Prediction of outcome in multiorgan resections for cancer using a bayes-network

The long-term success of multivisceral resections for cancer is difficult to forecast due to the complexity of factors influencing the prognosis. The aim of our study was to assess the predictivity of a Bayes network for the postoperative outcome and survival.

We included each oncologic patient undergoing resection of 4 or more organs from 2002 till 2005 at the Ulm university hospital. Preoperative data were assessed as well as the tumour classification, the resected organs, intra- and postoperative complications and overall survival. Using the Genie 2.0 software we developed a Bayes network.

Multivisceral tumour resections were performed in 22 patients. The receiver operating curve areas of the variables "survival >12 months" and "hospitalisation >28 days" as predicted by the Bayes network were 0.81 and 0.77 and differed significantly from 0.5 (p: 0.019 and 0.028, respectively). The positive predictive values of the Bayes network for these variables were 1 and 0.8 and the negative ones 0.71 and 0.88, respectively.

Bayes networks are useful for the prognosis estimation of individual patients and can help to decide whether to perform a multivisceral resection for cancer.