Multilevel Asynchronous Time Network for Medication Recommendation

Medication recommendation is a pivotal task for AI in the realm of healthcare. Previous works have primarily focused on recommending medication for intricate medical conditions solely based on patients’ electronic health records (EHR).Although extensive progress has been made, current research still faces the following limitations: ignoring data integrity in EHR, disregarding the impact of single-visit data on recommendation results; not adequate in learning about multilevel dependencies; and ignoring the effect of asynchronous relationships in irregular time intervals on recommendation results. To solve the above limitations, the Multilevel Asynchronous Time Network for Medication Recommendation (MLATNet) model is proposed. MLATNet first enhances the embedding of the EHR data using graph attention networks. Secondly, we use a transformer global fusion module to learn patients’ long-distance global information, and a multikernel CNN module to learn local time dependencies information, thus obtaining global-local multilevel dependency information. Then, we design an asynchronous time module to fuse irregular time series. Moreover, through a DDI loss, MLATNet effectively controlling DDI rate to achieve medication recommendation. MLATNet outperforms all baseline methods, achieving a Jaccard score of 52.96%, an F1-score of 68.93%, and a PR-AUC of 77.56% on the MIMIC-III test set.