Background/Objectives: Breast cancer recurrence risk stratification has relied on gene expression tests that are limited by long turnaround times and consumption of valuable tissue. Artificial intelligence (AI) utilizing digital pathology images elucidates novel morphological biomarkers with strong prognostic associations, but the use of such AI models requires a modified analytical validation approach. Here, we report analytical validation of a novel breast cancer prognostic test. Methods: Ataraxis Breast RISK (ATX) uses a survival analysis model based upon features from a pan-cancer foundation model. This model extracts morphological features (biomarkers) from H&E-stained slides. These features are combined with clinical variables, and the test outputs a calibrated recurrence risk score. We validated ATX across five axes: intra-operator repeatability, inter-operator reproducibility, limit of blank, limit of detection and inter-laboratory reproducibility. Additionally, we assessed robustness to clinicopathologic data perturbations and conducted a clinical validation bridging study. Experiments were performed in CLIA-certified laboratories. Results: Intra-operator repeatability yielded an intraclass correlation coefficient (ICC) of 0.99 with 100% risk category agreement. Inter-operator reproducibility was concordant (ICC 0.99, 100% agreement). Inter-laboratory reproducibility across multiple scanners showed an ICC of 0.97 with 94.7% agreement. Under simulated clinicopathologic data perturbation, ATX maintained an average C-index of 0.62 with 90.0% agreement. The bridging study confirmed that the performance of the CLIA version was comparable to the prior clinical validation version (C-index 0.63 vs. 0.62). Conclusions: ATX met all predefined analytical acceptance criteria. These results support the analytical readiness of ATX use in clinical testing.
Dantone et al. (Sun,) studied this question.