A 10-year mammography-based AI risk model achieved high performance in a study published May 20 in Science Translational Medicine.
The model showed significantly higher performance compared to established clinical risk models and Mirai in a diverse cohort, wrote a team led by Mikael Eriksson, PhD, from Karolinska Institute in Solna, Sweden and colleagues.
“For breast radiologists, this could become a practical tool for risk-stratified screening and prevention,” Eriksson told AuntMinnie.
Radiologists continue to explore AI’s capabilities in predicting breast cancer risk, with some recent studies showing how the technology can be a reading assistant and help predict five-year risk. Eriksson and colleagues highlighted a “critical need” for image-based risk models for long-term disease prediction over current clinical models where cancers can slip through the cracks.
The Eriksson team developed its own AI model for long-term risk prediction of invasive and in situ breast cancer. It used an ensemble machine learning method to extract image-based risk features from digital mammograms to predict breast cancer in a case cohort sampled from the KARolinska MAmmography Project for Risk Prediction of Breast Cancer (KARMA) screening.
The researchers performed an independent evaluation of 10-year discriminatory, calibration, and clinical risk classification performances in two population-based case cohorts. These consisted of women from KARMA as an internal validation set and from Olmsted Medical Center in Minnesota as an external validation set. The team also performed further validation in the hospital-based EMory BrEast imaging Dataset (EMBED), based in Atlanta.
The study included breast cancer data from 8,696 women in the Olmsted and KARMA cohorts, and 1,633 women with incident cancers. The model achieved similar performance in the Olmsted and KARMA cohorts.
AI model performance on cohorts | ||
Measure | Olmsted cohort | KARMA cohort |
Average 10-year risk | 3.83% | 3.14% |
Expected-to-observed events ratios | 0.99 | 0.99 |
AUC (invasive cancers) | 0.72 | 0.72 |
The researchers noted similar results in the EMBED cohort, including an AUC of 0.70 for predicting invasive cancers.
The model performed significantly better than the Mirai AI model (Jameel Clinic, Massachusetts Institute of Technology) in all three cohorts. Mirai achieved AUCs of 0.64 in Olmsted, 0.64 in KARMA, and 0.63 in EMBED.
In the top 10% of high-risk women in the KARMA cohort, the AI risk tool predicted 33% of breast cancers compared with the following models: Tyrer-Cuzick, version 8 (23%), Breast Cancer Surveillance Consortium, version 3 (20%), and Mirai (24%).
Potential clinical use
The AI tool could help identify women who may benefit from more personalized care, Eriksson said. He added that the goal of this tool is to use information already present in mammograms to support targeted prevention strategies.
“Our work suggests that mammograms are not only screening tools but could also be a source of long-term biological risk information,” Eriksson told AuntMinnie. “That opens the possibility of using AI to identify women at high risk years before cancer develops with the aim to potentially prevent the disease.”
The team will next test the model in broader and more diverse populations and healthcare systems to better understand how it performs in real-world clinical practice.
“We are also interested in studying how this type of AI-driven risk assessment can be integrated into prevention programs and risk-stratified screening pathways,” Eriksson said. “Another important area is to further develop risk reduction strategies, which high-risk women may benefit from in order to reduce their risk of developing breast cancer in the first place.”
Read the full study here.
