<p>Precision medication recommendations, enabled by deep analysis of electronic health records (EHRs), serve as a cornerstone of smart healthcare. They aim to maximize therapeutic efficacy while systematically mitigating adverse drug reactions. However, significant heterogeneity in medical records hinders effective patient representation, limiting the accuracy and clinical utility of existing recommendation models. To address these challenges, we propose CHGRec, a novel multilevel patient representation-enhanced medication recommendation framework. CHGRec constructs foundational patient representations through similarity learning and enhances their expressiveness via a causally inferred heterogeneous graph derived from EHRs. This graph quantitatively models medication effects on specific diseases based on patient histories, revealing latent relationships among medical entities. Furthermore, CHGRec incorporates a personalized medication refinement module that dynamically adjusts recommendation probabilities using edge weights from the causal heterogeneous graph. This dual enhancement of patient-specific representations and individual variability modeling ultimately delivers more accurate and personalized medication recommendations. Evaluations on MIMIC-III/IV datasets demonstrate CHGRec’s superior performance in recommendation accuracy and safety, underscoring its potential for real-world clinical deployment.</p>