Temporally Nonstationary Component Analysis; Application to Noninvasive Fetal Electrocardiogram Extraction

Area of Science:

• Biomedical Signal Processing
• Machine Learning
• Statistical Signal Analysis

Background:

• Temporally nonstationary signals are prevalent in biomedical applications.
• Signal nonstationarity can be leveraged for effective signal separation.
• Existing methods may struggle with complex mixtures of nonstationary sources and noise.

Purpose of the Study:

• To propose a semi-blind source separation algorithm for extracting temporally nonstationary components from multichannel mixtures.
• To develop a hypothesis testing framework for detecting and fusing nonstationary events using statistical properties.
• To validate the algorithm's performance on real-world biomedical data, specifically noninvasive fetal cardiac recordings.

Main Methods:

• A semi-blind source separation algorithm is introduced for nonstationary signal extraction.
• A hypothesis test utilizes ad hoc indexes to monitor the innovation process's first and second-order statistics.
• The framework is tested using local power variations and model-deviations detectors with an extended Kalman filter on fetal cardiac data.

Main Results:

• The proposed method demonstrates effective extraction of nonstationary components from signal and noise mixtures.
• Performance is assessed across various signal-to-noise ratios and in the presence of white and colored noise.
• The algorithm shows promise in identifying and separating nonstationary events within complex datasets.

Conclusions:

• The developed semi-blind source separation scheme is general and applicable to multivariate data.
• It facilitates the extraction of nonstationary events and deviations from presumed models.
• This approach addresses a common challenge in various machine learning applications, particularly in biomedical signal analysis.