Physiological Waveform Imputation of Missing Data using Convolutional Autoencoders

Daniel Miller, Andrew Ward, Nicholas Bambos, David Scheinker, Andrew Young Shin

Abstract: Machine Learning has great potential to improve automated real-time patient diagnostics. For the majority of machine learning algorithms, taking advantage of this potential requires a complete dataset with no missing data. In practice, missing values are estimated using a variety of imputation methods in the pre-processing stage. However, with time-series data, and physiological waveforms in particular, imputation can be difficult due to the unique patterns and shapes of each waveform, as well as how these patterns vary between patients, and even for a single patient over longer durations. We demonstrate that deep learning techniques can reconstruct missing data using patient-specific patterns present in the non-missing portions of the waveform. Using convolutional neural network (CNN) autoencoders trained on 288 15-minute samples from each of 138 pediatric patients, we develop a generalizable model to analyze and extract information from arbitrary physiological waveforms, and use this model to develop methods for mid-channel missing time-series imputation. We further show that the autoencoder can be used to compress the dense physiological waveforms to a low-dimensional representational space.