Prediction of ventilation requirements in an intensive care unit

Neural networks are applied to the problem of predicting patient requirements for mechanical ventilation in an intensive care unit (ICU). Two classes of neural network are considered: generalized single-layer networks (GSLNs) trained using a technique known as iterative fast orthogonal search with dynamic model resizing (IFOS-DMR), and feedforward sigmoid-activated multilayer perceptrons (MLPs) trained using backpropagation with weight elimination (BP-WE). It is found that (1) The GSLNs and MLPs implemented have similar correct classification rates on the test data, and (2) in contrast with BP-WE, IFOS-DMR accomplishes automatic determination of an appropriate model structure without reference to the test data. This work is undertaken as part of the Medical Intelligent Decision Aid Systems (Medical IDEAS) project