Cancer and Precision Medicine · global
Reading Recurrence Risk from Breast Cancer Slides: Agendia Announces AI Model Validation with Ten-Year Follow-Up
An investigational digital pathology model aims to estimate risk information from a 70-gene test using routinely stained slides. Long-term data show that it can distinguish prognostic groups, but evidence gaps remain before it could replace genomic testing or guide treatment decisions.
Whether additional treatment is needed after breast cancer surgery often depends on the likelihood of future tumor recurrence. If routine pathology slides could also provide risk clues that otherwise require genomic testing, more patients might gain access to this type of assessment. Agendia recently announced validation results for an investigational artificial intelligence model, showing that risk groups estimated from slide images were associated with long-term outcomes in patients with early-stage breast cancer.
The study was presented by company co-founder and Chief Research Officer Laura van ’t Veer at the 2026 European Society for Medical Oncology (ESMO) Congress. The model was trained using results from MammaPrint, a 70-gene recurrence risk test, and 70,000 de-identified hematoxylin and eosin (H&E)-stained whole-slide images. Through tissue recognition, classification, and image feature extraction, it generates a continuous risk score and then divides patients into digital high-risk and low-risk groups. It estimates the risk information reflected by the genomic test rather than directly measuring gene expression.
The company’s results announcement used data from 372 patients with early-stage breast cancer in the RASTER study, with a median follow-up of 10.3 years. The correlation coefficient between the model and MammaPrint results was 0.954, with approximately half of patients assigned to each of the high-risk and low-risk groups. This figure describes the degree of correlation between the two sets of results; it cannot be interpreted as a prediction accuracy of 95.4% for the model.
More clinically relevant evidence comes from long-term follow-up: ten-year breast cancer-specific survival was 98.1% in the digital low-risk group and 86.7% in the high-risk group. Among 274 patients with estrogen receptor-positive, HER2-negative disease, the corresponding figures were 98.1% and 84.7%; coverage of the same announcement by *Clinical Lab Products* also reported this subgroup result. However, breast cancer-specific survival measures the proportion of patients who have not died from breast cancer and cannot be treated directly as the proportion who have remained recurrence-free.
The conference abstract states that the study evaluated performance using Kaplan–Meier survival analysis and multivariable Cox models adjusted for clinicopathological factors. The company subsequently reported that digital risk classification remained significantly associated with breast cancer-specific survival in multivariable analysis, with a hazard ratio of 0.084 and a p-value of 0.0002. However, separation between prognostic groups does not establish that increasing or decreasing chemotherapy based on the model’s results can improve patient outcomes.
Two further points require clarification when interpreting the data. RASTER itself was a prospective study, but this AI model validation was a retrospective analysis of existing data and cannot be regarded as a prospective trial testing AI-guided treatment. In addition, the conference abstract lists 376 patients, while the company’s results announcement reports 372; the publicly available text does not currently explain the difference. The abstract still describes the study in terms of “expected results” and does not include the full numerical results reported in the subsequent announcement.
The appeal of this approach lies in using slides that are already routinely prepared to explore a more accessible route to risk assessment; the announcement has not yet provided comparative evidence on whether it can actually save time or reduce costs. The model remains for research use. Its consistency across hospitals, staining workflows, and scanning equipment, as well as how cases with results that disagree with genomic testing are handled, will influence subsequent clinical validation and regulatory assessment. The current results are not sufficient to support replacing MammaPrint with the model or using it to make treatment decisions directly.