Abstract PS3-04-08: Independent Validation of a Pathology-Based Multimodal Artificial Intelligence Biomarker for Predicting Risk of Distant Metastasis in Postmenopausal, Estrogen Receptor-Positive, Early-Stage Breast Cancer Patients: Analysis of the ABCSG Trial 8

Background: Genomic assays are currently used to determine prognosis and predict treatment outcome in patients with estrogen receptor-positive (ER+), HER2-negative early breast cancer (EBC). However, their use is expensive, time-consuming and not available in many clinical facilities. We previously developed a multimodal artificial intelligence (MMAI) model using clinical and histopathological data from more than 12,000 patients from six phase III trials, offering a faster, non-genomic alternative for personalized risk stratification. We present the first external independent validation of this MMAI model using the ABCSG trial 8 - a large, prospective phase III trial of a low- to moderate-risk cohort of ER+ postmenopausal breast cancer patients receiving endocrine therapy (ET) only - to assess its prognostic performance for distant metastasis (DM). Methods: ER+ HER2- patients from the prospective-randomized trial ABCSG 8 with digitized baseline H&E-stained sections, complete clinical data (age, tumor size and nodal status), and long-term follow up were included in this pivotal validation study. The locked MMAI model includes a continuous score and predefined risk groups. Prognostic performance was assessed using univariable and multivariable Cox proportional hazard models, with hazard ratios (HR) and corresponding 95% confidence intervals (CI). Results: MMAI raw scores were generated for 2,109 patients. The median follow-up time was 9.5 years. Despite the overall lower risk profile (G1 and G2 only) of the ET-only population, the MMAI classified 77% patients as low risk, 9% as intermediate risk, and 14% as high risk. The estimated 10-year DM-free rates were 94.7% (93.3% - 95.8%) for low-risk patients, 89.1% (82.5% - 93.3%) for intermediate-risk patients, and 77.3% (71.3% - 82.2%) for high-risk patients. In univariable analysis (UVA), both the continuous MMAI raw score and MMAI risk groups were significantly associated with risk of DM: MMAI raw score (HR [95% CI] = 2.22 [1.91 - 2.58], p<0.001) and MMAI risk group (intermediate vs. low risk HR [95% CI] = 2.19 [1.34 - 3.57], p=0.002; high vs. low risk HR [95% CI] = 4.45 [3.19 - 6.19], p<0.001). In multivariable analysis (MVA) adjusting for age, tumor size, nodal status, both continuous MMAI raw score (p<0.001) and MMAI risk group (p<0.001) remained independently associated with DM. The MMAI raw score remained significantly associated with the risk of DM across clinically meaningful subgroups, including lymph node status (N0 and N1), tumor grade (Grade 1 and 2), histology (invasive ductal carcinoma and invasive lobular carcinoma), and trial-defined Ki67 levels (≤5%, 6-29%, ≥30%). Additionally, the image-only component of the MMAI was independently associated with risk of DM (p<0.001 in both UVA and MVA). Conclusion: This study validates the MMAI model, originally developed for HR+ HER2- EBC, as a strong prognostic tool for DM in lower-risk postmenopausal women with ER+ HER2- EBC. This MMAI technology promises to improve personalized risk stratification and adjuvant treatment decisions in ER+ HER2- EBC as a cost-effective, non-tissue consuming, and faster alternative to genomic assays. Citation Format: M. Filipits, D. Hlauschek, J. Zhang, M. Balic, R. Kates, R. Greil, F. Fitzal, N. Toro-Bauer, G. Rinnerthaler, H. Pinckaers, K. Sotlar, Z. Bago-Horvath, W. Hulla, P. Regitnig, A. Piehler, C. E. Geyer, H. Kreipe, J. Griffin, N. Harbeck, N. Wolmark, M. Gnant. Independent Validation of a Pathology-Based Multimodal Artificial Intelligence Biomarker for Predicting Risk of Distant Metastasis in Postmenopausal, Estrogen Receptor-Positive, Early-Stage Breast Cancer Patients: Analysis of the ABCSG Trial 8 [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-04-08.