Evaluating the prognostic performance of an AI-based model in clinically low-risk HR+/HER2− early breast cancer.

552 Background: Accurate prediction of recurrence risk in early HR+/HER2- breast cancer remains central to adjuvant treatment decision making. Clinically approved prognostic models largely rely on genomic features, yet their performance remains insufficient for optimal clinical stratification. To address these limitations, AI-based approaches have increasingly been developed to improve risk prediction. Ataraxis Breast (ATX) is a multimodal artificial intelligence test integrating clinical information with morphological information from H&E-stained whole slide images for HR+/HER2- patients. Here, we evaluate ATX’s ability to prognosticate recurrence risk in an external cohort of clinically low risk early breast cancer patients from Dordrecht, the Netherlands. Methods: Clinical information and H&E images were accessed for 887 patients with HR+/HER2- breast cancer. ATX scores were generated using a locked model and a prespecified cut-off was applied to classify patients into ATX high vs low groups. The primary endpoint was recurrence-free interval (RFI); recurrence-free probabilities were estimated for the two groups using a Kaplan-Meier estimator. Additionally, performance of ATX was assessed using C-index and hazard ratio (HR per 0.1 increase in ATX score). No patients from this dataset were used in the training of ATX. Results: The cohort of 887 patients was characterized by low clinical risk, where 631tumors were classified as T1 and 239 as T2, and the majority of patients were node-negative (N0: n = 639). Tumors were predominantly low to intermediate grade (grade 1: n = 296; grade 2: n = 496), with relatively few high-grade cases (grade 3: n = 88). Histological analysis identified invasive lobular carcinoma in 131 patients. The median tumor size was 15 mm (interquartile range, 11-22 mm). Accordingly, most patients (n = 773, 87%) were classified as ATX low risk. Patients in the ATX high risk group had lower estimated 5-year probability of meeting the RFI endpoint (0.88, 95% CI = 0.80-0.93) than patients classified as ATX low risk (0.98, 95% CI = 0.96-0.98). When modeled as a continuous variable, ATX demonstrated strong discriminatory performance for recurrence risk (C-index = 0.740, 95% CI, 0.68-0.80) and was associated with a significantly increased hazard of an RFI-contributing event (HR = 2.68, 95% CI, 2.09-3.45; p < 0.001). To account for clinical confounding, we fitted a multivariate Cox proportional hazards model adjusting for T stage, N stage, grade, age at diagnosis, adjuvant endocrine therapy, adjuvant chemotherapy, and tumor size. We found that ATX score remained significant (HR = 1.99, 95% CI = 1.38-2.86, p < 0.001) and tumor size was the only other variable that was statistically significant (p < 0.05). Conclusions: ATX accurately prognosticates recurrence risk in a clinically low risk group of HR+/HER2- breast cancer patients from an external validation cohort.