Globally, people who inject drugs (PWID) continue to experience some of the most explosive HIV epidemics. Network-based approaches are a powerful tool to understand transmission and implement interventions efficiently; however, detailed social network studies are limited and the non-Euclidian nature of networks poses analytical challenges. Advances in neural architectures in deep learning, namely graph neural networks (GNNs), offer new analytical tools to better elucidate network patterns to identify optimal intervention strategies.
We recruited a cohort of 2,512 PWID in New Delhi, India using a name generator network recruitment method that captured social (injection partners) and spatial (injection venues) network information. Biometric data was used to establish cross-network linkages. Longitudinal HIV incidence from November 2017 – March 2020 was 21.3 per 100 person-years (p-y). The Louvain algorithm was used to identify communities in the network i.e., groups of PWID and places they inject that are more connected to each other than other groups in the network. Neural Overlapping Community Detection (NOCD) using GNNs was used to determine community overlap.
PWID reported injecting in 181 different injection venues across a diameter of over 20 km in New Delhi and formed one large network when accounting for social and spatial ties. Injection networks were highly dynamic (75% gained/lost a partner over a median 12 months of follow-up), but greater stability was observed in spatial networks (48% reported changing venues). Injecting at the highest risk venue (#40 in Fig) was the strongest predictor of incidence (adjusted incidence rate ratio [AIRR]= 3.11) and also exhibited the highest network stability (9% of those who injected at the venue switched venues). Seven distinct communities of PWID were identified. HIV incidence ranged from 0 – 40.9 per 100 p-y across these communities, with the highest incidence observed in community 3 (Fig). NOCD revealed that while only 2 of 2,512 person nodes overlapped >2 communities, 8 of 181 spatial nodes overlapped 6 of the 7 communities, suggesting that these 6 communities could be reached by rapidly scaling-up services in just one of these 8 venues.
In this setting with explosive HIV incidence, deep learning methods suggest network-based interventions that target spaces, which may represent unmeasured network connections, could be the optimal strategy to interrupt transmission among PWID while conserving resources.