External validation of precisebreast, a digital prognostic test for predicting breast cancer recurrence, in an early-stage cohort from the Netherlands

Abstract

Background: Current clinical guidelines recommend gene expression profiling to guide treatment in early-stage breast cancer. PreciseBreast (PDxBR) is a digital prognostic tool that integrates artificial intelligence (AI)-derived features from hematoxylin and eosin (H&E) slides with clinicopathologic data to predict recurrence risk. This study externally validated PDxBR in an independent cohort and compared its performance to other risk models. Methods: We retrospectively analyzed PDxBR in a cohort of 739 patients with early-stage hormone receptor-positive, HER2-negative breast cancer (median follow-up of 8.8 years). For each case, one H&E-stained slide was digitized and analyzed to generate recurrence risk scores using the full PDxBR model, as well as image-only and clinical-only variants. A subset of patients who underwent MammaPrint testing was also evaluated. Model performance was assessed by AUC/C-index, hazard ratios, sensitivity, specificity, and negative and positive predictive values (NPV and PPV, respectively). Results: PDxBR showed prognostic accuracy in this external cohort (AUC/C-index 0.71, 95% CI: 0.66–0.75). Applying the PDxBR threshold (≥ 58 versus < 58) yielded a hazard ratio of 3.05 (95% CI: 2.1–4.4, p < 0.001), sensitivity 0.70, specificity 0.59, NPV 0.90, and PPV 0.27. PDxBR outperformed the modified Adjuvant! Online clinical model (MINDACT model, p < 0.00001) and effectively reclassified grade 2 tumors into distinct risk groups. Conclusions: PDxBR demonstrated robust prognostic performance in an independent cohort, supporting its potential as a scalable, reproducible alternative to genomic assays for individualized risk stratification in early-stage breast cancer.