Determine which modality drives progression-performance loss under cohort shift

Identify which input modality—structured clinicopathological data, histopathology whole-slide images, or RNA-sequencing data—contributes most to the loss of progression-prediction performance when models are transferred from the Erasmus Cohort B distribution to the Urolife cohort.

Background

The benchmark found that progression-prediction models performed worse on the Urolife cohort than on Erasmus Cohort B and that cohort shift was detectable in structured variables, histopathology features, and transcriptomic data. However, the observed performance degradation could not be attributed to any single modality because the submitted models differed in preprocessing, representation, fusion, and missing-data handling. Controlled experiments using matched architectures and systematic modality combinations are therefore needed to isolate the modality responsible for the transportability loss.

References

Cohort shift was evident across multiple input domains, but the analyses cannot identify which modality contributed most to the performance loss.

— CHIMERA Challenge Task 2 and 3: Response Subtypes Classification and Progression Survival Prediction in Bladder Cancer Patients using Multimodal Datasets  (2609.09510 - Chia et al., 8 Sep 2026) in Section 4, Discussion, subsection “Cohort shift and model transportability”