Handling Missing Values in Multivariate Time Series Classification

Multivariate time series data are littered with missing values across many domains. Often the patterns of missing data points reflect conscious decisions by the data collector, and thus may contain correlations that can be used to more accurately classify the data. Previous works propose machine learning models that are capable of detecting these correlations in order to achieve highly accurate results. We review the literature and identify a rich variety of strategies along the machine learning pipeline for handling missing values, including forward imputation, appending additional missing indicators, and others. We also study different deep learning methods for handling time series data, namely GRU and LSTM cells. We then design an experimental study for investigating the relative effectiveness of different state-of-the-art methods for handling missing values to more fully comprehend the intricacies of missing data. For the experimental evaluation study, we utilize MIMIC-III, a publicly available critical care dataset composed of Electronic Health Records (EHR) for over 58,000 hospital admissions collected at Beth Israel Deaconess Medical Center from 2001 to 2012. In particular, our experimental study compares variety of missingness-aware machine learning methods for predicting patient mortality, a benchmark task for this database. We validate in general that extracting information from missing values indeed improves predictive accuracy, and we observe that GRU and LSTM cells perform comparably with the forward imputation method further impacting the resulting accuracy for both models.