Addressing Time-series Signal Quality in Healthcare Data

Abstract

Healthcare data time-series signal quality assessment (SQA) plays a vital role in the accuracy and reliability of machine learning algorithms to analyze health metrics. However, these signals are often corrupted with different kinds of noises and artifacts, including Baseline Wander, Muscle Artifacts, Powerline Interference, and Equipment Failure. This can lead to vital, potentially deadly, errors in the medical domain. This can include inaccurate calculation of basic health features like Heart Rate, clinical alarm burnout from bedside monitors, as well as disrupting general downstream machine learning tasks. While some work has been done in the area of open-source signal quality analysis in general, there are very few open source implementations of signal quality analysis frameworks that attempt to reproduce and expand on existing results on open source datasets.

First, we propose an open-source implementation of signal quality indices (SQIs) for analysis of electrocardiogram (ECG), plethysmography, and more. We aim to codify and reproduce SQIs and results from The Physionet Signal Quality Classification 2011 Challenge. We show that these SQIs may be used for signal quality outlier detection in a real world clinical dataset from University of Pittsburgh Medical Center (UPMC). Secondly, in the case of another common healthcare SQA issue: ECG denoising, we compare Wavelet, EMD, and Convolutional Autoencoder denoising techniques. We show that Convolutional Autoencoder denoising performs the best on the open MIT-BIH Arrhythmia Noise Stress Test dataset, and evaluate it on the UPMC dataset. We also perform a case study on ECG denoising for real vs artifactual alert classification. To our knowledge, we are the first to provide an open source implementation of these two SQA tasks that is validated on public datasets. Ideally, this work serves as an accessible, open source, toolkit for signal quality analysis and ECG denoising.