COGNET-AMO: Personalized Prescription Recommendation using Attention over Medical Order Information

Abstract Prescription recommendation is widely discussed in smart healthcare research and applications. Current recommendation systems focus on mining structured data such as lists of diagnostic and surgical codes in patient records, without considering combining textual descriptions from Medical Orders. However, medical orders typically provide a more concise and important description of the patient, which reflects the doctor’s prescribing in tentions. Meanwhile, such information is also useful for building patient profiles and making recommendations based on profile similarities. In this paper, we propose a novel Conditional Generation Net with Attention over Medical Orders for Personalized Prescription Recom mendation (COGNET-AMO). COGNET-AMO employs an attention mechanism to model the patient’s health status with a focus on chief complaints and other important information in medical orders. Then, it constructs a patient case library that clusters patients based on their similarities over medical order information and Electronic Health Records. Finally, it combines them to generate more personalized and accurate prescription recommendations. COGNET-AMO is evaluated on the public MIMIC-III dataset, and the experimental results indicate that it can achieve 7%, 5%, 4% and 8% relative improvements than state-of-the-art approaches in Jaccard, PRAUC, F1, and average precision