Summary
A medRxiv preprint reports that Ataraxis Breast CTX, a multimodal AI model using clinical data and H&E images, separated five-year disease-free survival in a selected group of 556 clinically high-risk, node-positive HR+/HER2- early breast cancer patients. An exploratory analysis found that everolimus benefit varied by the model’s score.
A medRxiv preprint describes a multimodal artificial-intelligence model that separated residual recurrence risk among clinically high-risk patients with node-positive, hormone receptor-positive and HER2-negative (HR+/HER2-) early breast cancer. In a selected group of 556 patients, five-year disease-free survival was 93% in the model’s low-risk group and 80% in its high-risk group.
The study also reports an exploratory finding linking the model’s score to benefit from adjuvant everolimus, a treatment used after initial therapy. The work was posted on September 21, 2026, as a preprint.
What the model assessed
HR+/HER2- breast cancer is a subtype whose biology includes hormone-receptor signalling but not HER2 overexpression. Node-positive disease means cancer has been detected in nearby lymph nodes, a factor associated with a higher risk profile. The study focused on patients considered clinically high risk even after receiving adjuvant chemotherapy and endocrine treatment.
The model, called Ataraxis Breast CTX, combines clinical variables with features extracted from haematoxylin-and-eosin (H&E) histopathology images. H&E staining is a standard method used to show tissue structure and cellular features under a microscope. The researchers describe two outputs: CTX-prognostic, which estimates recurrence risk, and CTX-benefit, which estimates treatment benefit.
The evaluation was a prospective-retrospective validation using data from the phase III UNIRAD trial. The analysis included patients who had not received neoadjuvant chemotherapy, had received adjuvant chemotherapy and had H&E slides available. Disease-free survival was the primary endpoint.
Risk separation and everolimus signal
Across the analysed patients, those classified as low risk had a five-year disease-free survival of 93%, compared with 80% for those classified as high risk. The authors report that this difference was statistically significant. In practical terms, the model separated a group with favourable outcomes under chemoendocrine therapy from a group with higher residual risk.
The researchers then examined whether CTX-benefit was associated with the effect of adjuvant everolimus. In an exploratory multivariable analysis, the relationship between the score and everolimus treatment benefit remained statistically significant after adjustment for age, tumour size, nodal status, grade and menopausal status. The reported interaction p-value was 0.01.
An interaction analysis tests whether a treatment’s apparent benefit varies across groups defined by another factor—in this case, the model’s score. The finding therefore supports further study of CTX as a tool for selecting patients for treatment escalation, rather than treating all clinically high-risk patients as having the same residual risk.
The study’s immediate contribution is the combination of routine clinical information and digital pathology features in a model intended to guide post-treatment risk assessment. The authors also identify possible applications involving additional therapies such as everolimus or CDK4/6 inhibitors, although the supplied results specifically report the exploratory everolimus analysis.
Evidence level and study boundaries
This is a retrospective reanalysis of a selected subset from the UNIRAD clinical trial, presented as a medRxiv preprint. The everolimus analysis was exploratory, while the primary study endpoint was disease-free survival. These features make the results relevant for prospective evaluation of risk-guided treatment escalation, but the preprint does not describe a clinical decision threshold for using CTX in routine care.
The abstract also does not provide an effect estimate or confidence interval for the everolimus finding. The study was funded by Ataraxis AI, and several authors report holding equity in the company; the other authors declared no competing interests. The underlying data are not publicly available because of institutional and ethical constraints, although access can be requested from Unicancer.