RESUMEN
A test of the performance of two probabilistic classifiers (random forests and multinomial logit models) in automatically defining cancer cases has been carried out on 5608 subjects, registered by the Venetian Tumour Registry (RTV) during the years 1987-1996 and manually checked for possible second cancers that occurred during the 1997-1999 period. An eightfold cross-validation was performed to estimate the classification error; 63 predictive variables were entered into the model fitting. The random forest allows to automatically classify 45% of subjects with a classification error lower than 5%, while the corresponding error is 31% for the multilogit model. The performance of the former classifier is appealing, indicating a potential drop of manually checked cases from 1750 to 960 per incidence year with a moderate error rate. This result suggests to refine the approach and extend it to other categories of manually treated cases.