Debiased medication recommendation through fusing frequent pattern and temporal medical records

Medication recommendation systems have gained significant attention in healthcare, aiming to deliver precise and personalized drug prescriptions tailored to patients' health conditions. However, existing approaches struggle with the skewed distribution of drug prescriptions, where models are biased toward common medications, compromising predictions for rare ones. To address these limitations, we propose a Debiased Medication Recommendation Network (DMRNet) to enhance both the fairness and accuracy of prescription formulation. Specifically, DMRNet introduces two key innovations: (1) a frequent pattern mining algorithm that integrates multiple drug combinations to mitigate imbalanced medication distribution by leveraging diverse frequency patterns to improve recommendation balance, and (2) a drug recalibration mechanism that enhances low-frequency drug recommendations by selectively fusing temporal prescription records since certain drugs are likely to appear in a patient's medication history. To further improve recommendation accuracy, we devise a cross-view drug prediction mechanism to capture patient-drug interactions from multiple perspectives. Experimental results on two benchmark datasets demonstrate that DMRNet achieves state-of-the-art performance, notably improving the PRAUC metric by 1.26% and 2.83%. The code of our model is available at https://github.com/Lxb0102/DMRNet.