<p>Medication recommendation aims to provide personalized and safe regimens by leveraging patients’ clinical data. However, existing methods heavily rely on external knowledge (e.g., drug interaction graphs, molecular structures), leading to challenges in annotation cost and data reliability. To address these issues, we propose DCSD-MR, a single-source data-driven model that captures intrinsic patterns between current and historical clinical visits. DCSD-MR introduces two key innovations: (1) a clinical presentation-driven historical matching mechanism that identifies relevant past visits based on the current disease status, with separate modeling of clinical features and medication characteristics to enhance patient-specific representations, and (2) a dynamic constraint strategy that combines a classification-calibrated loss and a drug count normalization method to explicitly regulate the number of recommended medications, which helps control prescription complexity and indirectly reduce potential drug–drug interactions risks, thereby improving practical applicability. Moreover, the proposed framework involves large-scale historical visit matching and quantity-aware optimization over longitudinal patient records, which naturally requires parallel and high-performance computing support to enable efficient training and real-time inference in large clinical systems. Extensive experiments on the MIMIC-III dataset demonstrate that DCSD-MR outperforms existing baseline methods across multiple metrics, achieving a Jaccard score of 53.37%, PRAUC of 77.39%, and F1 score of 68.69%. These results underscore its effectiveness and strong potential for real-world clinical deployment.</p>