Deep Imputation of Missing Values in Multivariate Time-Series Health Data

The imputation of missing values in multivariate time series data has been explored using a few recently proposed deep learning methods. Evaluations of these state-of-the-art methods have been limited to one or two data sets, low missing rates, and completely random missing value types. These limited experimental conditions did not comprehensively benchmark imputation methods on realistic data scenarios with varying missing rates and not-at-random missing types. To address these limitations, this thesis work took a data-centric approach to benchmark state-of-the-art deep imputation methods across five time series health data sets and five experimental conditions. Our extensive analysis revealed that no single imputation method outperformed the others across all five data sets. The imputation performance depends on data types, individual variable statistics, missing value rates, and types. In this context, state-of-the-art methods jointly performed cross-sectional (across variables) and longitudinal (across time) imputations of missing values in time series data. However, variables with high cross-correlation were better imputed by cross-sectional imputation methods alone. In contrast, the ones with time series sensor signals were better imputed by longitudinal imputation methods alone. The findings of this study emphasized the importance of considering the specifics of the data when choosing an appropriate missing value imputation method for multivariate time series data