[Paper Review] A Social Network Analysis Framework for Modeling Health Insurance Claims Data

Ana Paula Appel, Vagner Figuerêdo de Santana+3 more|arXiv (Cornell University)|Feb 20, 2018

Healthcare Systems and Technology19 references3 citations

TL;DR

This paper proposes a social network analysis framework to model health insurance claims data by constructing physician networks based on shared patients, enabling the identification of mutual referrals, patient retention patterns, and physician centrality. The approach reveals structural insights into physician behavior, with high consistency across centrality metrics and dense subnetworks indicating key clinical hubs.

ABSTRACT

Health insurance companies in Brazil have their data about claims organized having the view only for providers. In this way, they loose the physician view and how they share patients. Partnership between physicians can view as a fruitful work in most of the cases but sometimes this could be a problem for health insurance companies and patients, for example a recommendation to visit another physician only because they work in same clinic. The focus of the work is to better understand physicians activities and how these activities are represented in the data. Our approach considers three aspects: the relationships among physicians, the relationships between physicians and patients, and the relationships between physicians and health providers. We present the results of an analysis of a claims database (detailing 18 months of activity) from a large health insurance company in Brazil. The main contribution presented in this paper is a set of models to represent: mutual referral between physicians, patient retention, and physician centrality in the health insurance network. Our results show the proposed models based on social network frameworks, extracted surprising insights about physicians from real health insurance claims data.

Motivation & Objective

• To address the limitation of traditional claims analysis that focuses only on providers, by incorporating the physician's perspective and patient flow.
• To model physician relationships using shared patients as a proxy for professional collaboration or referral.
• To identify patterns indicating physician excellence or potential abusive practices through network-based metrics.
• To support health insurance companies in managing provider networks and improving patient care coordination.
• To validate the framework using a real-world 18-month claims database from a major Brazilian health insurer.

Proposed method

• Constructed a physician-physician network using shared patients as a proxy for professional relationships.
• Applied four centrality measures—degree, closeness, betweenness, and eigenvector centrality—to assess physician importance in the network.
• Defined mutual referral as a symmetric referral pattern between two physicians based on shared patient referrals.
• Modeled patient retention as the frequency of return visits to the same physician-patient pair over time.
• Computed network density to assess connectivity within subgroups, especially top-ranked physicians.
• Validated results through expert review with physicians, process analysts, and IT specialists from the partner health insurer.

Experimental results

Research questions

• RQ1 How can physician relationships be modeled from health insurance claims data when direct referral data is absent?
• RQ2 To what extent do centrality measures in the physician network correlate, indicating robustness of network-based insights?
• RQ3 What patterns in patient retention and mutual referral indicate high-performing or potentially problematic physician behavior?
• RQ4 How does the network structure of physicians evolve over time, and what does it reveal about clinical collaboration?
• RQ5 Can social network analysis of claims data detect clinically meaningful subnetworks, such as dense clusters of interconnected physicians?

Key findings

• The four centrality measures (degree, closeness, betweenness, eigenvector) showed high consistency, with 10–20% overlap in the top 100 physicians, indicating robust network insights.
• The top 10 physicians by eigenvector centrality in Q2 2014 were highly consistent across metrics, with 14 of them in the same rank order.
• The network density was 0.001 overall, but rose to 0.6 for the top 100 physicians and 0.9 for the top 40, indicating strong local connectivity.
• A subset of 21 top physicians (Q2 2014) had a network density of 0.933, showing near-complete mutual connectivity.
• The top physicians were geographically clustered, with four from the same region (São Paulo), suggesting location as a key factor in network formation.
• The framework successfully identified clinically relevant patterns, including physician hubs and referral clusters, with real business value for fraud detection and care coordination.