Temporally Local Maximum Signal Fraction Analysis for Artifact Removal From Biomedical Signals

In this correspondence, we present a novel spatio-temporal extractor of source components, termed temporally local maximum signal fraction analysis (TLMSFA), which is used for artifact removal from biomedical signals. Different from classical maximum signal fraction analysis (MSFA) that uses only global spatial covariances to define signal-to-noise (SNR) expression, TLMSFA considers temporally local information in the covariance formulations used in the SNR modelling. The local time-dependent spatial covariances contain more information for source separation of biomedical signals that slowly change over time. TLMSFA is a temporal generalization of MSFA in the sense that MSFA can be formulated under the TLMSFA umbrella as a special instance. We show that TLMSFA is actually to approximately maximize the temporally local autocorrelation of the observations. By designing a time-dependent weighting function, TLMSFA is solved as a generalized eigenvalue problem. So, TLMSFA is computationally as competitive as MSFA. Empirical evaluation and experiments of source separation on real electrocardiogram data and artifact removal on real electroencephalogram data show the effectiveness of the proposed TLMSFA technique.