Generative Neural Networks for Data Imputation in Longitudinal Epidemiological Studies

Longitudinal epidemiological studies often face challenges with incomplete follow-up and missing data, which can bias results and reduce statistical power. Conventional imputation methods may not adequately capture the complex patterns and dependencies in such multivariate time series data. While more recently developed generative machine learning models offer improved solutions, few methods are available which can handle inconsistently spaced intervals between measurements across long time periods and completely missing time steps, characteristics which are common in real-world studies evaluating long-term health outcomes. This paper introduces a variational autoencoder-based generative neural network designed for imputing partially and fully missing information in irregular time series with extensive missingness. Our approach exploits both correlations between features at a single time step and trends of the same feature over time to reconstruct missing values. Experiments on synthetic data designed to resemble the characteristics of longitudinal epidemiological studies and a case study on a real-world dataset demonstrate the effectiveness of our approach. We show superior performance and parameter stability across varying degrees of missingness and missingness patterns compared to prior work.