Clinical validation of a multi-modal Ataraxis AI platform for recurrence prediction in early-stage breast cancer across multiple patient cohorts.
Jan Sylwester Witowski, Khalil Choucair, Jailan A. Elayoubi, Elena Diana Chiru, Nancy Gain Chan, Young-Joon Kang, Frederick Matthew Howard, Irina Ostrovnaya, Freya Ruth Schnabel, Waleed Abdulsattar, Yu Zong, Lina Daoud, Marcus Vetter, Jia Wern Pan, Arvydas Laurinavičius, Brian Piening, Carlo Bruno Bifulco, Adam Brufsky, Francisco J. Esteva, Lajos Pusztai · Journal of Clinical Oncology · 2025
549 Background: Breast cancer (BC) treatment selection is traditionally guided by clinical characteristics. However, as clinical characteristics cannot capture the complexity of a disease, genomic tools have been developed. Recent advances in artificial intelligence (AI) have allowed pathology imaging to be used to build more accurate and comprehensive prognostic/predictive models. In this study, we validated an AI test, powered by a pan-cancer histopathology foundation model, that integrates digital pathology images with clinical variables to predict breast cancer recurrence. Methods: The Ataraxis AI prognostic model (ATX) was developed using 4,659 stage I-III BC patients from 10 distinct cohorts. Ataraxis AI platform first extracts novel morphological features from digitized H&E slides using a pre-trained AI foundation model. These morphological features are then integrated with common clinical characteristics, such as TNM staging, ER/PR/HER2 status, age at diagnosis, or lobular or ductal histology to generate a risk score between 0 and 1. We evaluated ATX on 3,502 patients from 5 external cohorts, including 858 patients with available Oncotype DX (ODX) scores. The primary endpoint of this study was disease-free interval (DFI), defined as the time until first recurrence, with deaths prior to recurrence censored. Results: Across 3,502 patients spanning five validation cohorts, ATX accurately predicted DFI with a C-index of 0.71 [0.68-0.75] and hazard ratio (HR) of 3.63 [3.02-4.37, p < 0.01], computed for every 0.2 unit increase in the test score. Compared to ODX (n = 858), the ATX was more accurate, achieving a C-index of 0.67 [0.61-0.74] versus 0.61 [0.49-0.73]. Additionally, ATX added independent prognostic information to ODX in a multivariate analysis (HR: 3.11 [1.91-5.09, p < 0.01]). ATX demonstrated robust accuracy in TNBC (n = 230, C-index: 0.71 [0.62-0.81], HR: 3.81 [2.35-6.17, p = 0.02]) and HER2+ (n = 353, C-index: 0.67 [0.55-0.80], HR: 2.22 [0.99-5.01, p = 0.05]) groups. Conclusions: (1) ATX is predictive of breast cancer recurrence, (2) ATX improves upon the accuracy of ODX, (3) ATX demonstrates robust performance in all main BC subtypes. ATX evaluated across 5 cohorts individually and pooled, for both Harrell’s C-index and hazard ratio. Cohort N C-index HR Karmanos 168 0.62 [0.49-0.75] 3.82 [1.33-10.98, p=0.01] Basel 269 0.67 [0.58-0.77] 3.98 [1.92-8.25, p<0.01] TCGA 911 0.70 [0.63-0.77] 3.0 [2.1-4.28, p<0.01] Providence 1733 0.74 [0.7-0.79] 4.02 [3.09-5.23, p<0.01] Chicago 421 0.70 [0.60-0.80] 3.25 [1.45-7.31, p<0.01] Pooled 3502 0.71 [0.68-0.75] 3.63 [3.02-4.37, p<0.01]