Study: AI Tool Uses Multiple Years of Mammograms to Predict Risk

Breast cancer doesn't happen overnight. Abnormal cells in the breast multiply and grow over time. What if there was a way to detect breast changes before they show up on a mammogram?

Artificial intelligence (AI) may be able to do just that. A recent study found that when an AI model reviewed a person’s last three years of mammograms, the model did a better job at finding people at greatest risk of breast cancer than when people answered questions about their health histories.

Knowing your risk of breast cancer can help you and your doctors create a plan for any additional screenings, treatments, and lifestyle changes you may need, says Shu (Joy) Jiang, PhD, who was the lead author of the study. Jiang is a professor of surgery at Washington University.

Most tools that healthcare professionals use to predict breast cancer risk today are based on a single snapshot in time — for instance, how you answer questions in a breast cancer risk calculator or what your last mammogram found.

Many experts recommend that women at average risk of breast cancer get mammograms each year beginning at age 40. Jiang wanted to create a tool that could use information collected from years of mammograms to better predict breast cancer risk.

Her team built an AI model that uses information from up to three years of previous mammograms to predict a person’s risk of developing breast cancer within five years. 

The AI model studied mammograms from over 10,000 women who received breast cancer screenings from 2008 to 2012. The group of women was followed to 2020. During the follow-up period, 478 were diagnosed with breast cancer. Using this data, the model looked for patterns, such as changes in breast density, to predict breast cancer risk.

Next, the team used their model to review mammograms from more than 18,000 different women and separate the women into groups based on their risk of developing breast cancer within five years. In this group, 332 were diagnosed with breast cancer.  

The AI model assigned 61% of women who developed breast cancer in the next five years to the highest risk category. Meanwhile, the Tyrer-Kuzick questionnaire (a breast cancer risk calculator) assigned only 26% of women to the high-risk category. By using multiple mammograms, the AI model was 2.3 times more accurate than the questionnaire.

Jiang says the researchers are working on testing the AI model on diverse groups of women. To date, the model has been tested on Black, Asian, and white women, but the researchers hope to test it on more Hispanic women and women of Pacific Islander or Native American descent. 

Currently, the team can only look back to the last three years of mammograms to predict the next five years of risk. A question that remains is, “What is the optimal number of years of mammograms that we should look at [to predict risk]?” says Jiang. Answering this question will require more follow-up studies.