Biomarker Robustness

High classification accuracy alone cannot show that a model has learned robust or neurobiologically meaningful biomarkers. The findings interpret classification accuracy as inadequate for biomarker discovery. Cross-entropy optimization does not force models to prioritize neurobiologically faithful features. Models can achieve high accuracy while producing non-reproducible or weakly supported salient features. Future biomarker studies should assess robustness and neurobiological faithfulness instead of relying on classification performance.