Using administrative data to identify naturally occurring networks of physicians

Abstract

Background: Physicians naturally form networks. Networks could form a rational basis for Accountable Care Organizations (ACOs) for defined populations of Medicare beneficiaries.

Objectives: To use methods from network science to identify naturally occurring networks of physicians that might be best suited to becoming ACOs.

Research design, subjects, and measures: Using nationally representative claims data from the Medicare program for CY 2006 on 51 hospital referral regions (HRRs), we used a network science-based community-detection algorithm to identify groups of physicians likely to have preestablished relationships. After assigning patients to networks based upon visits with a primary care physician, we examined the proportion of care delivered within communities and compared our results with potential ACOs organized around single hospitals.

Results: We studied 4,586,044 Medicare beneficiaries from 51 HRRs who were seen by 68,288 active physicians practicing in those HRRs. The median community-based network ACO had 150 physicians with 5928 ties, whereas the median hospital-based network ACO had 96 physicians with 3276 ties. Among patients assigned to networks via their primary care physicians, seventy-seven percent of physician visits occurred with physicians in the community-based networks as compared with 56% with physicians in the hospital-based networks; however, just 8% of specialist visits were to specialists within the hospital-based networks as compared with 60% of specialist visits within the community-based networks. Some markets seemed better suited to developing ACOs based on network communities than others.

Conclusions: We present a novel approach to identifying groups of physicians that might readily function as ACOs. Organic networks identified and defined in this natural and systematic manner already have physicians who exhibit close working relationships, and who, importantly, keep the vast majority of care within the networks.

Figure 1. Example of community-network construction for a very small HRR

We start from the claims data, which can be represented as a bipartite network connecting physicians to their patients (not shown). This network is then projected to a unipartite or one-mode network, which in this case consists of physicians only (the nodes), and ties connect two or more physicians if they have shared patients. This typically results in a very tightly knit network as shown in panel A. Each network tie is then assigned a “strength”, which is a number that quantifies the number of patients any two physicians share. In order for a tie to exist between them, by definition the physicians must share at least one patient, but many share several more. Because most of the connections between physicians turn out to be weak, corresponding to just one or two shared patients, we filter the network by keeping, for each physician, only the strongest 20% of their ties. Note that for a tie to be retained between any two physicians, the tie has to fall in the top 20% for each physician. The tie thresholding process results in a sparser network with arguably only the more influential connections left in place (panel B).. Finally, we apply the method of modularity maximization to detect network communities, which are sets of densely connected nodes. In panel C, we show the outcome of community detection, and color the nodes based on their community assignments. For example, the green nodes in the upper left corner form a fairly densely connected group with few connections to nodes in other communities. This is an example of a community-network.

Figure 2. Depiction of Hospital Networks and Community Networks in Tallahassee, FL (Panels A and B), and Norfolk, VA (panels C and D)

The figure depicts networks in the same HRRs based on hospital affiliation (panels A and C) and community detection (Panels B and D). Colors indicate hospital/community affiliation.