Independent Component Analysis (ICA) methods for neonatal EEG artifact extraction: Sensitivity to variation of artifact properties

Abstract

Independent Component Analysis (ICA) is becoming an accepted technique for artifact removal. Nevertheless, there is no consensus about appropriate methods for different applications. This study presents a comparison of common ICA methods: RobustICA, SOBI, JADE, and BSS-CCA, for extraction of ECG artifacts from EEG signal. Algorithms were applied to the data created by superimposing artifact free real-life neonatal EEG and synthetic ECG. Their sensitivity to variation of noise property was compared: we examined variability of Spearman correlation coefficients (SCC) for various Heart Rates (HR) in each of ICA methods. Results show that SOBI and BSS-CCA methods were less sensitive than RobustICA and JADE to artifact alterations (mean SCCs were 0.85 and 0.85 compared to 0.80 and 0.73, respectively) being quite successful in source signal extraction.