Estimating Ventilation Requirements in Critical Care Medicine

H.C.E. Trigg, M. Stevenson, M. Frize, F.G. Solven

World Congress on Medical Physics and Biomedical Engineering *Nice'97.

Introduction Artificial ventilation requires the use of scarce and costly resources. If its use could be estimated, planning of major surgeries and admissions to the intensive care unit (ICU) could provide optimal use of equipment and facilities. Despite ongoing research on estimating mortality, little has been done to estimate duration of ventilation. The main focus of this work has been to estimate outcomes on a single patient basis, rather than for the entire unit, as many scoring systems do.

Methods A feed-forward artificial neural network (ANN) was trained using the back-propagation algorithm to estimate 'duration of ventilation'. The ANN was trained and tested using adult ICU patient records collected at the Dr. E. Chalmers Hospital (1322) patients). Two-thirds of the records were used to train the network, and a third to test its performance. Over-fitting was experienced with the first series of tests. A new approach using weight-elimination was designed and tested on a slightly larger database (1491 records), using 51 input variables. Weight-elimination was turned on and off and results compared. Weight-elimination was then combined with another technique that presented 'high' and 'low' values of continuous and integer-valued medical parameters to a pair of input nodes, rather than presenting them to a single node.

Results The use of weight-elimination improved both the generalisation and the overall performance of a fully-optimised network trained to estimate 'ventilation duration' with the current data set; the weights associated with parameters that had little significance in determining the network's output were driven to zero, simplifying the network's structure and resulting in improved network generalisation and performance. The 'high/low' node data presentation technique facilitated the independent interpretation of high and low values of each input parameter by the ANN model, and hence the weights selected by such a network are expected to be more representative of the true significance of each input parameter in determining such an outcome than those identified by a network which does not employ a 'high/low' node technique.

Conclusion Weight-elimination provided the network with its own means of screening out unimportant variables and eliminated the need for making pre-conceived judgments as to what medical parameters are most instrumental in determining a particular patient outcome. In these experiments, analysing post-operative patients, the network identified five variables (out of 51) as the major effect on estimating this particular outcome. This system represents a new class of 'intelligent monitoring systems' for critical care and is now being applied to other patient environments.