AI model improves recurrence prediction in early breast cancer trial

An artificial intelligence model called IICM+ predicted distant recurrence more accurately than the 21-gene Recurrence Score in patients with hormone-receptor-positive, HER2-negative, lymph node-negative early-stage breast cancer, according to a model-development and validation study using data from the ECOG-ACRIN TAILORx clinical trial. The researchers analyzed tumor specimens, clinical information and long-term outcomes from 4,429 participants, with median follow-up exceeding 11 years in both the development and independent validation groups. IICM+ had better discrimination than the genomic score for overall distant recurrence and for recurrence classified as early or late. The findings support improved prognostic performance in this trial population, but they do not establish that the model can safely guide treatment selection or improve clinical outcomes.

The study focused on HR-positive, HER2-negative disease, which the researchers described as the most common breast cancer subtype and as accounting for about half of all breast cancer cases in the United States. IICM+ was designed partly to address a limitation of the 21-gene Recurrence Score, also known as the Oncotype DX Breast Recurrence Score. The established genomic test is more effective at predicting recurrence within five years of diagnosis than at predicting recurrence later. More than half of distant recurrences in this breast cancer population occur after the five-year mark, with some reported 10, 15 or even more than 20 years after initial surgery. The investigators therefore tested whether combining pathology images, broader molecular information and clinical factors could improve prediction across both early and later recurrence periods.

IICM+ integrates several data types into a single prognostic model. Its inputs include digitized histopathology images of tumors and features extracted from those images with deep-learning methods. The model also uses clinical variables such as patient age, tumor size and tumor grade, together with molecular information measured through an expanded 42-gene panel. This multimodal approach enables the algorithm to assess tumor morphology, molecular characteristics and patient-level information together rather than relying only on the 21 genes represented in the standard recurrence assay. For the analysis, researchers converted archived TAILORx pathology slides into digital images, creating an opportunity to apply contemporary artificial intelligence to tumor specimens and clinical data collected years earlier.

The research team used information and tumor specimens from 2,808 TAILORx participants to develop IICM+, then evaluated the completed model in an independent group of 1,621 participants whose data had not been used during development. Median follow-up was more than 11 years in both groups. Investigators assessed performance for overall distant recurrence, defined as cancer returning in parts of the body away from the original tumor, early distant recurrence within five years and late distant recurrence after five years. They compared IICM+ with the 21-gene Recurrence Score using the C-index, a discrimination measure for which a higher value indicates a stronger ability to distinguish patients with different recurrence risks.

In the independent validation group, IICM+ achieved a C-index of 0.735 for overall distant recurrence, compared with 0.578 for the 21-gene Recurrence Score. The difference was statistically significant, with P<0.001. Performance was also better for early distant recurrence, although the margin was narrower: IICM+ produced a C-index of 0.791, versus 0.722 for the genomic score, with P=0.046. The largest separation appeared for late distant recurrence. IICM+ achieved a C-index of 0.710, while the Recurrence Score achieved 0.514, a statistically significant difference with P<0.001. These results indicate that the integrated model discriminated between different recurrence risks more accurately than the genomic score in the validation cohort, particularly for outcomes occurring after five years.

The model also separated groups within the broad categories supplied by the Recurrence Score. Among patients with a Recurrence Score of 0–25, representing lower genomic risk, IICM+ identified a subset with a higher observed risk of distant recurrence. In the other direction, among patients with a Recurrence Score of 26–100, representing higher genomic risk, IICM+ identified a subset whose observed recurrence risk was lower than the genomic score alone would suggest. The analysis thus shows that the additional information could refine recurrence-risk estimates beyond the broad genomic categories. The evidence demonstrates differences in prognostic estimation, not that any individual prediction is sufficient to determine treatment.

The project drew on the unusually detailed resources accumulated through TAILORx, a National Cancer Institute-supported trial that enrolled 10,273 women with HR-positive, HER2-negative, lymph node-negative early breast cancer. TAILORx helped define which patients could be treated effectively without chemotherapy. Its biorepository combined preserved tumor specimens with detailed molecular and clinical information, treatment records and long-term follow-up outcomes. In a public-private partnership with Caris Life Sciences, archived pathology slides were transformed into digital images and analyzed with modern artificial intelligence and deep-learning methods. The study demonstrates how a long-running clinical trial can support new prognostic research using specimens and outcomes assembled for earlier investigations.

Joseph A. Sparano, MD, of the Icahn School of Medicine at Mount Sinai, said IICM+ uses artificial intelligence to integrate clinical, molecular and histopathology data in an effort to provide more reliable prognostic information about breast cancer recurrence. George W. Sledge Jr., MD, executive vice president and chief medical officer of Caris, said combining clinical data, next-generation sequencing data and imaging data through modern artificial intelligence techniques created a new way to interrogate breast cancers for prognostic information. Peter J. O’Dwyer, MD, ECOG-ACRIN Group Co-Chair, said applying current technologies to the TAILORx research resource creates opportunities to further individualize breast cancer care. These statements describe the model’s research purpose and potential; they do not constitute evidence of treatment benefit.

Several limitations remain despite the independent validation. The model was developed and evaluated in participants from TAILORx, all of whom had HR-positive, HER2-negative, lymph node-negative early-stage disease, so its performance in other patient populations is not established by this study. The research also does not determine whether IICM+ can guide chemotherapy or other treatment decisions, whether its predictions improve outcomes, or whether the model is ready for routine clinical use. Additional studies are needed and reportedly under development to validate the approach in other populations and assess whether it can inform treatment decisions and improve patient outcomes. The peer-reviewed findings were published in npj Breast Cancer and should currently be interpreted as evidence of improved recurrence-risk prediction in a defined trial population, not as proof of a new clinical pathway.

Subject of Research: AI-assisted prediction of early and late distant recurrence in early-stage breast cancer

Article Title: A new AI-powered test offers more accurate prediction of breast cancer recurrence risk

Article References: A new AI-powered test offers more accurate prediction of breast cancer recurrence risk. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, breast cancer, recurrence prediction, TAILORx trial, digital pathology, molecular profiling, prognostic modeling, 21-gene assay, long-term outcomes, machine learning, AI-powered, test