Authors
Ngakan Putu Anom Harjana, Pande Putu Ida Oktayani, Brigitta Dhyah Kunthi Wardhani, Pande Putu Januraga
Published in
BMC medical informatics and decision making. Volume 26. Issue 1. Aug 20, 2026. Epub Aug 20, 2026.
Abstract
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., and is sensitive to how the network itself is constructed and how node features and network structure are kept complementary.
We analysed survey data from 186 MSM recruited by snowball sampling in Bali, Indonesia. We constructed a bipartite graph in which the 186 respondents and their 355 reported sexual partner codes form separate node sets (541 nodes total, 369 edges). Respondent node features were restricted to 32 individual-level variables, explicitly excluding partner-aggregate behavioral features, to preserve the complementarity assumption underlying GCN theory: the network carries structural information (who shares a partner with whom) and the features carry individual information. A two-layer GCN was compared with logistic regression, random forest, and XGBoost, and a validation-weighted ensemble of all four models was implemented and reported. Leakage was assessed via an inductive check (test-node edges removed at inference), 5-fold component-aware cross-validation, and a sensitivity analysis re-introducing status-adjacent variables.
Under the bipartite design, the GCN achieved a test AUC of 0.833, outperforming logistic regression (0.707), random forest (0.760), XGBoost (0.620), and the validation-weighted ensemble (0.787). The inductive check confirmed the advantage was not due to label leakage (AUC 0.847 with test-node edges removed; Δ=+0.013). Five-fold cross-validated pooled out-of-fold AUCs ranged from 0.597 to 0.654 across baselines (GCN: 0.624). Predicted risk was higher among HIV-positive, with mean P(HIV-positive) was 0.513 among HIV-positive vs 0.415 among HIV-negative individuals, targeting the top 20% by predicted risk captured 31.1% of positive cases.
When the network is constructed as a bipartite respondent-partner graph with individual-only node features (preserving complementarity with network structure), and evaluation is conducted with component-aware splitting, an inductive leakage check, and cross-validation, the GCN achieves a genuine and meaningful discriminative advantage over conventional classifiers for HIV risk prediction among MSM in Bali. The key methodological requirements, such as verifiable edges, complementary features and structure, and component-aware evaluation, are described and can be adopted in future network-informed HIV surveillance analyses.
PMID:
43040506
Bibliographic data and abstract were imported from PubMed on 17 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 9
- Comments 0