Calculating risk and outcome: The society of thoracic surgeons database

Various approaches to the calculation of medical risk are reviewed, including univariate analysis, additive methods, use of Bayes' theorem by The Society of Thoracic Surgeons, logistic regression, and neural networks. Strengths and weaknesses of the various approaches are evaluated. The use and importance of observed/expected ratios, the C statistic, and receiver operating curves are discussed. Specific requirements for the building of useful risk-calculation models are discussed, including the importance of the model set/test set method and the role of both numbers of patients and time frames in model building.