H3: A Healthcare Three-Hop Index for Physician Referral Network Prediction

Accurately predicting physician referral links is critical for optimizing care coordination and reducing fragmentation in healthcare. However, existing methods—from triadic closure heuristics to graph neural networks (GNNs)—struggle with the unique properties of physician referral networks: sparsity, disassortative degree mixing, and hub-dominated topologies. To address this, researchers from (institutions implied) propose H3, a healthcare three-hop index. H3 explicitly models indirect referral pathways through intermediate physicians, applying degree-based normalization and a redundancy penalty to filter out hub-mediated noise. The method is designed to be both computationally efficient and interpretable.

Evaluated on Medicare Physician Shared Patient Patterns data, H3 was tested under two regimes: within-period prediction (recovering contemporaneous links under sparsity) and cross-period prediction (testing robustness to temporal shifts). In both cases, H3 consistently outperformed classical heuristics (e.g., common neighbors, Adamic-Adar) and deep learning baselines (e.g., GNNs). Crucially, unlike black-box neural networks, H3 produces fully decomposable predictions—each predicted link can be traced back to specific intermediary physicians. This transparency is vital for clinical deployment, where trust and auditability are paramount. The paper also includes 13 pages, 4 figures, and 7 tables detailing the methodology and results.