Referenced in our Newsletter Volume 2, Issue 9 - September 2003
Medical Insurance Claims Analysis
This month's Link Chart of the Month discusses medical insurance claims analysis. This is sample data that will demonstrate the results one can expect to see when using a simple model for medical insurance claims.
The model we are using is shown in Figure 1. Although this sample model is quite simple, a production model would probably include other identifiers (SSN, for instance) and might include additional objects that could serve to provide additional trends. For instance, including diagnostic and/or treatment codes would allow an analyst to detect whether or not certain treatments are preferred by fraudsters.
Our simple model includes claims and information related to each claim. Each CLAIM object is identified (keyed) by an ID generated by the source claims management system. The PHYSICIAN and PATIENT objects are identified by their names. The PATIENT_PHONE object is keyed on the phone number itself and the PATIENT_ADDRESS object uses the Street, City, State and Zip Code as a composite key.
In our model, we will be looking for patients that are sharing either phone numbers or addresses.
We might expect to see this for spouses and dependents, but in general, each individual patient should
have a single address and, perhaps, two or three phones (home, work and mobile).
To begin our analysis, we queried our database for a particular date range and region to make the
quantity of data manageable. The result is shown in Figure 2.
As you can see, we have found a number of networks. In this Figure, each network is laid out with a PHYSICIAN object to the far left of each group, a CLAIM object in the second column followed by the PATIENT object, and finally, the PATIENT_PHONE and PATIENT_ADDRESS objects on the right.
Figure 3 (below) shows a more detailed illustration of this arrangement.
If we examine the large network diagram, the two networks in the upper-left display interesting structures.
In Figure 4, we zoomed in on the six networks in the upper left.
The entire bottom row and the network in the upper-right show normal structures. In each network we have a single physician connected to claims that are, in turn, connected to patients that all have single phones and addresses. Further, none of these patients has more than a single claim. These networks are normal and show no indication of fraudulent activity.
The top-center network shows some interesting crossed lines. Closer examination of this network reveals that these crossed lines are due to patients having filed multiple claims for the period we are investigating. The number of claims is low, and all of the patients have a single phone and address. This network warrants no further investigation.
This brings us to the top-left network. This network shows a very interesting structure. We can see from this structure that some of the patients are connected to more than one phone or address object. This is indicated by the fifth and sixth columns to the right of the network. Additionally, some of the patients have submitted multiple claims for the analysis period. We need to take a closer look at this network.
In Figure 5, we have zoomed in on the network and have rearranged the objects to make a little more sense. We can clearly see four distinct network structures. Each of these structures is detailed below. In the discussion that follows, the PHYSICIAN object is moved closer to the section under discussion and the other sections of the network are hidden for clarity.
The upper section of this network, magnified in Figure 6, is actually a normal network, showing a number of patients who are connected to a single claim and just one phone and/or address. We can discount this section of the network from our analysis.
The next section of the network, magnified in Figure 7, is also indicative of normal activity and shows no indication of fraud. The distinguishing characteristic in this group is that the patients have submitted multiple claims for the period. Since the number of claims is low, we need spend no further time on this section of our network.
The section of our network shown in Figure 8 shows a number of PATIENT objects that share a single PATIENT_PHONE object. Though this could be indicative of fraud, it could also be due to data entry errors. It could also be a large family that uses the same phone number in their patient records.
However, in this case, drilling down on the objects shows that the names are not at all similar. This would suggest that further investigation is in order for this section of the network.
We come now to the final section of our network diagram. In Figure 9 we see a group of patients that share an address. The possible explanations for this condition are the same as with phone numbers: name misspellings or these are members of the same family.
In this case, when we examine this group in detail, we find that these are college-aged patients. In fact, they appear to belong to the same fraternity and live together in the same residence, thus explaining the shared address. This reminds us that due diligence is required to ascertain the true nature of the data.
This leaves us with the section of our network shown in Figure 8. What might we do to extend our analysis? Following are just a few suggestions.
Additional steps might include deleting all of the objects from the full network diagram that are not shown in Figure 8. This will clean up the diagram and allow us to focus on the objects of interest. Next, we could expand our analysis on this network by executing single-level data walks on the objects displayed. This would add additional objects to the diagram for these patients and for this physician.
After each data walk, the objects returned can to be examined to determine if they contribute to our analysis, or if they are "noise" objects. "Noise" objects should be deleted.
Additional analysis activities might also include walking these objects to other data sources (models) or we might plot the objects on a temporal matrix to look for repeating temporal patterns. It may even be necessary to review the original source documentation for the objects in question.
|
