Prevalence effects in visual research: Theoretical and practical implications

点击翻译按钮获取中文摘要

Low prevalence searches form an important and problematic class of visual search tasks. These are tasks where the search target is rare. Many socially important tasks like airport security or cancer screening are low prevalence tasks. Previous work, much of it from our lab, has shown that low prevalence can have undesirable effects. Most notably, miss (false negative) errors are markedly elevated at low prevalence. This is a clear problem if the purpose of the search is to detect something rare but important like cancer or a terrorist threat. Our previous work has documented this pattern of increased miss errors in a number of expert domains including cytology (cervical cancer screening), airport baggage screening, and breast cancer screening. False alarm (false positive) error rates typically decline at low prevalence, moving in the opposite direction from miss errors. This indicates a shift in the observer’s decision criterion. At low prevalence, observers become more reluctant to call something a target. Several studies – ours and others - have shown that this “conservative” criterion shift is not adequate to explain the entire prevalence effect. Wolfe and VanWert (2010) developed a “Dual- Threshold” model that better captures the important aspects of the prevalence effect data by proposing two effects of low prevalence: (1) the conservative shift in the criterion for deciding if an attended item is a target, and (2) a lowering of the “quitting threshold.” The quitting threshold determines when observers end a search. Quitting too soon also increases the chance that the observer will miss a target. Prevalence effects have been studied in experimental isolation from other aspects of search. However, in tasks like breast cancer screening, other factors interact with prevalence. The four projects in the present proposal each investigate one of these interactions. Project 1 examines the relationship of prevalence to the “vigilance decrements” that are seen as time elapses in a task. In search, observers must maintain an internal, mental representation of the search target (or targets). Project 2 is concerned with the impact of prevalence on these “target templates”. Advances in artificial intelligence (notably deep learning) are producing tools to assist expert searchers. However, once deployed, these AI tools have been less effective than theory predicts. Project 3 tests the hypothesis that part of the problem is another side-effect of low prevalence and the project tests a potential intervention. Finally, clinicians, searching for one type of target (e.g. pneumonia) are supposed to report signs of other possible problems (e.g. lung cancer). Project 4 probes the role of prevalence in the failure to report such “incidental findings”. Again, we test several interventions. This is “use-inspired, basic research” whose results will provide guidance for experts performing socially important low prevalence tasks.