AI-based prognosis and chemotherapy benefit in node-positive, HR+ breast cancer: a blinded external validation on SWOG-8814 trial
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
The benefit from adjuvant chemotherapy varies among patients with postmenopausal, node-positive, hormone receptor-positive breast cancer. Genomic assays identify patients likely to benefit from chemotherapy but are expensive and tissue consuming. AI-enabled biomarkers may provide scalable alternatives. We performed a blinded external evaluation of Ataraxis Breast CTX, a causal multi-modal AI model that integrates H&E images with clinical variables, in the phase III SWOG-8814 trial.
Methods
The analytic cohort comprised 677 patients (46% of the 1477 enrolled) with digitized H&E slides. Two CTX scores were evaluated: a histology-only score and a multi-modal score integrating pathology features with clinical variables. CTX weights and thresholds were locked before evaluation and SWOG-8814 data were not used for retraining. Predictions were generated blinded to outcomes and provided to SWOG for validation. The primary endpoint was disease-free survival (DFS), matching the primary endpoint of the trial.
Findings
The histology-only score was prognostic for DFS (HR = 1.36, 95% CI = 1.22-1.51) and added prognostic information beyond clinical variables alone (p < 0.001). The multi-modal score showed a stronger association with DFS (HR = 1.60, 95% CI = 1.44-1.78, p < 0.001) and greater discrimination than the histology-only score (C-index 0.66 vs 0.59). In exploratory analyses, modeled as a continuous score, CTX-benefit showed a treatment-by-biomarker interaction of HR of 0.82 (0.64-1.04, p = 0.053) for CET versus ET. The chemotherapy effect on DFS was numerically larger in the high-benefit subgroup (n = 480, HR = 0.79, 95% CI = 0.59-1.05) compared to the low-benefit subgroup (n = 197, HR = 0.91, 95% CI = 0.49-1.68).
Interpretation
In this blinded validation, CTX was significantly associated with DFS, supporting its prognostic value. In exploratory analyses, CTX demonstrated suggestive evidence of a treatment-by-biomarker interaction for chemotherapy benefit. These findings support further prospective validation of CTX as a tissue-sparing tool for recurrence-risk assessment and treatment selection.
Funding
National Cancer Institute, The Hope Foundation for Cancer Research, and Ataraxis AI.
Research in context
Evidence before this study
We searched PubMed for studies published between Jan 1, 2016, and Sept 1, 2026, using the terms (“digital pathology” OR “deep learning” OR “artificial intelligence”) AND (“hematoxylin” OR “haematoxylin” OR “H&E”) AND “breast cancer” AND (“chemotherapy benefit” OR “predictive biomarker”), without language restrictions. Multiple AI models based on H&E-stained whole-slide images, with or without clinical variables, have been developed to estimate recurrence risk in early hormone receptor-positive (HR+) breast cancer. However, robust evidence for AI-based prediction of chemotherapy benefit remains limited as most models either infer the 21-gene Recurrence Score or estimate baseline recurrence risk. To date, no AI-based markers have been shown to be predictive of chemotherapy benefit in a fully blinded, prespecified analysis of randomized clinical trial data.
Added value of this study
This is a blinded, pre-specified external validation of an AI-based multi-modal biomarker in SWOG-8814, a phase III randomized chemotherapy trial in postmenopausal, node-positive, hormone receptor positive breast cancer. The AI model weights and thresholds were locked prior to evaluation, no data from the trial was used for model development, and predictions were analyzed independently by the SWOG statistical center. CTX was significantly prognostic for SWOG-8814’s primary endpoint (disease-free survival) across the full cohort, within both treatment arms, and in the 1-3 positive-node subgroup. CTX also added prognostic information beyond established clinical covariates. In exploratory analyses of chemotherapy benefit, the treatment-by-biomarker interaction did not reach statistical significance, though the magnitude of chemotherapy benefit appeared to vary by subgroup.
Implications of all the available evidence
These findings support CTX as an AI-based approach to recurrence-risk assessment in postmenopausal women with node-positive HR+ breast cancer, a group in whom decisions about adjuvant chemotherapy remain challenging. As CTX derives prognostic information from routinely collected H&E-stained slides, it avoids the tissue consumption, cost, and turnaround delays that limit access to genomic assays. Prospective validation in contemporary node-positive cohorts receiving taxane-based chemotherapy and aromatase inhibitors will be required before CTX-guided treatment selection can be recommended in routine practice.
