Circulatory shock is a critical condition in Intensive Care Units (ICU), where timely intervention is vital for survival. While deep learning models have achieved high predictive accuracy, their “black box” nature often hinders medical adoption due to a lack of interpretability. Conversely, standard Retrieval-Augmented Generation (RAG) frameworks struggle to capture the temporal dynamics essential for identifying patient deterioration. To bridge this gap, we apply Time-aware Dynamic-window RAG (TD-RAG), a framework proposed in our previous work and proven effective in environmental monitoring, to the medical domain. By transforming physiological time-series into semantic temporal windows, TD-RAG retrieves historical trajectories of decompensation to ground its predictions. Evaluation on the MIMIC-IV dataset demonstrates that TD-RAG acts as a high-sensitivity safety net, achieving a recall of 82.70%, significantly outperforming both Vanilla-RAG ( \(p < 0.01\) ) and traditional deep learning baselines ( \(p < 0.05\) ). Our approach balances performance with transparency, offering doctors both accurate risk alerts and context-aware rationales.