Benchmarking histology foundation models to identify Prosigna high-risk breast cancer patients.

e12572 Background: Currently, a variety of commercial molecular tests, including MammaPrint, Oncotype DX, and Prosigna, are used to guide adjuvant therapy decisions for early-stage breast cancer patients with hormone receptor-positive/HER2-negative phenotypes. Despite their effectiveness, these tests present challenges in terms of complexity, turnaround time, and cost, motivating the development of AI-based digital pathology tools as alternative prognostic methods. Methods: We benchmarked two self-supervised histology foundation models (UNI-v1 and Prov-GigaPath) for prediction of Prosigna-derived recurrence risk (ROR-PT) from paired H&E whole-slide images. Training and testing were performed using more than 7 million image tiles (224 × 224 pixels) derived from 534 breast cancer patients collected at the University Hospital of North Norway. Both models were fine-tuned using end-to-end task adaptation with slide-level supervision to classify patients into high-risk versus intermediate/low-risk groups. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC) and standard classification metrics. Results: In the overall cohort, molecular subtypes were distributed as follows: luminal A 65% (n = 347), luminal B 32% (n = 173), basal 2% (n = 8), and HER2-enriched 1% (n = 6). The Prosigna ROR-PT scores were dichotomized into low-risk and high-risk groups, comprising 427 (80%) and 107 (20%) patients, respectively. In the hold-out test set (n = 107, 20%), the UNI and Prov-GigaPath models achieved sensitivities of 0.82 and 0.73, specificities of 0.70 and 0.78, and AUCs of 0.79 and 0.81, respectively. While Prov-GigaPath demonstrated slightly higher precision (PPV), the UNI model showed higher sensitivity, suggesting a broader identification of high-risk patients (Table). In addition, model ensembling did not show performance gains over the best unimodal model. Furthermore, a strong correlation (Pearson’s r = 0.65, p < 0.001) was observed between the predictions of both foundation models and the continuous Prosigna ROR scores. Conclusions: The findings of this initial study demonstrate that AI-driven analysis of routine H&E images represents a promising, cost-effective, and rapid surrogate to Prosigna for prediction of recurrence risk in breast cancer. These results will be further validated in independent external cohorts. Model AUC (CI, 95%) Sensitivity Specificity PPV NPV UNI 0.79 (0.65-0.89) 0.82 0.70 0.42 0.94 Prov-GigaPath 0.81 (0.76-0.96) 0.73 0.78 0.46 0.92