Multimodal AI Improves Breast Cancer Recurrence Risk Prediction Beyond Standard Genomic Testing

Widely used genomic assays are most informative during the first five years after diagnosis, yet many recurrences occur later, particularly in hormone receptor-positive, human epidermal growth factor receptor 2-negative (HR+/HER2-) disease. To improve risk assessment across both early and late periods, investigators have developed an artificial intelligence model designed to provide more accurate estimates of distant recurrence risk.

The innovation, called IICM+, was created by the ECOG-ACRIN Cancer Research Group (Philadelphia, PA, USA) in collaboration with Caris Life Sciences (Irving, TX, USA). It is designed for patients with HR+/HER2-, a subtype that accounts for about half of cases in the United States. The work is detailed in a study published in npj Breast Cancer.

IICM+ integrates multimodal inputs to generate a single prognostic score. The model combines digitized tumor histopathology with clinical variables such as age, tumor size, and grade. It also incorporates molecular features from an expanded 42-gene panel. The approach uses modern AI techniques to fuse these data modalities into a unified prediction of recurrence risk.

Development and validation drew on tumor specimens and more than a decade of outcome data from 4,429 participants in the TAILORx breast cancer trial. Investigators developed the model using data from 2,808 participants and independently validated it in a separate cohort of 1,621 participants, with median follow-up exceeding 11 years in both groups. Performance was evaluated for overall distant recurrence as well as early recurrence, occurring within five years, and late recurrence, occurring after five years.

In the independent validation cohort, IICM+ outperformed the 21-gene Recurrence Score based on the C-index, with higher values indicating better discrimination. For overall distant recurrence, the C-index was 0.735 for IICM+ versus 0.578 for the Recurrence Score; for early distant recurrence, 0.791 versus 0.722; and for late distant recurrence, 0.710 versus 0.514, with statistically significant differences reported. 

The model also identified higher-risk subsets among patients classified as having low genomic risk (Recurrence Score 0–25) and lower-risk subsets among those classified as high genomic risk (26–100), suggesting it could further refine interpretation of the Oncotype DX Breast Recurrence Score. The investigators reported that the findings do not establish IICM+ for selecting or changing treatment at this time. Further studies are in development to validate the model in additional populations and to determine whether use of the tool can guide treatment decisions and improve outcomes.

"Powered by artificial intelligence integrating clinical, molecular and histopathology data, this new test provides more reliable prognostic information for breast cancer recurrence," said Joseph A. Sparano, M.D., of the Icahn School of Medicine at Mount Sinai.

"TAILORx continues to yield important new insights years after its original findings—a testament to the enduring value of Cooperative Group research. Applying today's technologies to this exceptional research resource creates opportunities to further individualize breast cancer care," said Peter J. O'Dwyer, M.D., ECOG-ACRIN group co-chair.

Related Links
ECOG-ACRIN Cancer Research Group 
Caris Life Sciences