NLMS (Normalized LMS) Adaptive Filter for PPG
The Normalized Least Mean Squares (NLMS) algorithm adapts the LMS step size by dividing by the input signal power, eliminating sensitivity to input scaling and providing more consistent convergence across varying motion intensities in PPG wearables.
NLMS modifies the LMS update rule to w(n+1) = w(n) + (μ / (ε + ||x(n)||²)) · e(n) · x(n), where the step size is normalized by the squared Euclidean norm of the input vector plus a small regularization constant ε. This normalization ensures that convergence speed is independent of input signal power, which varies dramatically in PPG applications as motion intensity changes.
For wearable PPG, NLMS is the de facto standard adaptive filter because it requires no manual step-size tuning across different activities. During walking (low-power accelerometer reference), NLMS automatically uses a larger effective step size for fast adaptation. During running (high-power reference), the effective step size shrinks to prevent excess noise. The TROIKA framework by Zhang et al. (2015) uses NLMS as the first stage in its signal reconstruction pipeline, achieving 2.36 bpm mean absolute error on the IEEE Signal Processing Cup 2015 dataset.
Block NLMS variants process multiple samples simultaneously, reducing per-sample computational overhead on DSP architectures with SIMD support. Affine projection algorithms (APA) generalize NLMS by projecting the weight update onto multiple past input vectors, improving convergence for colored noise at the cost of O(P²N) complexity where P is the projection order.
Frequently Asked Questions
How does NLMS differ from standard LMS for PPG?
NLMS normalizes the step size by input power, making convergence independent of signal level. Standard LMS requires manual step-size tuning for each activity type, while NLMS adapts automatically.
What regularization constant should be used in NLMS for PPG?
Typical values are ε = 0.001–0.01 times the expected input power. Too small values cause numerical instability when the input approaches zero; too large values reduce the normalization benefit.
Is NLMS sufficient for high-intensity exercise PPG?
NLMS alone achieves 5–10 bpm error during vigorous exercise. Combined with spectral peak tracking (as in TROIKA), errors reduce to 2–4 bpm. For extreme motion, nonlinear methods or deep learning outperform NLMS.