115 Evaluating an AI-Based Prognostic Test to Guide Adjuvant CDK4/6 Inhibitor Selection in Early HR+/HER2– Breast Cancer

Background

The phase 3 NATALEE clinical trial, which defined risk of recurrence (ROR) criteria, showed that ribociclib improves invasive disease-free survival among patients with stage II to III moderate and high-risk hormone receptor–positive (HR+)/ HER2-negative (HER2–) early breast cancer. As these ROR criteria rely exclusively on clinical features, we sought to determine whether incorporating pathology data enhances risk stratification. To do so, we evaluated Ataraxis Breast RISK (ATX-RISK), an artificial intelligence test which integrates morphological information from H&E-stained images with clinical information, to establish whether it can stratify patients by recurrence risk within clinical risk groups and whether its clinical utility is higher than clinical risk.

Methods

ATX-RISK scores were generated for 2228 patients with HR+/HER2– breast cancer. Patients received standard-of-care treatment which did not include CDK4/6 inhibitors. Patients were classified into clinical high and low risk groups based on NATALEE trial criteria. ATX-RISK generated 5-year risk of recurrence scores and a cut-off of 10% was used to classify patients into ATX high vs low groups. The primary end point was disease-free interval, evaluated using Kaplan-Meier curves with log-rank testing. ATX-RISK was not trained on any of the data sets used in this analysis.

Results

A total of918 (41%) patients were classified as clinical high risk and 1130 were classified as clinical low risk. In the clinical high group, ATX high patients had a significantly lower disease-free interval at 5 years (0.81; 95% CI, 0.76-0.85) than ATX low patients (0.93; 95% CI, 0.89-0.95; P_log-rank <.001). Similarly, patients in the clinical low group could also be stratified by ATX score. ATX high patients had a lower disease-free interval of 0.88 (95% CI, 0.78-0.94) than ATX low patients (0.96; 95% CI, 0.95-0.97; P_log-rank = .004).

A time-dependent decision curve analysis comparing ATX-RISK score, clinical risk criteria, and combined ATX score with clinical risk criteria was performed (Figure). At lower thresholds (0.05), the combined ATX score and clinical risk criteria conferred the highest net benefit (ATX-RISK only = 0.022, clinical risk only = 0.022, combined model = 0.023). However, as the threshold increased (0.10), ATX-RISK alone outperformed both the clinical risk criteria and combined model (ATX-RISK only = 0.008, clinical risk only = 0.000, combined model = 0.000).

Conclusions

ATX-RISK stratifies patients by recurrence risk, potentially redefining patient eligibility for adjuvant CDK4/6 inhibitors relative to current clinical criteria.