LMS Adaptive Filter for PPG Signal Processing
The Least Mean Squares (LMS) adaptive filter is the most widely used method for motion artifact removal in wearable PPG. It uses an accelerometer signal as a noise reference to iteratively estimate and cancel the motion-induced noise component from the PPG, adapting its filter coefficients based on the error between the filtered output and the desired clean signal.
The LMS algorithm minimizes the mean square error between the filter output and a desired signal using a stochastic gradient descent update rule: w(n+1) = w(n) + 2μe(n)x(n), where w is the weight vector, μ is the step size (learning rate), e(n) is the error signal, and x(n) is the reference input. For PPG, the reference input is the accelerometer signal, the primary input is the corrupted PPG, and the error output approximates the cleaned cardiac PPG component.
The step size μ controls the tradeoff between convergence speed and steady-state error. Large μ values converge quickly but exhibit high misadjustment noise; small μ values produce cleaner output but adapt slowly to nonstationary motion patterns. The normalized LMS (NLMS) variant adapts μ based on the input power, providing more robust performance across varying motion intensities without manual tuning.
Key limitations of LMS for PPG include the linearity assumption (motion noise is modeled as linearly correlated with accelerometer output), which fails for complex non-linear artifacts from skin-sensor interface micromotion. During high-intensity exercise, LMS noise cancellation performance degrades significantly, with heart rate estimation errors exceeding 10 bpm despite LMS processing. Multi-reference LMS using triaxial accelerometer data (three separate adaptive filters combined) substantially improves performance over single-axis reference.
Frequently Asked Questions
How does LMS compare to RLS for PPG artifact removal?
RLS converges faster and achieves lower steady-state error than LMS but has higher computational complexity (O(N²) vs O(N)). For battery-constrained wearables, LMS/NLMS is preferred. RLS outperforms LMS in rapidly changing motion scenarios.
Can LMS work without an accelerometer reference?
No. LMS requires a noise reference signal. Without an accelerometer, blind source separation methods (ICA, SVD) must be used. Attempts to use the second PPG wavelength as a motion reference have shown mixed results.
What filter order should be used for PPG LMS?
Typical implementations use filter orders of 16–64 taps. Longer filters can model more complex motion interference but increase computational load. Order selection via BIC or AIC criteria optimizes the tradeoff.