Abstract
Artificial Intelligence (AI) has demonstrated success in computationalpathology (CPath) for disease detection, biomarker classification, andprognosis prediction. However, its potential to learn unintended demographicbiases, particularly those related to social determinants of health, remainsunderstudied. This study investigates whether deep learning models can predictself-reported race from digitized dermatopathology slides and identifiespotential morphological shortcuts. Using a multisite dataset with a raciallydiverse population, we apply an attention-based mechanism to uncoverrace-associated morphological features. After evaluating three dataset curationstrategies to control for confounding factors, the final experiment showed thatWhite and Black demographic groups retained high prediction performance (AUC:0.799, 0.762), while overall performance dropped to 0.663. Attention analysisrevealed the epidermis as a key predictive feature, with significantperformance declines when these regions were removed. These findings highlightthe need for careful data curation and bias mitigation to ensure equitable AIdeployment in pathology. Code available at:https://github.com/sinai-computational-pathology/CPath_SAIF.