Use of an artificial neural network to predict cancer development in patients with inflammatory myopathy: Comment on the letter by Selva O'Callaghan et al [9] (multiple letters)

To the Editor:
We read with great interest the comments of Selva O'Callaghan et al (1) pertaining to our report on use of the artificial neural network to improve differentiation between Churg‐Strauss syndrome and Wegener's granulomatosis (2). The authors describe the application of a neural network to predict the development of cancer in patients with idiopathic inflammatory myopathy (IIM).

From the methodologic point of view, the approach is very convincing, particularly the feature selection and the employment of cross‐validation. We assume that the authors chose the smallest admissible network size that provides a solution, so that regularization methods or early stopping were not needed in order to avoid overfitting. The accuracy obtained is very impressive. As a suggestion for further work, one might calculate a receiver operating characteristic curve for explicitly adjusting the sensitivity and specificity (3).

Nevertheless, we have some concerns with respect to the results. In Selva O'Callaghan et al's study, 13 patients developed cancer. The authors obtained a sensitivity of 98.46%. They probably correctly classified 12 of the 13 patients with cancer; however, 12 of 13 is 92.3%. How do the authors explain the difference? Perhaps they should recalculate this point. Moreover, the percentage of men with cancer in relation to the sex is inconsistent. In their Table 1, if there are 6 men with cancer (of 19), then the percentage should be 32%. This percentage still seems to be very high, higher than that described among male patients in other studies (4, 5). In this connection it would be of interest to know when these patients developed their cancer. The highest risk for cancer in patients with IIM has been described to occur around the time of diagnosis, i.e., within the first year since the diagnosis was established. The risk for malignant disease diminished with time and disappeared beyond 5 years (4, 5). This point seems to be of great practical importance in determining the extent of the search for malignancy at the time of diagnosis of an IIM. The occurrence of cancer beyond the first year of diagnosis of polymyositis/dermatomyositis could also be associated with the type of therapy.