Low variance adaptive filter for cancelling motion artifact in wearable photoplethysmogram sensor signals

The photoplethysmogram (PPG) is an extremely useful wearable sensing medical diagnostic tool. However, the PPG signal becomes highly corrupted and unusable when the sensor wearer is in motion. This paper investigates how confidently Widrow's Adaptive Noise Cancellation can eliminate motion artifact and recover a motion corrupted PPG signal for a wearer engaged in jogging motions. It has previously been shown that Widrow's Adaptive Noise Cancellation can recover a motion corrupted PPG signal for certain data sets by using a collocated accelerometer to measure the corrupting motion. However, wearer motion is band limited, and provides little information for estimating motion-to-PPG noise transfer dynamics. This means that, without proper care, recovery results can be unreliable. In the present work, both Finite Impulse Response (FIR) and Laguerre series black box transfer dynamics models are evaluated for how confidently they can be identified. Model confidence is quantified in terms of variance of the transfer dynamics estimate at the motion frequencies. For typical jogging motion, it is found that standard deviation of the FIR model transfer dynamics is 30% of the mean value at the motion input frequency. The standard deviation of the Laguerre model transfer dynamics is only 1%. Time domain data shows how a Laguerre model outperforms a FIR model in accordance with the computed model variance.