Risk Profiling Used with Mammography Diagnosis Should Reduce Number of Missed Tumors

Mehmet U.S. Ayvaci, from the University of Texas Dallas (USA) presented his research group’s findings about the role of risk profiling in the interpretation of mammograms at Advances in Decision Analysis, a conference sponsored by the INFORMS Decision Analysis Society (DAS). Formed in 1980 with 1,000 current members, DAS supports the development and use of logical methods for improving decision-making in public and private enterprise.

The researchers found that providing radiologists with the patient’s risk profile information for breast cancer at the most advantageous time when examining the mammogram, together with statistical weighting based on profile risk, reduces false-negative results by 3.7%, thereby forewarning women whose cancer would have gone undiagnosed at an early stage, when treatment is most effective. It also reduces false-positive findings by 3.23%, thereby slashing superfluous healthcare costs and sparing patients’ needless worry.

Risk factors include family history, reproductive history, age, and ethnicity, and others forming the risk profile information. The researchers examined the complicated questions of whether providing risk profile data about women being screened for cancer biases radiologists and, if there is bias, whether this bias actually helps make readings more effective.

Historically, available clinical evidence has been inconclusive on the use of profile information when interpreting mammograms. One position is that profile information helps radiologists make better decisions and should be employed when reading mammograms. A contradictory position holds that profile information may bias the radiologists. However, whether bias always causes harm is unclear.

The investigators searched profile information and potential bias in mammography interpretation using a decision science technique called linear opinion pooling, which assigns weights to better aggregate probability estimates. They analyzed the decision performance of three groups: (1) a mammogram-only reading, with no risk profile information about the patient; (2) an unbiased reading, in which radiologists consult the risk profile after examining the mammogram; and (3) lastly, biased or “influenced” readings, in which radiologists consult a woman's risk profile as they examine the mammogram. Then they examined the conditions in which profile information could help improve biopsy decisions.

Numeric analysis using a clinical dataset from the Breast Cancer Surveillance Consortium revealed that use of profile data with an appropriate weight could reduce the false-positives and the number of missed tumors when compared to cases where profile information was not examined.

Related Links:
Institute for Operations Research and the Management Sciences
University of Texas Dallas