A Blind Source-Based Method for Automated Artifact-Correction in Standard Sleep EEG

TY - GEN

T1 - A Blind Source-Based Method for Automated Artifact-Correction in Standard Sleep EEG

AU - Waser, Markus

AU - Garn, Heinrich

AU - Jennum, Poul J.

AU - Sorensen, Helge B.D.

PY - 2018/7/1

Y1 - 2018/7/1

N2 - Electroencephalogram (EEG) is a common tool in sleep medicine, but it is often compromised by non-neural artifacts. Excluding visually identified artifacts is time-consuming and removes relevant EEG information. Blind source separation (BSS) techniques, on the other hand, are capable of separating "brain" from "artifact source components". Existing algorithms for automated component labeling require either a priori morphological information or adaptation to individual recordings. We present a method for the automated identification of artifact components based on their autocorrelation and spectral properties. It requires no tuning for individual recordings. The method was tested on 100 one-minute EEG segments during rapid eye movement sleep. EEG source components were estimated by second order blind source identification and, as reference, manually labeled as "brain" or "artifact component". The algorithm identified electro-cardiogram components by autocorrelation peaks between 0.5-1.5 seconds and -oculogram components by linear discriminant analysis of spectral band-power. Using 5-fold cross-validation, we observed 97% accuracy (95% sensitivity, 98% specificity), as well as minimized correlation of artifacts and the EEG. The approach has demonstrated its potential as promising tool for a broad range of sleep medical applications.

AB - Electroencephalogram (EEG) is a common tool in sleep medicine, but it is often compromised by non-neural artifacts. Excluding visually identified artifacts is time-consuming and removes relevant EEG information. Blind source separation (BSS) techniques, on the other hand, are capable of separating "brain" from "artifact source components". Existing algorithms for automated component labeling require either a priori morphological information or adaptation to individual recordings. We present a method for the automated identification of artifact components based on their autocorrelation and spectral properties. It requires no tuning for individual recordings. The method was tested on 100 one-minute EEG segments during rapid eye movement sleep. EEG source components were estimated by second order blind source identification and, as reference, manually labeled as "brain" or "artifact component". The algorithm identified electro-cardiogram components by autocorrelation peaks between 0.5-1.5 seconds and -oculogram components by linear discriminant analysis of spectral band-power. Using 5-fold cross-validation, we observed 97% accuracy (95% sensitivity, 98% specificity), as well as minimized correlation of artifacts and the EEG. The approach has demonstrated its potential as promising tool for a broad range of sleep medical applications.

U2 - 10.1109/EMBC.2018.8513619

DO - 10.1109/EMBC.2018.8513619

M3 - Article in proceedings

C2 - 30441706

SN - 978-1-5386-3646-6

T3 - Proceedings of the International Conference of the IEEE Engineering in Medicine and Biology Society

SP - 6010

EP - 6013

BT - 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)

PB - IEEE

ER -