Predicting laboratory testing in intensive care using fuzzy and neural modeling

Laboratory testing is a frequent activity for patients in intensive care units (ICU). Recent studies demonstrate that frequent laboratory testing does not necessarily relate to better outcomes. We hypothesize that unnecessary laboratory testing can be reduced by predicting which tests are unlikely to influence clinical management. Reducing unnecessary tests could reduce morbidity and hospitalization costs. We analyzed an ICU database containing 26,665 patient records at Beth Israel Deaconess Medical Center, Boston, and selected a subset of patients with gastrointestinal bleeding. Database knowledge discovery was applied involving data preprocessing, feature selection, and classification. Conventional soft computing tools such as fuzzy models and neural networks were utilized in this work, combined with statistical and mathematical tools. The input variables included bedside monitor trends, lab tests, arterial/central catheter information, urine collections, transfusions, indexes and scores calculated for the patients. The outcome variable was a binary classification based on falling levels of hematocrit. Feature selection was performed by a bottom-up strategy, maximizing the area under the ROC curve (AUC), the integrated discrimination improvement (IDI) and a multiobjective function primarily pondering the sensitivity of the models. A leave-one-out cross validation process was used to evaluate the overall models´ performance, as well as the additional predictive value of the variables selected. Urine output was selected by all models as the best predictor of useful hematocrit testing. Our results show that it is possible to correctly classify the usefulness of a hematocrit lab test up to 81% of the time by using fuzzy models and neural networks.