ICA (Independent Component Analysis) for PPG Signal Separation
Independent Component Analysis (ICA) is a blind source separation technique that decomposes a mixture of signals into statistically independent components without requiring a reference signal. For PPG, ICA applied to multi-wavelength or multi-site recordings can separate the cardiac signal from motion artifacts and other physiological noise sources.
ICA assumes that the observed signals are linear mixtures of independent source signals: x = As, where A is the unknown mixing matrix and s are the independent sources. The ICA algorithm estimates the unmixing matrix W such that y = Wx approximates the original sources. For PPG, two or more sensor channels are required — commonly two PPG wavelengths (green + red) or two spatially separated sensors.
FastICA, the most popular ICA algorithm, alternately maximizes non-Gaussianity of the estimated components using Newton optimization with negentropy approximations. JADE (Joint Approximate Diagonalization of Eigenmatrices) uses fourth-order cumulants for source estimation, showing better performance on super-Gaussian sources like cardiac PPG. SOBI (Second-Order Blind Identification) exploits temporal correlations rather than statistical independence, making it particularly effective for temporally structured biosignals.
The JOSS (Joint Optical-Spectral Separation) framework published by Salehizadeh et al. demonstrated that ICA combined with spectral power selection can achieve <3 bpm heart rate error during moderate walking (3–5 mph), without any reference accelerometer data. This "zero-reference" capability is a key advantage over LMS/RLS. The limitation is that ICA requires simultaneous multi-channel recording and fails when the number of independent sources exceeds the number of sensor channels.
Frequently Asked Questions
Does ICA for PPG require an accelerometer?
No. ICA for PPG source separation operates on multiple PPG channels (different wavelengths or spatial locations) without any reference sensor. This is its key advantage over LMS/RLS which require accelerometer references.
How many PPG channels are needed for ICA?
At minimum two channels are required to separate two sources (cardiac + motion). Three channels (e.g., RGB from a standard camera) allow separation of three independent components, providing better artifact rejection.
Can ICA separate multiple motion artifact sources?
ICA can separate as many sources as there are input channels. With three channels, it can theoretically separate three independent sources. However, real motion artifacts are rarely statistically independent from each other, limiting practical separation quality.