Summary
A blinded evaluation of the AI model Ataraxis Breast CTX in SWOG-8814 found its scores were associated with disease-free survival. Results suggesting it could identify chemotherapy benefit were exploratory.
A medRxiv preprint posted October 23, 2026, reports that Ataraxis Breast CTX, an AI model analysing breast-tumour images and clinical variables, produced scores associated with disease-free survival in a blinded evaluation using data from the phase III SWOG-8814 trial. The study found a stronger prognostic signal from the combined image-and-clinical score than from an image-only score. Its findings on identifying patients more likely to benefit from chemotherapy were exploratory.
The analysis concerned postmenopausal patients with node-positive, hormone receptor-positive breast cancer. Genomic assays can help assess recurrence risk and chemotherapy benefit, but the preprint describes them as costly and tissue-consuming. The authors present CTX as a potential tissue-sparing approach; this evaluation tested its scores against outcomes in an existing trial cohort.
How the model was evaluated
CTX analyses digitised haematoxylin-and-eosin (H&E) slides, which show tissue structure and cell appearance under a microscope. Researchers assessed an image-only score and a multimodal score that combined pathology features with clinical variables.
The analytic cohort included 677 patients, or 46% of the 1,477 enrolled in SWOG-8814, who had digitised H&E slides. The model’s weights and thresholds were locked before evaluation, the trial data were not used for retraining, and predictions were generated without access to outcomes before being sent to SWOG for validation. Disease-free survival (DFS), the trial’s primary endpoint, was the primary outcome for this analysis.
Prognosis was clearer than chemotherapy-benefit prediction
Both scores were associated with DFS. The reported hazard ratio for the histology-only score was 1.36 (95% CI 1.22–1.51), and it added prognostic information beyond clinical variables alone (p<0.001). The multimodal score had a stronger reported association with DFS, with a hazard ratio of 1.60 (95% CI 1.44–1.78; p<0.001). Its C-index, a measure of how well a model distinguishes patients’ outcome rankings, was 0.66, compared with 0.59 for the histology-only score.
The treatment-benefit analysis asked whether the estimated effect of chemotherapy differed according to the CTX score. In an exploratory analysis treating the score as continuous, the treatment-by-biomarker interaction for chemotherapy plus endocrine therapy (CET) versus endocrine therapy (ET) had a hazard ratio of 0.82 (95% CI 0.64–1.04; p=0.053). The chemotherapy effect on DFS was numerically larger in the model’s high-benefit subgroup (480 patients; HR 0.79, 95% CI 0.59–1.05) than in its low-benefit subgroup (197 patients; HR 0.91, 95% CI 0.49–1.68). The interaction result was suggestive rather than conclusive, and the subgroup estimates were imprecise.
The clearest result is the association between CTX scores and prognosis in this trial cohort. The authors say prospective validation is needed to assess CTX as a tool for recurrence-risk assessment and treatment selection. The preprint has not been peer reviewed. It reports funding from Ataraxis AI, among other funders, and says several authors held equity in the company.