Evaluation of surgical risks by means of neural networks in the presence of uncertainties

Abstract Surgical risks in elderly patients are often rather high and therefore a need exists for preoperative tests that are able to predict the postoperative risk of mortality. In most cases no test has been found to be completely able to predict postoperative severe complications although reasonable results can be obtained by employing a multi-test approach. A neural network can conveniently be employed to perform the required data-fusion thus producing an overall “risk index”; however the surgical outcome being of a binary type, the network output tends to be a step-like function that does not give much information on the risk level. In this paper a modified training approach that takes the parameter uncertainties into account and which trains the network avoiding the step-like behaviour is proposed. The results that can be obtained with this approach are eventually explained by applying it to the estimation of a surgical risk index for lung resection procedures in patients affected by lung cancer.