Spatio-temporal graph learning with epidemiological factors for HIV epidemic short-term prediction

Abstract

HIV/AIDS is a major epidemic in the 21st century, with high mortality rates and no effective preventive vaccine. It significantly impacts the economy, mental well-being and health systems and shortens national lifespans. Early detection helps reduce transmission and allocate medical resources effectively. However, predicting outbreaks remains challenging due to the influence of temporal, spatial and epidemiological factors, which complicate the spread of the disease across regions and pose difficulties for predictive models. Very few studies use deep learning models to tackle the HIV epidemic. To address this gap, we suggest using a graph data structure to simulate HIV transmission between neighboring areas and integrate epidemiological factors into this framework. We develop a spatio-temporal graph neural network model to predict short-term infection trends. This model incorporates important factors from HIV modeling, including temporal dynamics, geographic regions, and epidemiological variables such as age groups, career groups, gender groups, risk population groups, and transmission routes within an area. Our approach uses self-attention in the graph architecture to gather node-level information across the infection graph at each step during time series processing. We employ a GRU mechanism to update the graph information over time, allowing for a comprehensive evaluation of transmission probabilities between regions and improving predictive accuracy. Our proposed model was tested on HIV datasets from districts in Ho Chi Minh City, Viet Nam, and demonstrated superior performance compared to existing spatio-temporal models applied to the same dataset.