A Toolkit for Motion Artifact Signal Generation

Abstract

In the research and development stages of biomedical signal quality analysis tools, testing and validation help to ensure they work as intended and are robust enough to be used in all sorts of environments. Large datasets of biomedical signals (e.g., electrocardiogram, electromyogram) and signal contaminants (e.g., motion artifact, power line interference) are required for rigorous testing; however, obtaining a large, diverse database of real-life signals and contaminants is a challenging process. By accurately simulating signals and contaminants, researchers are able to more easily create large amounts of data, with known levels of contamination, which can be used for testing and validation of signal quality analysis tools. The Motion Artifact Signal Generation Toolkit allows for the synthesis of motion artifacts using one of three models: 1) autoregressive, 2) Markov chain, and 3) recurrent neural network. Each of these has been prepared for three use-cases: 1) pre-simulated motion artifacts, 2) pre-trained models that can be used to simulate motion artifacts, and 3) training a model using a motion artifact sample and using that model to simulate motion artifacts. The three model types were tested on nonstationary data, exposing some current limitations; specifically, the models' ability to model real-world, non-cyclical data. The recurrent neural network does appears to produce reasonable simulated motion artifact that exhibit similarities, in both the time and frequency domains, to short time segments of real-world motion artifact.