Research On Methods Of Drug Recommendation Based On Mining Clinical Log

Scientific medication plays an important role in improving the efficacy of disease treatment.However,prescribing currently relies mainly on the personal expertise and experience of physicians.In fact,large amount of historical prescription logs have been stored in medical information systems.Furthermore,the drugs and treatment process of the same disease usually follow a certain pattern.Based on the prescription data mining and analysis of required drugs and treatment patterns,we can provide drug recommendation when doctors formulate medication plans.This has great significance.For this purpose,following scheme was proposed in this paper:(1)LD A clustering:Assuming that drug efficacy required by a certain disease is of different topics.The daily medications of the patient serve a number of efficacy topics according to a Multinomial distribution,and the drugs of each topic also follow a Multinomial distribution.Based on this assumption,LDA topic model is used to train the distributions from the prescription log of patients.(2)Treatment sequences:According to the similarity of topic distributions,clustering the drug efficacy combinations in each clinical day.Each resulted cluster is labeled with the same tag.Hence,the medication process of each patient is analogous to a sequence of efficacy combination tag.(3)PST(Probabilistic Suffix Trees)constructing:Assuming that the efficacy combination of daily medications follows a variable-length Markov model,PSTs are trained based on the efficacy combination label sequence.Then,medication is recommended for patients in treatment according to the PSTs.(4)Drug recommendation with sign data:Vector of body sign data is accessible.PST prediction distribution is combined with sign vector as features.Xgboost is used to construct the classifier,after which drugs can be recommended.The proposed methods are evaluated using the prescription data of patients with septicemia or diabetes recorded in the MIMIC-III clinical database.It turned out that method of using PST is fit for the disease with fix routine.For disease greatly relies on sign,the other achieved more satisfactory results.Both of the methods proposed in this paper performed better in Accuracy,Recall and F1-Score contrast with traditional recommendation methods.