MeSIN: Multilevel selective and interactive network for medication recommendation

Recommending medications for patients using electronic health records (EHRs) is a crucial data mining task for an intelligent healthcare system. EHRs data, which comprise multiple temporal sequences (e.g., lab, diagnosis, and treatment sequences), encompass multilevel structure information, including flat structure information (e.g., sequential medical codes and their relationships) in each sequence and hierarchical structure information between sequences (e.g., information flow from lab sequence to diagnosis sequence or hierarchical codes relationships), which implicitly reflects a physician’s decision process. However, existing studies focus on mining flat structure information, which is incomplete and hence biased. Therefore, to realize more accurate medication recommendations, this paper proposes a multilevel selective and interactive network (MeSIN) that exploits complete structure information of EHR data; in particular, at the core of our framework is an interactive long short-term memory network (InLSTM), which is developed to model the temporal dependencies in each flat-structured sequence via a recurrent mechanism and capture the hierarchical relationships of different sequences via a reinforced interactive network. Moreover, to reduce the impact of irrelevant information on the multilevel interaction process, an attentional selective module (ASM) is introduced to filter out noisy information and thereby focus on important sequence information at each timestamp. Finally, a global selective fusion module (GSFM) is proposed to infuse multisourced information embeddings into multilevel structural relationship-aware patient representation. Experimental results on a practical clinical dataset show that our framework consistently outperformed all baseline models in the medication recommendation task.