While news about clever chatbots, image generators, and supposedly deadly bots dominate discourse around artificial intelligence, it can also be incredibly helpful—especially when used as a tool with plenty of human oversight. In fact, it can even be used to help predict the most common form of cancers in the world.
A multi-university team of U.S. researchers published a study Tuesday in the journal Radiology that found that AI was able to outperform the standard clinical risk model for predicting the five-year risk of breast cancer in a study of thousands of mammograms. The authors believe that the findings could lead to faster and more accurate cancer diagnoses in the future.
That’s great news of course, with one shocking wrinkle: The researchers are not entirely sure how the AI was able to do it.
“We’re looking for an accurate, efficient and scalable means of understanding a woman's breast cancer risk,” lead author Vignesh Arasu, a radiologist at Kaiser Permanente Northern California, said in a statement. “Mammography-based AI risk models provide practical advantages over traditional clinical risk models because they use a single data source: the mammogram itself.”
Breast cancer risk is calculated using clinical models. These typically utilize self-reported demographic and health data about the patient such as their age, family health history, and breast density to assess the probability that they’ll develop breast cancer via a risk score.
However, these models have flaws that sometimes cause inaccurate assessments. The patient might not know all of the information. They might get something wrong. Whatever the case, it could mislead both doctor and patient about the likelihood they’ll get the disease.
“Clinical risk models depend on gathering information from different sources, which isn’t always available or collected,” Arasu explained. “Recent advances in AI deep learning provide us with the ability to extract hundreds to thousands of additional mammographic features.”
The study’s authors drew on a cohort of 12,628 women in Northern California who had mammograms performed on them in 2016. Using this group of imaging, the researchers then gave them risk scores for breast cancer over a five-year period using five AI models—two academic algorithms, and three commercially-available algorithms—and the Breast Cancer Surveillance Consortium (BCSC) risk model, which is commonly used to assess breast cancer risk.
The results skewed heavily in favor of the bots, which were even able to pick up on missed breast cancer indicators despite the fact that the authors don’t know how it did it. In fact, AI models that were trained in short time spans as quick as three months were able to predict cancer risk up to five years when none was detected during breast cancer screenings.
“All five AI algorithms performed better than the BCSC risk model for predicting breast cancer risk at 0 to 5 years,” Arasu said. “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.”
As with all things AI, though, this comes with a few big grains of salt. For one, the data was pulled from women who lived in Northern California, and the majority of them were non-Hispanic white. AI of all types often come with deep-seated biases that result in sexist, racist, and generally problematic behavior.
Also, there’s reason to be wary of technology that we don’t fully understand, especially if the people who made them don’t understand it either. The “black box” nature of these AI models often creates harmful results for its users. This is especially dangerous when it comes to AI being used in medicine—a field that has a sordid history with scientific racism and sexism.
Still, the study’s authors believe that the technology represents a hopeful—if not a bit mysterious—future for breast cancer screenings. “AI for cancer risk prediction offers us the opportunity to individualize every woman’s care, which isn’t systematically available,” Arasu said. “It’s a tool that could help us provide personalized, precision medicine on a national level.”