Enhancing and Evaluating Diagnostic Accuracy
We develop two themes. The first is that diagnostic accuracy can be enhanced by decision aids. Our experimental finding that the enhancement was greater for cases more difficult to diagnose led to the second theme: an evaluation of the comparative accuracies of two diagnostic approaches, such as two alternative imaging techniques, may lead to various and contradictory conclusions depending on the spectrum of difficulty in the set of test cases.
A pair of decision aids was suggested by a conception of the diagnostic process in which pieces of evidence are individually assessed and then merged into an overall diagnosis. In our implementation, the pieces of evidence were the perceptual features of film mammograms and the diagnosis was the estimated probability of breast cancer. The first aid is a checklist of features that prompts the image reader to give a scale value for each feature, either a subjective rating of the degree to which the feature is present or a physical measurement of the extent of the feature. The second aid is a computer program that accepts the scale values from the reader, merges them with relative weights appropriate to their degrees of diagnostic relevance, and issues an advisory estimate of the probability of malignancy. In our tests, a set of approximately 150 proven cases was read twice by six radiologists, first in their usual manner and later with the decision aids.
In regard to the enhancing effect of the decision aids, this experiment extends one we reported earlier.1 The main differences are that this experiment investigated film mammography, now more common than the xeromammography used in the earlier experiment, and it required detection of lesions as well as a benignmalignant classification. Both experiments build on an extensive literature in computer-based decision making. Dawes and Corrigan advanced the idea that