A new preprint on arXiv highlights a fundamental problem in cross-cohort evaluation of breast cancer molecular subtyping models. The authors point out that source and target cohorts often use different annotation standards, making direct comparisons unreliable. For example, METABRIC labels six classes while PAM50 defines five, so a model trained on one system may produce predictions that have no equivalent in the other.
The abstract specifically notes that a claudin-low prediction cannot be meaningfully evaluated in a cohort whose label space does not include that class. This mismatch distorts performance metrics and can lead to incorrect conclusions about how well a classifier generalizes.
The preprint does not propose a full solution in the provided text, but the implication is clear: researchers must verify that label spaces are aligned before performing cross-cohort evaluation, or risk reporting artifacts as scientific findings.