A multimodal artificial intelligence (AI) model incorporating multiscale histopathology features was able to predict recurrence risk in patients with early-stage hormone receptor–positive and HER2-negative breast cancer more accurately than the 21-gene recurrence score, Oncotype DX, according to findings published in npj Breast Cancer. The model, called IICM+, was validated on patients from the TAILORx trial with over a decade of follow-up.
“Powered by AI integrating clinical, molecular, and histopathology data, this new test provides more reliable prognostic information for breast cancer recurrence,” said lead study author Joseph A. Sparano, MD, of the Icahn School of Medicine at Mount Sinai.
Background and Study Methods
The ECOG-ACRIN TAILORx trial demonstrated the benefit of the 21-gene recurrence score for stratifying recurrence risk in patients with breast cancer, and guiding chemotherapy use. However, the recurrence score has been more effective for predicting recurrence within the first 5 years of diagnosis than later recurrence. The researchers sought to make a model capable of predicting both early and late distant recurrences in patients with breast cancer.
The IICM+ model integrates clinicopathologic variables (patient age, tumor size, tumor grade, etc), transcriptomic features, and multiscale histopathology-derived image representations from tile-level embeddings and slide-level representations generated by a custom foundation model that was pretrained on paired histopathology images, RNA sequencing, and pathology data. It was developed on data and tumor specimens from 2,808 participants in the TAILORx trial with five-fold cross-validation. The remaining participants' data were used to evaluate and validate the model.
Patients in the TAILORx trial had received long-term follow-up of over a decade.
Key Findings
In the validation cohort (n = 1,621), the IICM+ model achieved a concordance index (C-index) of 0.735 (95% confidence interval [CI] = 0.681–0.782) for overall distant recurrence, 0.791 (95% CI = 0.715–0.858) for early distant recurrence, and 0.710 (95% CI = 0.645–0.773) for late distant recurrence. Comparatively, the Oncotype DX recurrence score achieved C-indices of 0.578 for overall distant recurrence (P < .001), 0.722 for early distant recurrence (P = .046), and 0.514 for late distant recurrence (P < .001).
Additionally, the IICM+ model stratified patients by high and low risk for overall distant recurrence (hazard ratio [HR] = 5.25; 95% CI = 3.50–7.86; P < .001). The model was still prognostic even after adjusting for clinicopathologic covariates and recurrence score categories (HR = 3.56; 95% CI = 2.22–5.70; P < .001).
The model was also able to identify groups discordant with recurrence score or IICM+ that showed different observed recurrence risks, which the authors noted supported additional prognostic stratification beyond recurrence score alone.
“Using multimodal data (clinical data, next-generation sequencing data, and data obtained from imaging) integrated through modern AI techniques allowed us to create a truly novel means of interrogating breast cancers for important prognostic information,” said senior author George W. Sledge, Jr, MD, Executive Vice President and Chief Medical Officer of Caris, which converted archived pathology slides from the TAILORx trial into digital images for the sake of the study as well as using deep learning methods to gain new insights from the samples.
The study authors added the findings do not indicate use of the IICM+ model for selecting or changing treatment. Further studies are currently in development to validate the model's use in other patient populations and to see if it can be used to guide treatment decisions.
DISCLOSURES: TAILORx was designed and conducted by ECOG-ACRIN with funding from the National Cancer Institute, part of the National Institutes of Health, and participation from the Alliance for Clinical Trials in Oncology, the Canadian Cancer Trials Group, NRG Oncology, and the SWOG Cancer Research Network. Additional funding was provided by the Breast Cancer Research Foundation, Susan G. Komen, and the Breast Cancer Research Stamp (U.S. Postal Service). For full disclosures of the study authors, visit nature.com.
