Awaken the Giant: Activating LLMs via Deep Model Guidance for Boundary-aware Medication Recommendation

Accurate and safe medication recommendations from Electronic Health Records (EHRs) are essential for clinical decision support. While Large Language Models (LLMs) have shown strong semantic reasoning capabilities in healthcare, they tend to make coarse binary predictions, overlooking medications near the decision boundary and leading to overprescription. In contrast, deep models offer fine-grained probability outputs but lack contextual reasoning needed for complex boundary cases. To address this, we propose a boundary-aware medication recommendation framework (GiantMed) that activates the potential of LLM "giant" under deep model guidance. Specifically, GiantMed leverages a deep model to identify boundary medications and directs the LLM to focus on these clinically ambiguous yet informative cases. Furthermore, we augment contextual knowledge of boundary medications by retrieving relevant historical EHRs and integrating Drug-Drug Interaction (DDI) constraints. The final recommendation is generated by incorporating the LLM-refined boundary medications with confident predictions from the deep model. Extensive experiments on two real-world EHR datasets demonstrate that GiantMed 1 achieves state-of-the-art accuracy while reducing DDI rates.