Necessity of laboratory blood tests in intensive care unit using data mining

Reducing unnecessary lab tests is an essential issue in intensive care unit (ICU). In this paper we analyze lab tests ordered for ICU patients using data mining methods. The selected dataset is extracted from Multi-parameter Intelligent Monitoring in Intensive Care II (MIMIC-II) database. Calcium test is selected as the target test which is one of the frequent tests for gastrointestinal bleeding patients. We labeled samples as necessary or unnecessary tests, and divided them in upper and lower GI categories. Five classification techniques, namely Fuzzy TS model, SVM with RBF kernel, Decision Tree, MLP neural network and KNN, are used to predict necessary or unnecessary lab tests. Sensitivity, specificity and mean class weighted accuracy (CWA) are used as performance measures for model evaluation. The best sensitivity and CWA is achieved by fuzzy TS model for Upper GI patients. For lower GI patients, SVM is slightly better than fuzzy TS model in both sensitivity and CWA measures. Results show the ability of classification models to be exploited as a part of CDSS for reducing unnecessary lab tests.