Equivalence of histology-based ROR-P predictions to transcriptomic assay in homogeneous treatment settings

Ascertain whether attention-based multiple instance learning models using pathology foundation model features to predict PAM50-based ROR-P from H&E-stained whole slide images can match the clinical performance of the transcriptomic PAM50-based ROR-P assay in guiding treatment decisions and in predicting outcomes in cohorts with uniform treatment protocols, thereby verifying equivalence in homogeneously treated populations.

Background

Although the histology-based models aligned well with transcriptomic ROR-P by ROC AUC, Pearson correlation, and C-index within CBCS, the authors caution that this does not establish equivalence to the assay in guiding treatment or predicting outcomes under homogeneous treatment conditions.

Determining whether image-based models match the clinical utility of the transcriptomic assay requires evaluation in settings such as randomized clinical trials or uniformly treated cohorts, which were not available in CBCS.