Towards robust drug recommendation based on dual perspective encoder and iterative denoising mechanism

Drug recommendation aims to provide optimal therapeutic drug combinations based on a patient’s health condition. Current drug recommendation methods primarily center on mining a patient’s longitudinal visit information in electronic health records. However, they often ignore sufficient interaction between diseases and drugs and struggle to capture the patient’s essential medical history because not all information is pertinent to the current visit. Consequently, these models typically exhibit suboptimal recommendation performance and robustness in complex medical scenarios. In this paper, we design a novel approach called DPID, which proposes a dual perspective encoder and an iterative denoising mechanism for robust drug recommendation. Specifically, we utilize a dual perspective drug representation module to capture disease–drug interaction from different perspectives and develop a novel iterative denoising module to explicitly filter out noisy history information for the current visit. Furthermore, we employ graph representation learning to acquire comprehensive drug representations, integrating diverse forms of drug knowledge to enhance recommendations and bolster model robustness. Experimental results on two publicly available datasets consistently demonstrate that our model achieves state-of-the-art performance, particularly achieving improvements of 1.10%, 1.02%, 0.98% on MIMIC-III in Jaccard, PRAUC and F1-score, respectively.