The surgical treatment of an ovarian tumour is a problem. Since to the naked eye it is often impossible to determine whether the tumour is benign or malignant, inappropriate radical surgery may be performed where the tumour turns out to be benign, or inappropriate conservative surgery where the tumour turns out to be malignant. This problem has resulted in the search for methods to determine pre-operatively the nature of an ovarian tumour so that the correct surgical treatment may be given. These methods rely on mathematical models derived retrospectively from multiple logistic regression analysis of various risk factors associated with ovarian cancer, but it is not clear how well these models perform prospectively. Naaila Aslam and her colleagues (pages 1347–1353) compared three mathematical models for the prediction of ovarian cancer: the risk of malignancy index, the revised risk of malignancy index, and Tailor's regression model. These models were tested prospectively in 63 women with adnexal masses, in 23 of whom the masses were malignant. Not surprisingly none of the models performed as well in this prospective study as they did in previous retrospective studies. There was a systematic difference between the models, for the sensitivity in Tailor's regression model was significantly less than in the two risk of malignancy models (χ2= 6.15; DF = 2; P < 0.05). Furthermore the performance of the unmodified risk of malignancy index approaches that of a “good test”, the likelihood ratio of a positive test being 9.25 and a negative test 0.28. This means that if the background risk of malignancy in a population of women with an adnexal mass is 10%, this risk will increase to 51% if the risk of malignancy index is positive, and decrease to 3% if the risk of malignancy index is negative. These results imply that if the risk of malignancy index is positive the woman should be referred to a cancer centre for radical surgery, but if it is negative conservative surgery is justified.

However this study suggests that further refinements of the predictive models are required. Fifteen of the 23 ovarian cancers in the study were advanced, such that the diagnosis of probable malignancy did not require the sophistication of a mathematical model. Since the real problem is the treatment of an isolated adnexal mass, future research into predictive models for the diagnosis of ovarian cancer should be confined to women with such an isolated adnexal mass. Furthermore in this study there was a mixture of histological types of ovarian cancer — invasive epithelial, borderline epithelial, and non-epithelial. It was not surprising that Tailor's regression model performed less well than the others, since it was derived from women with invasive epithelial cancers only. Since the morphological characteristics of ovarian tumours will vary with these different histological types, it may be necessary to develop mathematical models for each histological type. The science of predicting ovarian cancer has only just begun.