Online denoising of eye-blinks in electroencephalography

Abstract

The AJDC algorithm addresses limitations observed in AMUSE and outperforms it. No statistical difference is found between the manual and automatic approaches on a database composed of 15 healthy individuals, paving the way for an automated, operator-independent, and real-time eye-blink correction technique.

The Riemannian Potato Field: A Tool for Online Signal Quality Index of EEG

“…Let us consider an EEG recording X ∈ R C×N , in an online setting where new data become available sample-by-sample or block-by-block [8]. To meet the computational requirements of an online implementation, the recording is cut in overlapping epochs, resulting in a sliding window.…”

“…2. Correction of eye-blinks is performed by an online BSS method based on the approximate joint diagonalization of Fourier cospectra (AJDC) from 1 to 21 Hz [8], explaining why there is no potato dedicated to ocular artifacts in the RPF of Fig. 2.…”

“…Eight methods derive from the Cartesian product of the following two parameters: method (RP or RPF) and update mode (dynamic, semi-dynamic initialized on assessment (the subject's first available recording), semidynamic initialized on all other subjects (using a generic calibration), and static initialized on assessment). These Riemannian techniques are compared to another online method called Artifact Subspace Reconstruction (ASR) [66], originally designed for artifact correction, but which can be easily used for artifact detection 7 and can be considered as online static 8 .…”

“…In this context, the task is significantly more complex: it cannot be done retrospectively as in offline analysis, by a human operator or an automatic offline algorithm; denoising can only rely on automated pre-calibrated approaches that should be more generalizable as compared to an offline strategy. While some techniques have proven efficacious and reliable to correct some types of stationary artifacts including eye blinks [8], most types of noise affecting EEG remain transient and non-stationary (such as muscle activities and body movements) making correction strategies, specifically for real-time applications, only partly reliable. In this context, it is necessary to fall back on a detection strategy, informing of the data quality.…”

“…Belonging to anomaly detection [32], other approaches exist for detecting artifacts in EEG [33]- [36]. Most of them process block-online or offline instead of online (see [8], [37] for details on these distinctions), or apply BSS and give one index by source to remove artifacted ones [38]- [41]. A supplementary advantage of the online RP is that, since it is based on covariance matrices, it uses the same features used for classification in the new generation of BCI paradigms [17]: motor-imagery [42], event-related potentials [43], and SSVEP [31].…”