Geometric deep learning enables high-fidelity network imputation for HIV transmission modeling

Accurately mapping social and risk networks is critical for understanding and controlling infectious disease transmission, especially among hard-to-reach populations, such as people who inject drugs (PWID). Yet, empirical sociometric network ascertainment remains challenging and resource-intensive. Geometric deep learning may provide a scalable approach to inferring network structure from individual-level data, but its real-world performance and translation to epidemic modeling remain undercharacterized.

We trained a graph neural network (GNN) to predict injection partnerships from a longitudinal network study of 2512 PWID in New Delhi, India, using demographic, behavioral, and spatial injection-venue features. We compared the GNN with exponential random graph models (ERGMs), evaluated structural similarity between empirical and imputed networks, and assessed validity in an independent PWID network. To examine translational utility, we calibrated a network-based HIV transmission model on either the empirical or GNN-imputed injection network and compared HIV incidence two years after scaling interventions across venues.

The GNN achieves balanced predictive performance (accuracy 60.8%, precision 59.4%, recall 67.9%, F1 63.4%), outperforming ERGMs, and yields an imputed network with structural concordance to the empirical network (spectral similarity 0.87). Incorporating venue data increases accuracy from 51.0% to 60.8%. In the external cohort, the GNN maintains performance (F1: 61.3%) and captures structural changes. In the HIV model, calibrating on the GNN-imputed versus empirical network produces incidence curves that differ by at most 0.4 infections per 100 person-years.

GNN-based network imputation can recover sufficient epidemiologically relevant network structure to preserve conclusions about HIV interventions, illustrating how geometric deep learning can support network-informed epidemic modeling when full sociometric ascertainment is infeasible.