Towards Sustainable Compressive Population Health: A GAN-based Year-By-Year Imputation Method

Population health monitoring is a fundamental component of the public health system. Due to the high-cost nature of traditional population-wise health-data collection methods, a class of sparse-sampling-completion algorithms are proposed to exploit the spatio-temporal correlation buried under the observed examples. However, for the population health data, a huge challenge for the state-of-the-art completion methods is the unstationary environment. Specifically, the underlying temporal correlation of the population health data are evolving from year to year. To this end, we propose a GAN-based year-by-year completion framework: uncertainty-aware augmented generative adversarial imputation nets (UAA-GAIN), to address the problem of unstationary environment. To further restrain the error accumulation, we develop a stronger generator as well as a stronger discriminator in the min-max equilibrium. A by-product of the augmented GAIN model allows weighting the difficulty of examples. Inspired by the idea of curriculum learning, a better training schedule is implemented in the proposed framework. We evaluate the proposed method on three real-world chronic disease datasets and the results show that UAA-GAIN outperforms other baseline methods in various settings.