Dissertation Abstract

Effect of Computer-Generated Diagnostic Feedback on the Performance of Radiologists Interpreting Mammographic Film Images.

Elaborate interactive computerized learning systems are currently being developed in medicine. However, numerous basic questions about interactive learning in the individual medical specialties are just beginning to be addressed. In this study, the specific question addressed was: does feedback provided to a radiologist within an interactive computerized system influence the results of a diagnosis that has already been made (as measured by change in the diagnosis)? The radiologic procedure that was used in this study was the mammogram. The setting for the study was the mammography clinic of a major urban teaching hospital.

A database of cases was established (part 1 of the study), for which six radiologists trained in mammography interpreted a total of 429 cases (331 negative, 118 positive) using an interactive computer program designed specifically for this study. An algorithm that acted as a pattern classifier 'graded' responses to these inputs, along with those made in 255 cases (175 negative and 80 positive) that were interpreted during the segment of the study (part 2) in which a feedback loop was implemented. From this pattern classifier, feedback based on the database cases was provided to the six radiologists who participated in part 2 of the study. Three of the six radiologists made changes based on information received from the feedback mechanism during the mammographic interpretations they performed; three did not. Half the changes made as a result of feedback moved diagnoses closer to the 'truth,' as established from the known outcomes of cases in part 1 of the study. The other half moved diagnoses further from the truth. In no case was a change made in this study that would have overlooked a cancer and not sent the woman to biopsy when she required one. Nine per cent of the total cases in part 2 of the study were affected by the changes the radiologists made.

Overall, the feedback had little effect on the degree of accuracy (as measured by receiver operating characteristic analysis) of these radiologists, whose accuracy rates both before and after feedback were high. This was a pilot test of a system that, in theory, would help radiologists interpret mammograms. Because of logistical limitations, this theory could not be proved in this study. However, results hinted that the more practice one has with the system, the more likely it is to be of value in helping with the interpretation of a mammogram. It is also possible that practice with the system by less experienced or sophisticated readers would show a different result.

This dissertation citation and abstract are published with permission of ProQuest Information and Learning. Further reproduction is prohibited without permission.