Use of a neural network as a predictive instrument for length of stay in the intensive care unit following cardiac surgery.

A patient's intensive care unit (ICU) length of stay following cardiac surgery is an important issue in Canada, where cardiovascular intensive care resources are limited and waiting lists for cardiac surgery exist. A predictive instrument for ICU length of stay could lead to improved utilization of existing ICU and operating room resources through better scheduling of patients and staff. We trained a neural network with a database of 713 patients and 15 input variables to predict patients who would have a prolonged ICU length of stay, which we defined as a stay greater than 2 days. In an independent test set of 696 patients, the network was able to stratify patients into three risk groups for prolonged stay (low, intermediate, and high), corresponding to frequencies of prolonged stay of 16.3%, 35.3%, and 60.8% respectively. The performance of the network was also evaluated by calculating the area under the Receiver Operating Characteristic (ROC) curve in the training set, 0.7094 (SE 0.0224), and in the test set, 0.6960 (SE 0.0227). We believe the trained network would be a useful predictive instrument for optimizing the scheduling of cardiac surgery patients in times of limited ICU resources. Neural networks are a new alternative method for developing predictive instruments that offer both advantages and disadvantages when compared to other more widely used statistical techniques.