DeviCare: Dosage Deviation-Aware and Regulated Medication Recommendation

Medication recommendation based on electronic health records (EHRs) is a critical research area in healthcare, aiming to support clinical decision-making by leveraging patients' historical visit data. Although prior studies have acknowledged the importance of fine-grained medication administration events, they primarily rely on physician-prescribed dosages while overlooking the discrepancy between prescribed and actual administered dosages, resulting in the loss of essential clinical semantics. To address this issue, we propose DeviCare, a novel framework that explicitly models dosage deviation as a fine-grained clinical signal and integrates it into a dosage deviation-regulated learning framework for medication recommendation. Specifically, we extract dosage deviations from multi-event temporal sequences and introduce an adaptive calibration mechanism to regulate representation shifts induced by dosage variations, thereby enhancing recommendation performance. Extensive experiments on the MIMIC (Medical Information Mart for Intensive Care)-III and MIMIC-IV datasets demonstrate that DeviCare consistently outperforms state-of-the-art methods across multiple evaluation metrics, validating the effectiveness of modeling dosagedeviation in medication recommendation tasks.