Linghong Hong ¹ Shiwang Huang ² Xiaohai Cai ² Zhiming Lin ³ Yunting Shao ³ Longbiao CHEN ³ Min ZHAO ^4* Chenhui Yang ^2*

• ¹ Department of Drug Clinical Trial Institution, Xiang'an Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, China
• ² Fujian Key Laboratory of Sensing and Computing for Smart Cities, School of Informatics, Xiamen University, Xiamen, China
• ³ Fujian Key Laboratory of Sensing and Computing for Smart Cities, School of Information Science and Technology, Xiamen University, Xiamen, China
• ⁴ Big Data Center, The First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, China

According to the statistics of the World Health Organization (WHO), inappropriate medication has become an important factor affecting the safety of rational medication. In the gray area of medical insurance supervision, such as designated drugstores and designated medical institutions, there are lots of inappropriate medication phenomena about "big prescription for minor ailments". While the traditional Clinical Decision Support System (CDSS) are mostly based on established rules to regulate inappropriate prescriptions, which are not suitable for clinical environments that require intelligent review. In this paper, we model the complex relation among patient, disease and drug based on medical big data to promote the appropriate medication use. More specifically, we first construct the medication knowledge graph based on the historical prescription big data of tertiary hospitals and medical text data. Second, based on the medication knowledge graph, we employ Gaussian Mixture Model (GMM) to group patient population representation as physiological features. For diagnostic features, we employ pre-training word vector BERT to enhance the semantic representation between diagnoses. And in order to reduce adverse drug interactions caused by drug combination, we employ graph convolution network to transform drug interaction information into drug interaction features. Finally, we employ the sequence generation model to learn the complex relationship among patient, disease and drug, and provide an appropriate medication evaluation for prescribing by doctors in small hospitals from two aspects of drug list and medication course of treatment. In this study, we utilize the MIMIC III dataset alongside data from a tertiary hospital in Fujian Province to validate our model. The results show that our method is more effective than other baseline methods in the accuracy of medication regimen prediction of rational medication. In addition, it has achieved high accuracy in the appropriate medication detection of prescription in small hospitals.