AI driven digital pathology for the 70 gene recurrence signature: Development and validation in early stage breast cancer

Digital MammaPrint AI assay

Patient / Population Intervention / Exposition Comparaison Critère de jugement

In patients with early-stage breast cancer, an AI-based digital 70-gene assay from routine H&E slides identified a low-risk group with better 10-year breast cancer specific survival.

N

372patients

Schéma

Clinical validation of an AI digital pathology model, trained on 70,000 H&E whole-slide images with MammaPrint-derived supervision, in 372 patients with early-stage breast cancer from the prospective real-world RASTER study, comparing digital high-risk (D-High) with digital low-risk (D-Low) groups.

Critère

Breast cancer specific survival (BCSS) and distant recurrence free interval (DRFI) at a median 10.3-year follow-up, with concordance against MammaPrint.

RésultatConcordance with MammaPrint r = 0.954. 49% (n=183) were classified D-High and 51% (n=189) D-Low. Ten-year BCSS 98.1% for D-Low vs 86.7% for D-High (HR=0.12, 95% CI 0.04-0.39, P<0.0001). In ER+/HER2- patients (n=274), 10-year BCSS 98.1% vs 84.7% (HR=0.10, 95% CI 0.03-0.36). Among ER+/HER2- clinically high risk patients, D-Low patients had no breast cancer deaths. In clinically low risk patients, 10-year BCSS 83.8% for D-High vs 97.5% for D-Low. In multivariable analysis digital assay risk was the only independent predictor of BCSS (HR=0.084, 95% CI 0.02-0.31; P=0.0002). Similar results were obtained for DRFI. Chemotherapy was received by 74% of D-High and 24% of D-Low patients.

Van't Veer LJ, Hilbers FS, Kilickaya M, et al. Presented at ESMO Congress 2026, abstract LBA20, 24 October. Not yet published.

Discussion et critique

The image-based assay tracked MammaPrint closely and separated 10-year breast cancer specific survival in a cohort with long follow-up, including within clinical risk groups. Validation rests on a single cohort of 372 patients, the abstract gives no event counts or DRFI figures, and the wide confidence intervals suggest few events. Concordance is reported as a correlation coefficient rather than agreement on high versus low risk categories, and chemotherapy was given far more often to D-High patients, so the data show prognosis and not who benefits from chemotherapy. Many authors are employed by or paid by Agendia, which the source names.