SVD (Singular Value Decomposition) Applied to PPG
Singular Value Decomposition (SVD) decomposes a PPG signal matrix into orthogonal signal subspaces, enabling motion artifact removal by truncating or filtering singular components associated with noise. SVD-based methods are particularly effective for colored noise removal in resting-state PPG.
For PPG processing, SVD is applied to a Hankel or trajectory matrix constructed from delayed copies of the PPG signal. The SVD decomposes this matrix as X = UΣV^T, where U and V are orthogonal matrices and Σ is diagonal with singular values in decreasing order. The cardiac signal component typically concentrates energy in the first few singular vectors, while broadband noise and slow drift distribute across many components.
Singular Spectrum Analysis (SSA), which uses SVD of the trajectory matrix as its core step, is particularly powerful for PPG baseline wander removal and rhythm extraction. By selecting and reconstructing only the first K singular components corresponding to the cardiac oscillation, SSA achieves near-perfect baseline wander correction without distorting the pulse waveform morphology — a key advantage over IIR high-pass filters which can cause phase distortion.
Randomized SVD algorithms (based on randomized linear algebra) enable SVD computation on long PPG recordings at O(mNK) complexity instead of O(mN²) for full SVD, where m is the embedding dimension, N is signal length, and K is the number of retained components. This makes SVD-based PPG analysis practical for continuous monitoring on edge devices. Recent work has combined SVD component selection with spectral constraints to improve separation between cardiac and motion artifact subspaces during moderate exercise.
Frequently Asked Questions
How do you choose the number of singular values to retain for PPG?
The number of retained components K is selected based on the singular value gap (elbow in the Scree plot), or by choosing K components that collectively explain >95% of total signal variance. For resting PPG, K=2–4 typically captures the cardiac signal.
How does SVD compare to wavelet decomposition for PPG?
SVD is data-adaptive and parameter-free, making it more robust than wavelet methods that require manual selection of wavelet type and decomposition levels. However, wavelets provide better time-frequency localization for transient motion artifacts.
Can SVD work for real-time PPG processing?
Yes, using sliding window SVD with window lengths of 5–10 seconds and 50% overlap. Randomized SVD or rank-1 online updates enable real-time processing at typical PPG sampling rates (25–100 Hz).