How accurate are breast cancer risk prediction models in women with a family history of breast cancer?

Key messages

• We found four breast cancer risk prediction models that had been tested enough times to evaluate in detail. These were the Gail, Tyrer-Cuzick, BOADICEA, and BRCAPRO models.

• The BOADICEA model was one of the more reliable tools for estimating future breast cancer risk in women with a family history of the disease, meaning it may help them and their doctors to decide on treatment.

• Further research is needed to improve the accuracy of existing prediction models in distinguishing between women with a family history of breast cancer who will and will not develop the disease.

Why is it helpful to be able to predict a woman’s risk of breast cancer?

Women who have a history of breast cancer in their family have a higher likelihood of developing breast cancer themselves during their lifetime. In clinics, a woman's chance of developing breast cancer in a given time period is often estimated using statistical tools known as breast cancer risk prediction models.

Being able to estimate the risk of developing breast cancer accurately for a woman with a family history of breast cancer helps doctors and the woman decide how to manage or reduce her risk of breast cancer. Management may include:

• regular imaging with mammograms or magnetic resonance imaging (MRI; a type of scan that creates detailed pictures of the breast tissue) to detect breast cancer at an early stage;

• taking risk-reducing medications; or

• in some cases, surgical removal of both breasts to prevent breast cancer.

Currently, it is not clear which of the available breast cancer risk prediction models works best in women with a family history of breast cancer.

What did we want to find out?

We wanted to:

• identify breast cancer risk prediction models that have been developed or tested (or both) in women with a family history of breast cancer; and

• assess how accurately they predict future risk of developing breast cancer in these women.

What did we do?

We searched for studies that developed or tested these models. We looked at how accurately the models predicted breast cancer risk, focusing on:

• whether the predicted number of breast cancer cases was similar to the number that actually occurred (calibration); and

• whether the model could distinguish between women who did and did not develop breast cancer (discrimination).

When enough studies assessed the same model, we combined their results statistically.

What did we find?

We identified 12 models that estimate future breast cancer risk that had been tested in studies where all or most women had a family history of breast cancer. The models were tested using information from as few as 134 women to as many as 130,058. Most of the women lived in North America, Europe, or Australia, with a small number from Asia.

The studies were funded by governments (25 studies), universities (24), non-profit organisations (21), and industry (3). Six studies did not report funding sources, and some received funding from more than one source.

We were able to combine results from several studies for four models: the Gail, Tyrer-Cuzick, BOADICEA, and BRCAPRO models.

Calibration: was the predicted number of breast cancer cases similar to the actual number?

• The Gail and BOADICEA models accurately estimated the numbers of women in the included studies who would develop breast cancer in a given timeframe:

  • for every 100 breast cancers Gail predicted would occur, about 106 actually occurred in reality;

  • for every 100 breast cancers BOADICEA predicted, about 98 actually occurred.

• The Tyrer-Cuzick model estimated that more women in the studies would develop breast cancer than actually did. For every 100 breast cancers it predicted, only about 86 actually occurred.

• The BRCAPRO model estimated that fewer women in the studies would develop breast cancer than actually did. For every 100 breast cancers it predicted, about 144 actually occurred.

Discrimination: how well did the models distinguish between women who develop and do not develop breast cancer?

• All four models were moderately accurate in distinguishing between women who would and would not develop breast cancer in a given timeframe, but none did a great job.

• The Tyrer-Cuzick (version 8), BOADICEA, and BRCAPRO models correctly distinguished women who would develop breast cancer from those who would not about 64 to 65 times out of 100.

• The Gail model performed slightly less well, correctly distinguishing women who would develop breast cancer 61 times out of 100.

What are the limitations of the evidence?

We rated the quality of most studies included in our review as poor or unclear, which means we cannot be confident that these results are reliable. Our confidence was reduced for several reasons, including:

• in some studies, too few women developed breast cancer, making it harder to judge how accurate the prediction models were;

• not all studies reported the model performance information we sought;

• some studies had missing information or did not explain how they handled missing information.

How up to date is this evidence?
The review includes studies published up to December 2024.