Overview

Medical "ping ponging" can happen when a patient receives services from multiple medical providers. The Nebraska Department of Insurance defines the process as "The medical provider refers claimants unnecessarily to specific other providers and gets something in return (ping-ponging)."

Sybase defines the process a little differently: "Ping-ponging — the use of a single patient ID to generate billings across multiple providers ".

Detection

In any event, if this practice occurs, it can result in costly and sometimes result in unnecessary claims. However, it can be difficult to detect the practice, especially when there are large volumes of transactions.

A Non-Proprietary Solution

Described below is one potential approach to identifying ping-ponging through the use of non-proprietary software. The data and the extract procedures are contained in a zipped file which can be downloaded.

The key piece of software is that of Christian Borgelt which is free (GPL license) software available from his site. His site also has much more information as to the technical underpinnings of the software.

The example here is based on 2,785,343 actual claim details involving 174,995 patients and 12,981 providers.

An outline of the steps taken follows (note that your procedures will differ)

Step 1 - Using claims data, prepare a list of unique patient/provider combinations. If the data is in a database, this can be done using SQL. (Example output using actual data is provided). If the data is not in a database, then alternative software can be used, or else load it into a database.

Step 2 - Using Excel or equivalent (example workbook provided), take the data from step 1 and build a transaction file which is capable of being analyzed using Borgelt's "apriori" program. Note that the file contains a list of all provider numbers which a patient has seen during the review period. The apriori program requires that the provider numbers must be integers (won't accept alphanumeric text).

Step 3 - Run the apriori program on the data to identify dependencies.

On a 2.2 GHZ PC, processing this information generally takes less than 5 seconds (unless a support level of 0 is specified).

The detail steps taken

Step 1

Ran SQL query on the claims database to extract all unique patient/provider combinations during the period. SQL used was: select provider_number, patient_id, count(*) from claims_db group by provider_number, patient_id order by patient_id. The output was written to a file named extract.txt.

Step 2

Used an Excel macro to prepare the data for analysis. Selection criteria were that only patients seeing three or more providers would be chosen (this is an arbitrary number which can be changed). Output from this extract is a file named text.txt. (Included in zip file). This is the file analyzed by the apriori program.

Step 3

Run the apriori program specifying any parameters needed. In some cases the default values will not be appropriate and thus need to be overridden. There are at least two parameters of interest to the auditor:

1) the confidence level, i.e. what percentage of the cases have this relationship. This is necessary to support statements such as "When a patient sees Provider A and Provider B, there is a 95% chance that they will also see Provider "C".

2) the "support" level. This is a data mining term which means what portion of the population must the transactions constitute in order to be of interest. Since fraud may be involved, the actual percentage may be very low (hopefully), so a "support" level of one or two (i.e. 1% or 2%) is specified in order to have the program analyze most of the transactions. Also this results in additional processing time, it also means that a more thorough analysis will be conducted. Since the program runs unattended, this may not be a big factor...

Step 4

Review the output results. This may involve having a clinician or other professional review the actual results to determine if the relationship identified appears appropriate or not.

Results of the apriori program

The program identified a number of potential instances of "ping-ponging". In some cases, there could be a valid explanation. The suggestion is to rank these relationships by their "confidence", i.e. the probability of occurrence and then review some claims in more detail.

The output of the program specifies each provider number, then the support level and confidence and then the group of related providers. A brief example (only two lines) is shown below, along with an explanation:

Output from apriori

2042 <- 1588 3 6 (1.8, 95.7)
6 <- 3579 314 3 (1.9, 99.8)

The first line indicates that if a patient sees providers 1588, 3 and 6, there is a 95.7% probability that the patient will also see provider #2042.

Similarly, the second line indicates that if a patient sees providers 3579,314 and 3, then there is a 99.8% probability that they will also see provider #6.

Thanks to Chrstian Borgelt for his software (no we are not affiliated with him)!

Main Form