Attribute-Aware Adversarial Domain Augmentation for Zero-Shot Medical Domain Adaptation

Deep learning-based medical diagnosis has demonstrated remarkable performance in in-distribution environments, whereas it remains vulnerable in out-of-distribution (OOD) scenarios, often producing unreliable predictions when applied to patients from unseen domains. Medical data involves diverse domains characterized by patients' attributes, such as age, which implies a high risk of encountering OOD instances. Domain generalization is a practical approach to overcome domain gaps, which does not require target data often unavailable due to strict privacy regulations for medical data. However, the lack of knowledge about a target domain limits the generalization performance of diagnosis models. To address this issue, we propose Attribute-Aware Adversarial Domain Augmentation (AAADA), a zero-shot domain adaptation approach that produces synthetic data reflecting prior knowledge of a target domain. Given target attribute information, AAADA adversarially explores instances that capture target-specific characteristics based on feature correlations among source attributes. This approach enables models to tailor their decision boundaries to the target domain effectively. Experiments on large-scale health check-up datasets demonstrate that AAADA significantly enhances diagnostic accuracy compared to state-of-the-art domain generalization methods by effectively utilizing target attribute information.