Abstract: Artificial neural networks are trained to predict mortality for patients undergoing heart surgery. A database with patients from Lund is used for training and test. The results, presented as ROC areas, are compared to the commonly used risk model EuroSCORE and to linear models trained on the Lund database. No statistically significant improvement is found compared with EuroSCORE but it is confirmed that neural networks out-perform the linear models. It is proposed that combining different good risk models may lead to improved generalization performance. Also, different ways of weighting ensembles of neural networks are explored. Remarkably, no increase in performance is found by optimizing the weights and a possible explanation is provided.