Data Mining in Healthcare Information

Data Mining in Healthcare Information Systems

Case Study of a Veterans' Administration Spinal Cord Injury Population

The business problem that initiated the use of Business Intelligence and the original goal in this case were to reduce the length of stay (LOS) and the associated costs for the hospital administration. Since the cost totals were predicted to be at around $7.2 to $9.5 billion per year for the VHA, thorough analysis for the implementation of BI Software was needed.

The data sources used came from the national database, and local operational databases. The national database was in conjunction with the VA Hospital Information Systems and Technology Architecture (VistA), while the nursing and personal information were derived from the hospitals local operations.

The type of data repository used where data warehouse and real-time data. The data warehouse information was not organized and supplied no enhanced value to the project. The process of data cleaning needed to be performed before data could be used for further analysis.

The type of data mining used at the VHA included mainly current data stored locally and remotely. The use of text extractions to determine the nursing diagnoses and intervention was a significant challenge since the wording between nurses meant the same yet was formulated differently. Before data mining could be put in place several pre-mining tasks such as data acquisition, data validation, data aggregation and integration, data visualization, and the length of stay (LOS) had to be performed.

The type of data quality was very low and forced several runs of data cleaning to ensure valid information can be passed from one layer to the next layer of the business intelligence model. These data needed to be visualized and several data, which did not cohere with the expected outcome, where deleted.

The tools used, in this case, for knowledge discovery and data mining where based on artificial neural networks (ANN) and consisted of four different models. All models represented supervised learning models with a known output. The four models of the ANN were dynamic network, prune network, the multilayer perceptron, and the radial basis function network.

The main challenge for its implementation was that data needed to be cleaned so the data mining process can easily retrieve these data and process them according to the required use. All abnormalities needed to be eliminated since the outcome would have been impacted by invalid or erroneous information.