Prediction of the length of stay of patients in the neuro-critical care unit using data mining techniques

Today, cost reduction and resource planning play an important role in managing hospitals. Hospital,s intensive care departments have the most expensive beds. The prediction model for length of stay is a tool for optimal management of beds and scarce resources in intensive care units. Our goal in this study is to provide models for predicting the patient length of stay in neurocritical care unit using data mining techniques. In this study, RapidMiner data mining software was used for modeling in order to classify and construct the prediction model for patients admitted to the neuro-critical care unit of Loghman Hakim Hospital in Tehran. The data of this study were extracted from 592 patients admitted to the intensive care unit between 94-97. Artificial neural network, K-nearest neighbors, decision tree and random forest algorithms used to classify the patients. Finally, the confusion matrix was obtained to calculate accuracy.

Results: The findings of this study indicate that the type of surgery and the pneumonia as a complication have the greatest impact on the length of ICU stay. Also, the accuracy of the algorithms used to construct the prediction model was: decision tree 84.28%, random forest 83.96%, artificial neural network 83.79%, and the K-nearest neighbors, 77.90%. Models of four techniques used in this study were able to predict the length of ICU stay. But the findings from the confusion matrix showed that the Decision trees with accuracy 84.28% had a relatively better performance among the techniques studied and the rules extracted from the decision tree could serve as a model for predicting the patient's stay in the neuro-critical care unit to be used