Test-fairness deep learning with influence score

Performance disparities in AI systems can manifest across sensitive groups or across data sources, especially when training data are collected from specific populations. In this work, we propose a feature-selection-based method that improves test-fairness while preserving prediction performance. Built on deep learning models, the proposed approach adopts the influence score (I-score), a statistical measure that captures interaction effects among multiple features. We identify features strongly associated with dataset membership by training an auxiliary model to predict dataset origin and applying I-score-based subset selection; these dataset-associated features are then excluded (masked) from the original prediction model for follow-up inference. We conduct experiments on two skin lesion datasets, ISIC 2019 and ASAN, collected from different populations. The empirical results show that the resulting fair I-score model can maintain high classification performance for skin lesion prediction while reducing cross-dataset subgroup performance disparity under our test-fairness evaluation setting.Author summary: Artificial intelligence is increasingly used to support medical image interpretation, but models that look accurate overall can still perform differently across patient populations or healthcare settings. In practice, medical image datasets often come from specific hospitals or regions, and images may differ in camera type, lighting, workflow, and population characteristics. As a result, a model may learn patterns that are tied to the data source rather than to clinically meaningful signs of disease, which can lead to unequal performance when the model is applied elsewhere. In this study, we present a practical framework to evaluate and reduce such uneven behavior without collecting new data or changing the original datasets. Using skin lesion classification as an example, we identify internal model signals that are strongly linked to the dataset source and reduce the model’s reliance on them. We test the approach on two widely used datasets from different populations (ASAN and ISIC 2019) and assess external generalization on a third dataset (PAD-UFES-20). The results show that the revised model maintains strong diagnostic performance while reducing cross-dataset performance disparity under our test-fairness evaluation setting, supporting more reliable deployment across clinical contexts.