NIRS measurement is known to be susceptible to motion artifact, instrumental noise, etc. When analyzing NIRS time series, we almost always lack the ground truth for evaluation, and from many methods, it could be hard to pick the most appropriate method for one’s application.
In this work, we proposed and examined the following pipeline: First, generate ground truth synthetic NIRS signal. Second, add application specific noises/activations. Finally, train a Long Short-Term Memory deep learning model for time series prediction. Our results revealed that this approach can recover the uncorrupted NIRS/PPG signal accurately and efficiently from noisy signals.
|