AOI-Attention: Interpretable Region Weighting for Eye-Tracking Depression Assessment

Eye-movement heatmaps offer a practical, non-contact signal for depression assessment. However, existing models rarely clarify which image regions drive predictions. In this study, we propose a region-aware framework that integrates an Area-of-Interest (AOI) mask (partitioning each stimulus into background, body, face, and eyes) with a sample-adaptive AOI attention layer that pools Vision Transformer (ViT) tokens by region and learns explicit per-AOI weights. When evaluated on a clinical cohort with a standardized viewing protocol, the proposed framework improves upon the ViT baseline and yields clear per-region attributions: the classifier relies predominantly on background distributions, while the learned face/eye weights increase monotonically from major depression to healthy controls—patterns not evident in raw gaze proportions. In addition, combining the model’s predictions with statistical eye-movement measurements leads to consistent performance improvements. The proposed AOI attention enhances interpretability without compromising accuracy, offering a principled framework for explainable depression severity classification.