This paper addresses the challenge of modeling transitions in user activities within subforums of online health communities, like an online breast cancer forum, to infer changes in users' health stages. It proposes a novel dynamic graph-to-sequence neural networks architecture (DynGraph2Seq), which uses a dynamic graph encoder and an interpretable sequence decoder, along with dynamic graph hierarchical attention mechanisms. The model effectively maps time-evolving user activity graphs to sequences of health stages, providing both accuracy and interpretability in health stage prediction.