The study investigates how physician professional networks impact medical referral patterns between specialist care (SC) and primary care (PC) by utilizing large-scale healthcare data and machine learning methods. However, referral processes are usually hampered by administrative inefficiencies, coordination challenges, and communication gaps that can cause delays, fragmented care, and duplication of tests. Previous research highlights the significance of physician relationships in referral decisions; the underlying structural drivers of these patterns have not been adequately quantified.
The study addressed this gap with a new analytical model that combines graph neural networks (GNNs) and professional network characteristics. It aims to recognize the position of physicians within professional networks analyzed through embeddedness, centrality, and shared affiliations that may impact the referral formation process. The research offers methodological contributions through the use of empirical value by improving referral prediction and graph-based embeddings and through conceptual insight by connecting model outcomes to social and organizational factors like collaboration, trust, and institutional proximity.
This analysis utilized a large dataset from a Portugal-based private medical provider that covers over 5 years of electronic health records (EHRs), including 3,632 physicians, 1.48 million patients, and 12 million consultations. 2 types of networks were established: one is a professional network based on institutional affiliations and shared training, and another one is a referral network based on the transition of patients from PC to SC within one month. This environment is specifically suitable since physicians still have autonomy in making referrals, and observed trends are likely to be indicative of professional relationships instead of administrative limitations.
Exploratory analysis revealed that the majority of patients visit a small number of physicians and a few with complicated conditions visit numerous physicians. The workload of physicians showed significant concentration, and referrals followed a hub-and-spoke pattern dominated by highly connected specialists. The study utilized three GNN models that are Attri2Vec, GraphSAGE, and Node2Vec to model referral patterns differing in their use of attributes and network structure. Key characteristics include eigenvector centrality, degree centrality, and betweenness centrality.
Integrating professional network data improves the referral prediction among all models, with Attri2Vec performing best and GraphSAGE showing the highest gain, emphasizing the significance of relational context. Referral behavior reflects shared affiliations, trust, and previous interactions with central physicians, attracting more referrals and improving coordination through bridging roles; this may also cause concentration among a few professionals.
The study highlights that network-aware analytics can improve healthcare efficiency and equity by determining referral bottlenecks, key coordination hubs, and underused specialists from a practical perspective. Policy measures like monitoring systems and balanced referral strategies may help in reducing gaps, although caution is required to avoid supporting existing biases. The study emphasized that professional networks play a major role in referral behavior, improve predictive modeling, and help conceptualize healthcare as a social and organizational system. Even with these limitations, the study highlights the value of network-based approaches and calls for future research focused on fairness, diverse settings, and patient outcomes.
Reference: de Brito Duarte R, Han Q, Soares C. Dissecting medical referrals in health services: role of physician professional networks. BMC Health Serv Res. 2026. doi:10.1186/s12913-026-14099-9
