms selected from a predefined redundant Gabor dictionary were used to approximate the signal locally. Later, the approximations in overlapping segments are averaged to increase the smoothness. Finally, the ability to reconstruct an event is translated to informative features and fed into a random forest classifier. This technique was tested on 10 minutes of interictal intracranial EEG (iEEG) recordings recorded from 11 patients with epilepsy. In this framework, three experts visually inspected 4466 events captured by the amplitude threshold-based HFO detector in iEEG recordings and labeled them as real-HFO or Pseudo-HFO. We reached 89.77% classification accuracy in these labeled events. Furthermore, the success of the method assessed by calculating the spatial overlap between the detected HFOs and SOZ channels. Compared to conventional amplitude threshold-based HFO detector, our method resulted a significant 18.27% improvement in the localization of SOZ.