SIFT: A Deep Network for Irregularly-Sampled Multivariate Time Series Public

Downloadable Content

open in viewer

Due to the way hospitals record information, patient health records are composed of sparse and irregular multivariate time series data, with long breaks between visits and varying physiological variables recorded each time. Such data can provide valuable machine learning opportunities for improving patient care. Since standard Recurrent Neural Networks (RNNs) rely on fixed-length vector inputs, existing methods work by first converting irregular data into regular. However, such conversion requires a priori specifying a time interval to align values on, at the risk of negatively augmenting the original data and/or meaning of the data expressed by such irregularity. To address this open problem, we propose a novel end-to-end Reference-Network architecture (SIFT) for irregular multivariate time series to predict patient health outcomes. The architecture is composed of a Reference Network, an Interpolator Network, and a Discriminator Network. SIFT is able to extract the most informative data by adjusting the sampling interval to filter out noise and pass only the most useful signals to the classifier. SIFT adapts the sampling strategy (time interval data alignment) for values at which time points to interpolate, this way paying more or less attention to different time windows. The reference network is rewarded for sampling from the discriminative signal, and penalized for sampling from noisy data. We validate our approach on a range of recently proposed models, including GRU-D and IPN. Our experiments demonstrate that SIFT outperforms five comparable imputation and interpolation methods in various settings, in both AUC and Accuracy.

  • etd-4921
Defense date
  • 2020
Date created
  • 2020-12-11
Resource type
Rights statement


In Collection:


Permanent link to this page: