Using Data Mining Techniques to Explore Hospital Coding Practices of Adverse Events in Administrative Datasets

This research evaluates the value of using data mining techniques to identify patterns associated with reported and unreported adverse medical events within hospitals. Identifying the degree to which hospitals reliably report complications is important for surveillance systems and performance measures.;The 2006 National Inpatient Sample---freely available hospital administrative data---was used to create cohorts of patients undergoing elective surgical procedures (e.g., total hip replacement, colon resection, and coronary artery bypass surgery). Public software available from the federal government was used to identify coded complications. The resulting database provided a rich environment to explore patterns that might help detect real, yet unreported medical complications for specific patients.;There were two important methodological steps. First, following the methodology from health services researchers, suspected cases of unreported complications were flagged based on excessive hospital lengths of stay. This involved creating a risk-adjusted model based on chronic conditions, and detecting outliers using statistical quality control methods developed by operations researchers. Second, inpatient discharge records were clustered using text mining tools available in R, an open source statistical and data mining environment. Unlike more traditional healthcare outcomes methodologies that often rely on a small number of risk factors as defined by ICD-9-CM codes, this methodology considers all of the thousands of possible disease codes. First, the full matrix of codes was reduced into a smaller set of "factors" using the singular value decomposition. Second, based on the reduced dimensions, patients were clustered using a variety of algorithms. Third, the clusters were validated by looking at their association with patient records that had: (1) coded medical complications; and (2) no reported complications, but excessively long length of stays.;Although the cluster results and association rules seemed robust by methodological standards, it was not easy to identify unequivocal patterns between resulting data mining patterns and the reported and likely complications. The data mining tools are more useful when there is close teamwork between analysts and clinical domain experts. Regardless of the specific results, I follow the lead of other researchers and optimistically conclude that data mining techniques offer a rich environment to explore coded data available in administrative datasets. With increasing interest in pay-for-performance and transparency, it is likely that administrative data will remain an important data source for creating outcomes reports. As a result, health services researchers may benefit from using data mining tools to help them validate the quality of adverse event reporting.