MUSIC Algorithm for PPG Spectral Analysis
The MUSIC (Multiple Signal Classification) algorithm provides super-resolution spectral estimation for PPG by decomposing the signal correlation matrix into signal and noise subspaces. MUSIC resolves closely spaced spectral peaks that FFT cannot separate, enabling accurate heart rate estimation when cardiac and motion frequencies overlap.
MUSIC constructs the autocorrelation matrix R = E[x·x^H] of the PPG signal vector and performs eigendecomposition to separate signal eigenvalues (large) from noise eigenvalues (small, approximately equal). The noise subspace eigenvectors are orthogonal to the signal steering vectors, creating the MUSIC pseudospectrum P(f) = 1 / (a^H(f) · E_n · E_n^H · a(f)) with sharp peaks at true signal frequencies even when they are closely spaced.
For PPG during exercise, cardiac frequency (e.g., 2.5 Hz = 150 bpm) and running cadence (e.g., 2.7 Hz = 162 steps/min) may be separated by only 0.2 Hz — unresolvable by standard FFT with typical window lengths. MUSIC with model order p = 3–5 components reliably resolves these closely spaced peaks, enabling correct identification of the cardiac peak versus the motion peak. The SpaMA (Spectral peak Approximation with MUSIC Algorithm) framework achieves 1.8 bpm MAE during treadmill running by combining MUSIC spectral estimation with harmonic verification.
The main limitation is computational complexity: eigendecomposition of the p×p correlation matrix requires O(p³) operations per window. For real-time embedded implementation, model order p = 4–8 is typical, requiring correlation matrix estimation from L = 3p–5p samples. Root-MUSIC, which finds polynomial roots instead of scanning a pseudospectrum grid, reduces computation when only a few frequency estimates are needed.
Frequently Asked Questions
When should MUSIC be used instead of FFT for PPG?
MUSIC excels when cardiac and motion frequencies are within 0.5 Hz of each other, which occurs during running when step rate approaches heart rate. For well-separated frequencies, FFT is simpler and equally effective.
How is the model order selected for PPG MUSIC?
Model order equals the number of assumed sinusoidal components. For PPG, p = 2–4 captures the cardiac fundamental plus 1–2 harmonics plus motion. AIC or MDL criteria can estimate optimal order, though fixed p = 4 works well in practice.
What is the minimum data length for MUSIC on PPG?
MUSIC requires at least 2p+1 samples for correlation matrix estimation. With p = 4 and 25 Hz sampling, this is only 0.36 seconds. However, 3–5 seconds provides more stable correlation estimates and better frequency resolution.