Using Neural Nets for Decision Support in Prescription and Outcome Prediction in Anticoagulation Drug Therapy

In this paper we consider the use of Artificial Neural Networks (ANNs) in decision support in anticoagulation drug therapy. In this problem domain ANNs can be used to learn the prescribing behaviour of expert physicians or alternatively to learn the outcomes associated with such decisions. Both these possibilities are evaluated and we show how, through the prediction of outcomes that the prescribing practice of physicians may be improved – the ANN can learn to predict the outcome resulting from prescribing levels better than expert physicians. In evaluating the use of ANNs on these regression problems the possibility of ensembling these nets to improve performance is also considered. We show that ensembles of networks do not improve over individual networks, presumably because of the lack of diversity in the outputs of the ANNs.