Artifact Removal from EEG Using a Multi-objective Independent Component Analysis Model

Abstract

Independent Component Analysis (ICA) has been widely used for separating artifacts from Electroencephalographic (EEG) signals. Still, a few challenging problems remain.

First, in real-time applications, visual inspection of components should be replaced with an automatic identification method or a heuristic for artifacts detection. Second, as we will explain more in the paper, we expect to have a clear order relationship between an electrode and a corresponding component. Third, we need to minimize the EEG information loss during artifact removal while also minimizing the residue of the artifact in the cleaned signal.

In this paper, we propose a decomposition of the independent components. This decomposition separates each component into two vectors, one - we call local vector - maintains maximum information from the unique EEG information encoded by an electrode, while the other - we call shared vector - absorbs overlapping artifact information. We present an explicit Pareto-based multi-objective optimization formulation that trade-off similarity between the local vector and the original vector on the one hand, and the uncorrelatedness of all local vectors from all components on the other hand. We demonstrate that the proposed method can automatically isolate artifacts from an EEG signal while preserving maximum EEG information.