Motion-Robust Pulse Wave Measurement Using Piezoelectric Film Sensors With Machine Learning-Based Artifact Removal and Regression-Based Validity Assessment

Piezoelectric film sensors can be used to measure pulse wave, but motion artifacts often contaminate the signal, leading to measurement failures. This study proposes a method for estimating pulse waves from contaminated data by effectively removing motion artifacts through a bidirectional recurrent neural network. Additionally, we propose a regression-based method to assess the validity of the estimated pulse wave by predicting the cross-correlation coefficient between the estimated and the corresponding reference pulse waves, enabling the selection of reliable waveforms. According to the experimental results, the recall rates were 96%, 91%, and 91% when the threshold values for the predicted coefficients were set to 0.4, 0.6, and 0.8, respectively, allowing us to select the estimated pulse wave with high validity without significant data loss. Combining these two methods makes it possible to measure pulse waves during body movement and obtain essential biometric information.