Abstract
Mammography is an important tool in the early detection of breast cancer. However, the perceptual task is difficult and a significant proportion of cancers are missed. Visual search experiments show that miss (false negative) errors are elevated when targets are rare (low prevalence) but it is unknown if low prevalence is a significant factor under real world, clinical conditions. Here we show that expert mammographers in a real, low-prevalence, clinical setting, miss a much higher percentage of cancers than are missed when the mammographers search for the same cancers under high prevalence conditions. We inserted 50 positive and 50 negative cases into the normal workflow of the breast cancer screening service of an urban hospital over the course of nine months. This rate was slow enough not to markedly raise disease prevalence in the radiologists’ daily practice. Six radiologists subsequently reviewed all 100 cases in a session where the prevalence of disease was 50%. In the clinical setting, participants missed 30% of the cancers. In the high prevalence setting, participants missed just 12% of the same cancers. Under most circumstances, this low prevalence effect is probably adaptive. It is usually wise to be conservative about reporting events with very low base rates (Was that a flying saucer? Probably not.). However, while this response to low prevalence appears to be strongly engrained in human visual search mechanisms, it may not be as adaptive in socially important, low prevalence tasks like medical screening. While the results of any one study must be interpreted cautiously, these data are consistent with the conclusion that this behavioral response to low prevalence could be a substantial contributor to miss errors in breast cancer screening.
Highlights
Mammographic screening is an important tool in the early detection of breast cancer [1] but it is a difficult perceptual task and error-prone [2] with reported false negative rates of 20–30% [3,4]
Within the constraints imposed by the real world, we show that false negative errors are higher in the low prevalence clinical setting than in the high prevalence, lab setting, suggesting that a substantial portion of missed cancers may be missed because of the properties of the human ‘search engine’
Because we could not control which radiologist saw which case in the low prevalence arm, data from the low prevalence arm was treated as if the entire 14radiologist practice constituted one experimental observer
Summary
Mammographic screening is an important tool in the early detection of breast cancer [1] but it is a difficult perceptual task and error-prone [2] with reported false negative rates of 20–30% [3,4]. The signs of breast cancer are often ambiguous and/or hard to see, with some proportion of errors attributable to the perceptual difficulty of the task. If disease is detected in the current exam, it can be seen in retrospect on the previous exam. These ‘‘retrospectively visible’’ or ‘‘actionable’’ cancers could have been found but were missed on that previous exam [5,6,7]. They are either errors in perception (failures of search) [8], or alternatively errors in interpretation. We consider one contributor to those failures, namely the low prevalence of disease in screening mammograms
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have