Comparative Analysis of Data Mining Algorithms for Predicting Inpatient Length of Stay

It is well documented that efficient measurement of inpatient length of stay (LOS) can greatly enhance the planning and management of hospital resources. This paper describes the use of a data mining approach to predict the LOS using historical data. The data are first pre-processed using attribute-aggregation, attribute-generalization and relevance analysis procedures to reduce the numbers of attributes and compress the training data set. Two classifiers, Naïve Bayesian and decision tree, are used for predicting LOS. Further a new concept, named “prediction profit”, is introduced to compare the performance of these classifiers. We demonstrate that this new measure of performance can better discriminate between alternative classifiers when compared to existing ones such as the overall accuracy. The empirical tests show that the proposed data mining approach for predicting LOS is a viable one.