RESPONSE: Re: Changes in Breast Cancer Detection and Mammography Recall Rates After the Introduction of a Computer-Aided Detection System
David Gur, Jules H. Sumkin, Howard E. Rockette · JNCI Journal of the National Cancer Institute · 2004
We thank Feig et al. for their three comments. First, we clearly stated in our article (page 188) that because of the large 95% confidence intervals, “our findings are consistent with the range of improvement in detection rates estimated and reported by others” ( 1 ) . Namely, we did not rule out the possibility that a large improvement is indeed possible. Whether it is likely, however, is another matter that requires more data that are appropriately ascertained and, as we indicated in our correspondence ( 2 ), we welcome these additional data, whatever they may be. Second, depending on the assumptions one makes, it is frequently possible to find subsets of participants that produce the desired result. As Feig et al. demonstrate, it is true that there was an increase (albeit not a statistically significant increase) in cancer detection rates among the low-volume radiologists, who interpreted fewer mammograms than the high-volume radiologists. However, during the same time period, the low-volume radiologists showed a statistically significant 14% increase in recall rates, from 10.5% (1242 women recalled per 11 803 mammograms read) to 12% (2596 women recalled per 21 639 mammograms read) ( P <.001). The persistent similar changes in recall and detection rates is the very reason we argued that, for radiologists who have had a certain level of training, recall rates may be highly correlated with detection rates, regardless of whether computer-aided detection is used ( 3 ) . In other words, one could argue that in our group of low-volume radiologists, the effect of computer-aided detection may have been similar to that obtained if we had simply told the radiologists to be more conservative (or suspicious) in their reading of mammograms. Feig et al. assume that during the period in question, our low-volume radiologists were less trained or less experienced than the high-volume radiologists, which was not necessarily the case. Third, we discussed the issues surrounding adjustment for confounding factors in this type of observational study in our own correspondence in response to the editorial by Elmore and Carney ( 2 ) who, unlike Feig et al., argued that we should have adjusted for recall rates to account for the gradual decrease in the fraction of women undergoing a first screen versus those undergoing a repeat screen. Elmore and Carney used our data to highlight the fact that recall rates (not cancer detection rates) have actually increased with computer-aided detection. Because we expected that the changes in recall rates and cancer detection rates among these two groups (first and repeat screens) would be similar in magnitude ( 4 ), we chose not to correct for this confounder. We believe, in this observational study, this decision was appropriate ( 2 ) .