Features

First INFORMS Data Mining Competition
Sponsored by the INFORMS Data Mining Section

The Data Mining Section of the Institute For Operations Research and the Management Sciences (INFORMS) announces the First INFORMS Data Mining Competition: The Antibiotic Protocol Case, held in conjunction with the INFORMS National Meeting in Washington, D.C., October 12-15, 2008. This contest highlights the growing overlap between the INFORMS and KDD communities, both in terms of the increased use of optimization techniques in data mining, and in the increased emphasis on operational and managerial applications of KDD.

This contest will focus on a managerial problem in health care, using real data to design a health care management strategy.

Problem Description:

Medicare has announced that it will no longer reimburse hospitals for errors, and for nosocomial infections. To reduce the occurrence of resistant infections, one 1000-bed hospital wants to implement a protocol to prevent them. High-risk patients admitted for elective surgery will be identified. They will be admitted to the hospital 24 hours prior to surgery (the usual protocol has them admitted after surgery) and placed on IV vancomycin. The antibiotic will be continued until discharge.

Risk is not uniform and is based on a combination of patient demographics, diagnoses, and procedures. Further, the cost to treat infection is not uniform across patients. A superficial wound generally will not add to the overall length of stay. A deep skin wound required ten days to three weeks of IV antibiotics (the date of discharge will vary). An infection in the bone requires six weeks of antibiotics, and carries the additional risk to the patient of limb amputation or death. A resistant infection in the lungs is life-threatening and the patient will be moved to ICU where the daily costs to treat increase substantially.

The Challenge:

Contestants will be given two years of patient data, including whether or not the patient contracted an infection during a surgical procedure, and the cost to treat that infection. Near the end of the contest, a set of test patients will be distributed. The contestants will determine an optimal strategy for choosing patients from the test group to minimize the total cost of medication.

The contest will consist of two parts:

Part 1 (classification): Submit a list of patients who will be diagnosed with MRSA.

Part 2 (policy): Develop and justify a realistic cost model (including cost of prophylactic treatment, cost of MRSA treatment, and probabilities from your predictive model), and use it to maximize the total cost savings of the proposed strategy.

Parts 1 and 2 will be judged separately, and a prize will be awarded for each. Part 2 is optional; contestants are free to submit Part 1 alone, or both parts. Part 1 will be judged based on the AUC of your predictions on the test data. Part 2 will be judged based on a written description of your cost model and the resulting savings.

Important Dates:

• April 1, 2008: Training set released.
• August 1, 2008: Test set released.
• September 1, 2008: Entry submission due.
• October 12: Winners announced at the INFORMS conference.

Questions should be submitted to the contest blog at http://informsdataminingcontest.blogspot.com/.

Contest committee:
Nick Street, University of Iowa: nick-street@uiowa.edu
Patricia Cerrito, University of Louisville: pcerrito@louisville.edu
Vijay Desai, SAS: Vijay.Desai@sas.com
Tom Au, AT&T: sau@research.att.com