Adaptive median feature baseline correction for improving recognition of epileptic seizures in ICU EEG

S. Raghu, Natarajan Sriraam*, Erik D. Gommer, Danny M. W. Hilkman, Yasin Temel, Shyam Vasudeva Rao, Pieter L. Kubben

*Corresponding author for this work

Research output: Contribution to journalArticleAcademicpeer-review

2 Citations (Web of Science)
8 Downloads (Pure)

Abstract

Automated classification of epileptic seizures surrogates the manual interventions required for analyzing long-term electroencephalographic (EEG) signals and helps to speed up the treatment in epilepsy patients. Developing a patient independent algorithm is a great challenge due to the differences in EEG characteristics. Feature distribution among many subjects results in inter-subject variability, which leads to poor classification performance. Therefore, in order to overcome this limitation, we have proposed a novel adaptive median feature baseline correction (AM-FBC) method to update the feature distribution. Two recently proposed features referred to as successive decomposition index (SDI) and matrix determi-nant (MD) were extracted from 40 intensive care unit patients EEG at a segmentation length of 4 s with 50% overlap. We have investigated the influence of outliers removal and correction, AM-FBC, and post -processing of classifier output to improve the seizure detection results. The classification was performed using a support vector machine classifier with leave-one-subject-out cross-validation. With the applica-tion of above-mentioned methods, the highest area under the curve (AUC) of 0.9663 (sensitivity S+ = 0.9661, specificity (S- = 0.8446) and 0.9812 S+ = 0.9822, S- = 0.8705) was achieved using SDI and MD fea-tures respectively. Further, the AUC of 0.9593 (S+ = 0.9069, S- = 0.8695) was achieved when both SDI and MD features were used with the application of the outliers correction method. The findings of the study suggest (1) Outliers correction method does not improve results (2) AM-FBC enhances the results (3) Post-processing method improved the classification results at least 2 to 5% and reduced false detections (4) Lowering the outlier removal factor showed good AUC at the cost of loss of feature samples. (C) 2020 Elsevier B.V. All rights reserved.

Original languageEnglish
Pages (from-to)385-398
Number of pages14
JournalNeurocomputing
Volume407
DOIs
Publication statusPublished - 24 Sep 2020

Keywords

  • Adaptive median feature baseline correction
  • Epileptic seizures
  • Matrix determinant
  • Post-processing
  • Successive decomposition index
  • Support vector machine
  • SIGNAL CLASSIFICATION
  • APPROXIMATE ENTROPY
  • WAVELET TRANSFORM
  • SVM
  • REPRESENTATION
  • NORMALIZATION
  • LOCALIZATION

Cite this