Prognostic versus treatment-confounded stratification in CBCS

Determine whether the observed stratification of recurrence risk produced by attention-based multiple instance learning models using pathology foundation model features from H&E-stained whole slide images of the Carolina Breast Cancer Study cohort reflects purely prognostic information inherent to tumor morphology, or whether it is partially confounded by heterogeneous treatment regimens received by participants, in order to establish the causal basis of the risk stratification.

Background

The models were developed and evaluated on the Carolina Breast Cancer Study (CBCS), a cohort in which participants received heterogeneous treatments. While the models showed strong alignment with transcriptomic ROR-P and stratified patients by recurrence, the non-uniform treatment exposure in CBCS complicates causal interpretation of the observed stratification.

Because CBCS is not a randomized clinical trial or a uniformly treated cohort, treatment effects may confound the relationship between histology-derived predictions and recurrence outcomes. Clarifying whether stratification is prognostic versus treatment-confounded is essential for clinical translation and for understanding the biological signals captured by H&E-based models.

References

As a result, we cannot determine whether the observed stratification of recurrence risk reflects prognostic information alone or is partially confounded by treatment effects.

— Towards interpretable prediction of recurrence risk in breast cancer using pathology foundation models  (2508.12025 - Kaczmarzyk et al., 16 Aug 2025) in Discussion — Limitations

We cannot fully separate treatment intensity from severity in this dataset, but the inversion surviving every cohort policy tested makes exclusion the weaker candidate.

— Fixing a Model That Learned Worse Cancer Means Lower Risk: Monotonic Constraints in Bladder Cancer Recurrence Prediction  (2610.00858 - Abbas et al., 1 Oct 2026) in Discussion, subsection “Why the inversion shows in the unconstrained model”