Revealing Healthcare Networks Using Insurance Claims Data


As I noted in my post last week, every healthcare accountable care organization in the United States is trying to understand provider networks. Common questions include:

  • What is the “leakage” from our network?
  • What medical practices should we acquire?
  • What are the referral patterns of providers within the network?
  • Does the path that a patient takes through our network of care affect outcomes?
  • Where should we build the next outpatient clinic?

Much of this analysis is being done by using insurance claims data, and this post is about how such data is turned into a provider network analysis.  Here, I’ll discuss how billing or referrals data is turned into graphs of provider networks.  Most of us are now familiar with social networks, which describe how a group of people are “connected”.  A common example is Facebook, where apps like TouchGraph that show who you are friends with, and whether your friends are friends, and so on.  These networks are build with a simple concept, that of a relationship.

To describe a physician network, we first make a table from claims data that shows which physicians (D) billed for visits or procedures on which patients (P).  This is shown in the figure below.  Next, we tally which physicians billed for the seeing the same patient, and how many times, giving a common billing matrix.  The billing does not have to happen at the same visit or for the same problem, just over the course of the measurement period. Notice that the matrix is symmetrical, with the diagonal giving the total number of patient encounters for each doctor.  This type of matrix is referred to as a distance or similarity matrix.


The provider network graph plotted from the above example shows the network relationship between four doctors.  The size of the circle shows total number of patients billed for by that doctor, and the width of the line shows the strength of the shared patient connection.


Now, if we have this data for a large network, we can look at a number of measures using standard methods.  In the above example, we can see that the two orange providers are probably members of a group practice, sharing many of the same patients and referring to many of the same providers. See this humorous post by Kieran Healy identifying Paul Revere as the ringleader of the American Revolution using a similar analysis!  Providers in red are “out-of-network”, and with connections to a single in-network physician.  However, the graph itself does not reveal the reason that these out-of-network providers share patients with the in-network provider.   It could be that the out-of-network group offers a service not available within the network, such as gastric bypass, pediatric hepatology, or kidney transplantation.

It is not difficult to see that you could create network representations using many types of data.  Referral data would allow you to add directionality to the network graph.  You could also look at total charges in shared patients, as opposed to visits or procedures, to get a sense of the financial connectedness of providers or practices.  Linking by lab tests or procedures can show common practice patterns.  Many other variations are possible. Complexity of the network can increase with the more providers and patients in the claims data you have.

These simple graphs are just the beginning.  Couple to network graph with geospatial locations of providers, and you add another layer of complexity.  Add city bus routes, and you can see how patients might get to your next office location.  Add census data, and you can look at the relationship between medical practice density, referral patterns, and the average income within a zip code area.  The possibilities are incredible!

So why is this big data?  To build a large and accurate network, you  need to analyze millions of insurance claims, lab tests, or other connection data.  Analyzing data of this size requires large amounts of computer memory and, often cluster computers, and distributed computing software such as Hadoop (more on this in a future post).  We owe a very large debt to the “Healthcare Hacker” Fred Trotter, who created the first such open source, very large, network graph from 2011 Medicare claims data for the entire United States, called DocGraph. The dataset can be downloaded from NotOnly Dev for $1 here.  This graph has 49 million connections between almost a million providers.  Ryan Weald created a beautiful visualization of the entire DocGraph dataset, which I will leave you with here.


How Much Unstructured Big Medical Data Is There In The EMR?

EMR saladHow much unstructured big data is there in the EMR? Unstructured data is data that doesn’t fit into neat columns on a spreadsheet, or fields and look-up tables in a database, like the narrative text in an HPI. It used to be that we sat down with a pen and the paper chart, and wrote our progress notes in the office and in the clinic. Or, we dictated the notes, which were transcribed. But with the advent of the EMR, templates have crept in, as well as the wide-spread and controversial practice of copying and pasting text from a previous  encounter (see the recent NYT article).

This is  interesting in a quirky way. As physicians, nurse practitioners, and other providers have become reluctant data entry clerks, they use many shortcuts so that they will have time to take care of the patients, including templates with stylized or constrained vocabularies, self-generated “smart phrases”, and patient-specific narratives that can be recalled and modified.  The remainder of the note is populated with structured data already in the system (labs, test results, x-ray results).  Because medical changes are often not so dramatic from one day to the next,  the actual novel unstructured information content from one note to the next may only be a tiny fraction of the total bytes, and probably the change between the current and previous note may carry as much information than the actual content.  But, when people get hurried or sloppy, old information gets carried along that is no longer current, but has not been changed in the notes.  So, the key information extraction question is identifying the true changes, separating them from relatively static or outdated data that is carried along, and extracting the novel information.

How is this relevant to big data analytics in medicine?  If much of the content is captured by a stylized vocabulary, and filled with structured data already present in data tables, how much independent information will there be in a medical note?  If the data has dependencies because of this stylized nature and controlled vocabularies, how does this impact data mining and statistical analytics.  I am not sure if this type of problem has a formal technical term in machine learning, but if not it is likely to get one soon!