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

Area of Science:

• Epidemiology
• Network Science
• Machine Learning

Background:

• Mapping social and risk networks is crucial for controlling infectious diseases, especially in hard-to-reach populations like people who inject drugs (PWID).
• Traditional network ascertainment is challenging and resource-intensive.
• Geometric deep learning offers a potential scalable solution for inferring network structure from individual data.

Purpose of the Study:

• To evaluate the performance of a graph neural network (GNN) for inferring injection partnerships among PWID.
• To compare GNN performance against traditional methods like exponential random graph models (ERGMs).
• To assess the translational utility of GNN-imputed networks in an HIV transmission model.

Main Methods:

• Trained a GNN using demographic, behavioral, and spatial venue data from a longitudinal study of 2512 PWID in New Delhi.
• Compared GNN predictions to empirical networks and ERGMs.
• Validated the GNN on an independent PWID cohort and used imputed networks in an HIV transmission model.

Main Results:

• The GNN achieved balanced predictive performance (F1 score 63.4%), outperforming ERGMs.
• The imputed network showed high structural concordance with the empirical network (spectral similarity 0.87).
• HIV transmission models using GNN-imputed networks produced comparable incidence curves to those using empirical networks.

Conclusions:

• Graph neural network-based imputation effectively recovers epidemiologically relevant network structures.
• Geometric deep learning can support network-informed epidemic modeling, even when full network data is unavailable.
• This approach enhances the feasibility of using network data for public health interventions.