Prognostic performance of a multimodal artificial intelligence histopathology-based tool versus the 21-gene recurrence score in early-stage breast cancer.

559 Background: Genomic classifiers are widely used to assess risk of distant metastasis (DM) and inform chemotherapy (CT) benefit in patients with HR+/HER2- early breast cancer (EBC). While clinically validated, these assays are associated with high cost, long turnaround time, and tissue consumption. This study evaluates the performance of a histopathology based multimodal AI (MMAI) model in real-world institutional cohorts and compares its prognostic performance with the 21-gene recurrence score (RS). Methods: We conducted a retrospective analysis of 307 de-identified patients with node-negative HR+/HER2- EBC. Patients were treated at two institutions: Hoag Health System (n = 62) and the University of Virginia (n = 245). Median follow-up was 9.5 years. MMAI risk classifications (low vs. high) were compared to RS risk groups as defined in the TAILORx study (RS 0-10 (low), RS 11-25 (intermediate), RS≥26 (high)). The endpoint was time to DM and was analyzed using Kaplan-Meier (KM) methods and Cox proportional hazards models, including univariable (UVA) and multivariable analyses (MVA) adjusting for clinicopathologic factors. Results: In this study, 64% of patients were classified as MMAI low risk and 36% as MMAI high risk. The majority of patients received endocrine therapy alone (86% with low RS and 77% with intermediate RS), while 79% of patients in the high RS group were treated with CT. KM analyses demonstrated significantly lower 10-year DM rates for patients classified as MMAI low risk (1.6%, 95% CI: 0.4%-6.8%) compared with patients classified as MMAI high risk (10.8%, 95% CI: 5.5%-20.6%). Among patients with low RS, the MMAI model classified 84% as low risk, with no DM events observed in this group. In contrast, 16% of patients classified as MMAI high risk within the RS low group experienced a 10-year DM rate of 20% (95% CI: 5.4%-60.0%). MMAI further stratified intermediate RS patients into distinct risk groups: 66% were classified as MMAI low risk and 34% as MMAI high risk, with 10-year DM rates of 2.6% (95% CI: 0.6%-10.9%) and 11.4% (95% CI: 4.5%-27.2%) respectively. In UVA analysis of the intermediate RS group, MMAI high risk classification was associated with DM risk (hazard ratio (HR) of 4.2 (95% CI: 1.1-16.1, p = 0.03). In MVA, adjusting for RS groups, adjuvant systemic therapy, and age, MMAI remained independently associated with DM risk (HR = 5.8, 95% CI: 1.7-19.8, p = 0.005). Conclusions: MMAI demonstrated prognostic risk stratification within RS groups, particularly among intermediate RS. Concordance between MMAI low risk and favorable outcomes among patients with low RS suggests MMAI may identify patients with low risk of DM without requiring additional tissue consumption. Further studies are needed to evaluate outcomes among MMAI risk groups within high RS to define the optimal clinical integration of MMAI with existing genomic assays.