Evaluating performance of an AI-based recurrence score in early breast cancer patients treated with chemoendocrine therapy: A secondary analysis of the UNIRAD trial.

550 Background: Patients with node-positive HR+/HER2- early breast cancer are at high risk for relapse within 5 years of diagnosis, suggesting the potential need for treatment escalation. However, it is unclear which patients may experience worse recurrence-free outcomes and thus benefit from additional therapy. Ataraxis Breast (ATX) is an artificial intelligence (AI) test that integrates clinicopathologic variables with features extracted from whole-slide H&E images to estimate recurrence risk. Here, we perform a secondary analysis of the control arm of the UNIRAD trial, evaluating the ability of ATX to identify patients treated with standard-of-care therapy who may be candidates for treatment escalation. Methods: Clinical information and scanned H&E slides were sourced for 365 patients enrolled in the UNIRAD trial who were randomized to the control arm (standard-of-care therapy). ATX scores were generated using a locked model with pre-specified thresholds. No patients from UNIRAD were used in the training of ATX. Recurrence-free interval (RFI) was used as the primary endpoint. Kaplan-Meier estimators were used to predict the probability of meeting the RFI endpoint. To quantify relative differences in the hazard of experiencing an event contributing to the RFI endpoint associated with ATX scores, Cox proportional hazards models were fitted, from which hazard ratios (HRs) were estimated. The discriminative performance of ATX was evaluated using C-index. Results: Of the 365 patients randomized to the control arm of the UNIRAD trial with H&E slides available, 163 (45%) were classified as ATX low risk, and 202 (55%) as ATX high risk. Patients classified as ATX high risk had lower Kaplan-Meier-estimated probability of meeting the RFI endpoint (77%, 95% CI = 70-83%) than patients classified as ATX low risk (93%, 95% CI = 88-97%). Consistent with these findings, when modeled as a continuous variable, higher ATX scores were associated with a significantly higher hazard of an RFI-contributing event (HR = 1.57, 95% CI = 1.29-1.99, p-value < 0.001) and demonstrated strong discriminatory performance (C-index = 0.66, 95% CI = 0.59-0.72). This association remained significant after controlling for receipt of neoadjuvant therapy, tumor, and nodal stage (HR = 1.53, 95% CI = 1.05-2.23, p = 0.027). Conclusions: In the clinically homogenous UNIRAD trial cohort of patients with node-positive HR+/HER2- early breast cancer, ATX high risk patients treated with standard-of-care therapy had a significantly increased hazard of an RFI-contributing event. These findings suggest that AI-based risk stratification identifies biologically high risk patients who may derive benefit from adjuvant treatment escalation. Clinical trial information: NCT01805271 .