Explainability and Fairness in VLMs for Wellbeing Assessment

A new study from arXiv (2604.23786) investigates fairness and explainability in Vision-Language Models (VLMs) for wellbeing assessment and depression prediction. Researchers evaluated models across laboratory (AFAR-BSFT) and naturalistic (E-DAIC) datasets, finding significant performance disparities: Phi3.5-Vision achieved 80.4% accuracy on E-DAIC, while Qwen2-VL scored only 33.9%. Both models showed a tendency to over-predict depression on AFAR-BSFT, raising concerns about diagnostic reliability and demographic fairness. The work highlights the under-explored intersection of Explainable AI (XAI) and multimodal foundation models in clinical mental health monitoring.