Incorporating network structure improves HIV risk prediction: a graph-based machine-learning study among men who have sex with men in Bali, Indonesia

Abstract Background Identifying individuals at high risk of HIV infection is particularly challenging in populations where transmission is shaped by social and relational networks. Most machine learning approaches rely on individual-level characteristics and do not explicitly model network dependence. Graph convolutional networks (GCNs) enable integration of network structure with individual attributes, but their added value in applied HIV surveillance remains insufficiently evaluated. Methods We represented the study population as a graph, with nodes denoting individuals and edges encoding observed network relationships. Individual-level features were incorporated within a semi-supervised node classification framework. A two-layer GCN was trained to jointly model node attributes and network topology and was compared with logistic regression, random forest, and XGBoost models using individual-level features alone. Discriminative performance was assessed using the area under the receiver operating characteristic curve (AUC) and distributional separation of predicted risk scores. Results The GCN model achieved a high discriminative performance with an AUC of 0.983, substantially outperforming logistic regression (AUC = 0.778), random forest (AUC = 0.881), and XGBoost (AUC = 0.904). Predicted risk scores differed markedly between HIV-negative and HIV-positive individuals, with a mean predicted risk of 0.864 among HIV-negative nodes and 0.082 among HIV-positive nodes. Median predicted risks were 0.981 for HIV-negative and 0.006 for HIV-positive individuals. Targeting analysis indicated that prioritizing the top 20% of individuals ranked by GCN-predicted risk would identify more than four-fifths of HIV-positive cases while requiring substantially fewer outreach resources compared with traditional models. Conclusion Incorporating relational structure through graph convolutional learning not only improves predictive accuracy but also offers meaningful gains in operational efficiency, supporting more precise allocation of limited prevention resources in network-driven HIV epidemics.