Investigators evaluated the Ataraxis Breast CTX model using archived material from the phase III SWOG-8814 trial. The analytic cohort comprised 677 patients with digitized H&E slides; the authors noted this represented 46% of the original 1,477 enrolled. CTX weights and thresholds were locked prior to the validation and SWOG-8814 data were not used for retraining. Predictions were produced blinded to clinical outcomes and submitted to SWOG for independent validation.
The primary endpoint for the validation matched the trial’s primary endpoint: disease-free survival (DFS). The study team reported that the histology-only CTX score was prognostic for DFS and added prognostic information beyond standard clinical variables (investigators report statistical significance). The multi-modal CTX score—integrating pathology features with clinical data—showed a stronger association with DFS than the histology-only score. Data indicated improved discrimination for the multi-modal score, with a reported C-index of 0.66 versus 0.59 for the histology-only classifier.
These results suggest that an AI model trained to combine image-derived features with clinical variables can stratify recurrence risk in an archival, node-positive, hormone receptor–positive population. The authors emphasized that CTX’s prognostic signal persisted despite the retrospective, pre-specified locked evaluation design. The validation was blinded; this reduces certain sources of bias, although sample selection and availability of digitized tissue shaped the analytic cohort.
Two CTX outputs were compared head-to-head. One relied solely on histopathology images; the other fused those histology-derived features with clinical covariates. The histology-only output showed a measurable association with DFS. The multi-modal score strengthened that association and increased model discrimination, as reflected in the higher C-index reported by the authors.
Investigators provided effect estimates for both outputs and tested whether the histology-only score added information beyond routine clinical factors; they reported that it did. The multi-modal CTX, by design, leverages both tissue-level patterns and clinical context. That combined approach produced the larger effect estimate and superior discrimination in this cohort. The authors did not report additional calibration metrics or details of which clinical covariates drove the incremental gain; the source did not provide those specific methodological breakdowns.
The team conducted exploratory analyses to evaluate whether CTX predicted benefit from adjuvant chemotherapy. When modeled as a continuous variable, CTX produced a treatment-by-biomarker interaction with a reported hazard ratio of 0.82 (confidence interval crossing 1) and a p value near conventional significance thresholds (p = 0.053) for CET versus ET. Investigators described this as suggestive evidence of a predictive interaction rather than definitive proof.
They also examined chemotherapy effect separately in CTX-defined subgroups. The chemotherapy-associated DFS reduction was numerically larger in the high-benefit subgroup (reported HR 0.79) than in the low-benefit subgroup (reported HR 0.91). Confidence intervals for these subgroup estimates included unity for both groups. The authors treated these subgroup comparisons as exploratory and recommended cautious interpretation given the retrospective nature of the analysis and the absence of a prospective test of treatment selection based on CTX.
Data used for this external validation are not publicly available due to institutional and ethical constraints; the team notes that access to SWOG-8814 data can be requested directly from SWOG under data-use agreements. The preprint discloses competing interests: multiple authors are equity holders of Ataraxis AI. Other authors reported institutional research support and patent holdings; specifics are enumerated in the source. Funding sources named in the manuscript include the National Cancer Institute, The Hope Foundation for Cancer Research, and Ataraxis AI.
The manuscript is a preprint and has not undergone peer review. The study team stated that CTX thresholds and model parameters were pre-locked before this blinded validation, and that SWOG-8814 samples were held out from any model development—a design intended to provide an external test of performance rather than an internal cross-validation.
Several limitations are explicit in the source. The analytic cohort represented a subset (677 of 1,477) of the original trial population, determined by availability of digitized H&E slides; selection bias related to tissue availability cannot be excluded. The exploratory treatment-by-biomarker interaction yielded a p value of 0.053—suggestive but not definitive—and subgroup hazard ratios had confidence intervals that crossed one. The source did not report certain methodological details, such as which clinical variables contributed most to the multi-modal gain or additional calibration measures; those specifics were not provided in the preprint.
Investigators conclude that the blinded validation supports CTX’s prognostic value and offers suggestive evidence for predictive utility, but they call for further prospective validation. The authors recommend a prospective study to test CTX as a tissue-sparing assay for recurrence-risk assessment and treatment selection. The preprint does not specify regulatory milestones, timelines for prospective trials, or plans for public data release; those elements were not reported in the source.