Summary:In computer-aided prognosis theoretical arguments in favour of a sophisticated model often do not result in a markedly better performance. In this study, straightforward extensions of the simple and comprehensive independent Bayes model, taking interactions between variables into consideration, are investigated, using a clinical data set of upper gastrointestinal bleeding, four sets of variables and different measures of performance (discriminatory ability, sharpness, reliability). For all criteria of performance, the differences between the models were small in the sets with few variables. In the sets with many variables there were marked differences between the models; however, no model was superior in all aspects of performance. Incorporation of interactions in models based upon Bayes’ theorem are worthwhile if many variables are used and the discriminatory ability is considered.

SUMMARYIn this paper two problems of computer-aided diagnosis with 'independence Bayes' were investigated: selection of variables and monotonicity in performance as the number of measurements is increased. Using prospective data from patients with upper gastrointestinal bleeding, the stepwise forward selection approach maximizing the apparent diagnostic accuracy was analysed with respect to different kinds of bias in estimation of the true diagnostic accuracy and to the stability of the number and type of variables selected. The results of this study suggest first that the selection of variables should be evaluated against the estimated true diagnostic accuracy obtained using all variables, and secondly that the results of a single selected sequence may be severely biased.

No abstract

No abstract

No abstract