Research On The Prediction And Detection Of Adverse Drug Events Based On Data Mining And Network Model

Adverse drug events(ADEs)which have been unresolved as major issues in the medical field,pose a serious threat to public health.ADEs have resulted in high morbidity,high mortality and huge medical costs.However,traditional toxicity testing and clinical trials are limited by such issues as the sample size and the type of data collected in the pre-market stages,and then the risk management is continued in the post-market stages.With the development of information technology,computational operation and simulation play a significant role in the prediction and detection of ADEs owing to increased efficiency and low cost,and continue to predict and detect the drug-ADE signals as data is continuously updated.In recent years,the prediction models have often exploited the network approach to represent and integrate data,which can give a clear explanation on how incorporating network proximity together with large-scale patient longitudinal data can facilitate the prediction of ADEs.The network-based system pharmacological approaches have attracted more attention from researchers,none of them however have investigated whether the frequency information and sample size of the drug-ADE associations play an important role in predicting true drug-ADE associations.On the other hand,disproportionality analysis(DPA)as data mining algorithm is developed to detect drug-ADE signals in the spontaneous reporting system(SRS).DPA is designed to identify highly significant drug-ADE associations,and can estimate the effect size of drug-ADE associations.However,DPA approaches need to be selected reasonably,and different approaches will bring diverse effects.In addition,these signals generated by different approaches can also complement each other,which are of great value to pharmacological research.Aimed at the above problems,considered from the computational strategy,this paper mainly explores and studies both ADEs prediction and detection,and aims to obtain superior predictive performance and effective detection methods to identify ADEs as early and accurately as possible.Consequently,this paper mainly focuses on the following aspects:(1)Research on the prediction of adverse drug events based on disproportionality analysis guided pharmacological network model(DPA-PNM).Pharmacological network model(PNM)utilizes the existing drug-ADE associations data to predict the future drug-ADE associations.PNM treats the observed drug-ADE pairs as true signals,and does not consider how frequently drug-ADE pairs are reported in the current data set,nor the effect size of drug-ADE associations.However,DPA itself not only estimates the effect size,but also ranks the drug-ADE signals.Therefore,this paper proposes an approach to predict drug-ADE signals based on DPA-PNM.Different DPAs have diverse effects,information component guided pharmacological network model(IC-PNM)is proposed to predict ADEs by analyzing the performance of different DPAs(PRR,ROR,IC and EBGM)in combination with PNM.IC-PNM not only takes advantage of network pharmacology features in predicting drug-ADE associations,but also combines with the frequency information and sample size of drug-ADE associations.It can predict ADEs for the new drugs effectively,thereby improving the predictive performance of system pharmacology approach based on network model.(2)Research on the prediction of ADEs via feature fusion-based predictive network models(FFPNMs).The performance of the methods for prediction of ADEs depends largely on features,and the ones that effectively reflect the essential attributes of the data are crucial for the prediction of ADEs.At the same time,machine learning classifiers that match the features also play a key role in predicting ADEs.Therefore,this paper studies in detail the link prediction based on the topological structure of complex network,and proposes the feature fusion-based predictive network models through the combination of network analysis approach and machine learning algorithms.The similarity measures based on topological structure of complex network are introduced into the definitions of network features.An algorithm is proposed through the improvement of similarity measures,and then the efficient features JADF(Jaccard and AA drug fusion,JADF)and JAAF(Jaccard and AA ADE fusion,JAAF)are defined.Finally we evaluate the performance of FFPNMs with three different machine learning classifiers.The experimental results show that FFPNMs have superior predictive performance.Among them,FFPNM with random forest has the best predictive result with the accuracy of0.945.There are no redundant features in FFPNMs,which reduces the data dimension and has good robustness.It can predict ADEs for the existent drugs effectively,thereby improving the accuracy and stability of ADEs prediction.(3)Research on the detection of ADEs via Bayesian signal detection algorithm based on the predictive network model.ADEs detection studies heavily rely on applying statistical or data mining methods to extract signals from the historical data in data sources such as SRS.Information component(IC)is employed as a metric to measure the disproportionality in Bayesian confidence propagation neural network(BCPNN)tool.IC assumes that parameters follow Beta distribution to estimate the prior probability,and the values of hyper-parameter are all supposed to be one.However,FFPNM can generate probabilities of drug-ADE associations with logistical regression model.Therefore,this paper combines FFPNM and IC to propose a Bayesian signal detection algorithm based on predictive network model(IC_FFPNM).The probabilities obtained by FFPNM are used as the prior probabilities of IC transformed by Bayes rule,and a logistic regression-based PS-adjusted approach is proposed to control confounding bias.Compared with the classic signal detection algorithms,IC_FFPNMhas better performance and can reduce errors effectively.Furthermore,IC_FFPNMcan also complement other signal detection algorithms.In comparison with a signal detection algorithm alone,combining different signal detection algorithms can obtain higher accuracy for detection research.