Accurate detection of viral molecular transmission clusters is critical to the respond pillar of the Ending the HIV Epidemic initiative, which aims to reduce HIV transmission by 90% in ten years. Previous analysis with high data completeness has shown that artificially lowering molecular surveillance completeness can preclude cluster detection and tracking of HIV outbreaks. Here, we investigate whether lack of access to the entire dataset of all people living with HIV would influence the accuracy of reconstructed molecular transmission networks and consequently lower effectiveness of public health interventions.
We analyzed viral genomic data of 12,818 individuals with diagnosed HIV (2012-2019) from Houston Health Department surveillance data with 52% completeness. We used HIV-TRACE to reconstruct a transmission network using first reported pol sequence and identified priority clusters based on the CDC-specified criteria. To understand the impact of lower-completeness, we performed 10-fold random subsampling by removing records (in 10% decrements) without replacement. Next, we subsampled based on node influence. Expected Force (ExF) computes spreading power of individual nodes via adopting a relative influence of different walk and walk lengths based on local connectivity in a network. We computed ExF for each node in the network. Then we removed nodes with highest or lowest ExF from the full dataset. We reconstructed the network twice (0.5% & 1.5% genetic distance) for all completeness levels to measure the impacts for the above subsamples.
Randomly subsampled data, the detection rate of priority clusters decreased in low completeness (Figure 1). We also observed that the number of singletons expanded up to 95% of all sequences with lower completeness. Next, we compared detection of priority clusters in networks where we filtered out all high ExF (influence) nodes and only 4.7% of priority clusters were detected in comparison with full data. We repeated the experiment with removal of lowest ExF nodes, yet 60% of priority clusters were detected with nearly a quarter of sequences which is far greater than random subsampling detection.
Our results indicate that exhaustively expanding molecular surveillance data to 100% sequence data completeness will have diminishing returns on identification of priority clusters. Network metrics like node influence should be considered to optimize cluster detection.
