AI Outperforms Standard Risk Model for Predicting Breast Cancer

The risk of breast cancer in women is generally assessed using clinical models like the Breast Cancer Surveillance Consortium (BCSC) risk model. This model utilizes various patient data, including age, family history of breast cancer, childbirth history, and breast density, to produce a risk score. Now, a large study of thousands of mammograms has demonstrated that artificial intelligence (AI) algorithms can outperform this standard clinical risk model in predicting five-year breast cancer risk.

In the retrospective study, researchers at Kaiser Permanente Northern California (Oakland, CA, USA) used data from negative (indicating no visible signs of cancer) screening 2D mammograms carried out in 2016. From the 324,009 eligible women screened that year, a randomly selected subgroup of 13,628 women was examined. Additionally, all 4,584 patients who were diagnosed with cancer within five years of their 2016 mammogram were also included in the study. All women were monitored until 2021. The researchers split the five-year study duration into three separate time frames: interval cancer risk (diagnoses between 0 and 1 years), future cancer risk (diagnoses between 1 and 5 years), and all cancer risk (diagnoses between 0 and 5 years).

Image: AI algorithms outperformed standard clinical risk model for predicting five-year risk for breast cancer (Photo courtesy of Freepik)

Five AI algorithms, including two used by researchers and three commercially available ones, were employed to generate breast cancer risk scores over the five-year period using the 2016 screening mammograms. These risk scores were then compared to each other and the BCSC clinical risk score. The study revealed that all five AI algorithms outperformed the BCSC risk model in predicting 0 to 5 year breast cancer risk. Some AI algorithms excelled in identifying high-risk patients for interval cancer, which can often be aggressive and may necessitate a second mammogram reading, additional screening, or follow-up imaging at short intervals. For instance, when assessing women with the top 10% risk, AI predicted up to 28% of cancers compared to the 21% predicted by the BCSC. Interestingly, even AI algorithms designed for shorter time horizons (as low as 3 months) could predict up to five years of future cancer risk when no cancer was clinically detected by the screening mammogram. When combined, the AI and BCSC risk models further enhanced cancer prediction.

"Clinical risk models depend on gathering information from different sources, which isn't always available or collected," said lead researcher Vignesh A. Arasu, M.D., Ph.D., a research scientist and practicing radiologist at Kaiser Permanente Northern California. "Recent advances in AI deep learning provide us with the ability to extract hundreds to thousands of additional mammographic features."

"This strong predictive performance over the five-year period suggests AI is identifying both missed cancers and breast tissue features that help predict future cancer development. Something in mammograms allows us to track breast cancer risk. This is the 'black box' of AI," added Arasu.

Related Links:
Kaiser Permanente