Estimating county health indices using graph neural networks

Population health analytics is fundamental to developing responsive public health promotion programs. A traditional method to interpret health statistics at population level is analyzing data aggregated from individuals, typically through telephone surveys. Recent studies have found that social media can be utilized as an alternative population health surveillance system, providing quality and timely data at virtually no cost. In this paper, we further investigate the use of social media to the task of population health estimation, based on a graph neural network approach. Specifically, we first introduce a graph modeling method to construct the representation of each county as a graph of interactions between health-related features in the community. We then adopt a graph neural network model to learn the population health representation, ended by a regression layer, to estimate the health indices. We validate our proposed method by large-scale experiments on Twitter data for the task of predicting health indices of the US counties. Empirical results show a significant correlation with the reported health statistics, up to a Spearman correlation coefficient (ρ) value of 0.69, and that our graph-based approach outperforms the existing methods. These promising results also suggest potential application of graph-based models to a range of societal-level analytics tasks through social media.