Signaling Adverse Drug Reactions Via Machine Learning Methods

The problem of drug safety is gaining more attention.Specifically,the research of signaling adverse drug reactions is a core topic in the fields of drug safety and drug discovery.The adverse drug reaction is the main reason to develop approximately 30%medical malpractice,which results in hundreds of billions of dollars in economic losses.Therefore,conducting the research of signaling adverse drug reactions getting attentions in the world.To mining and capture signals from adverse drug events,researchers have already proposed many data mining and statistics methods to signal the association between drug and ADRs.Referring data mining methods,they usually employ the pattern match or association rule methods to mining the signal of adverse drug reactions.This type of methods has two following shortcomings.(1)It is hard to signal rare associations between drugs and ADRs.(2)Since it is negative to take advantage of the confounders,which cause difficult adjusting the confounding drugs' effect when signaling the association between drugs and ADRs and cause a high false positive error.Since statistics methods usually employ the discriminative table methods based on the small size of sampling dataset to signal ADRs,which have the following two drawbacks.(1)It causes high bias error when signaling the association between drugs and ADRs due to the problem of sampling strategy.(2)It is hard to signaling the personalized adverse drug reactions referring different patients.In order to solve the above-mentioned problems,we work on studying to develop the machine learning algorithms to signal adverse drug reactions based on the massive clinical dataset.The first research point in this thesis is about signaling personalized adverse drug reactions.The definition of personalized adverse drug reactions is that since the same drug could induce variant adverse reactions due to patient's identity demographics,which requires the signaling method need to capture specific signals from different patients.We propose a feature-based similarity multitask learning model based on FAERS dataset.In addition to that,we propose to borrow the personalized recommendation methods from Recommendation study to solve drug safety problems.More specifically,we proposed a multitask learning method based on the FAERS data to generate personalized ADR ranking list by computing the association between patients and ADRs.Besides,we proposed a new measuring metric to evaluate our model comparing against the competing methods regarding the prediction accuracy about the signals of personalized ADRs.The experimental results show that our model performs better than competing methods on predicting adverse drug reactions per different patients.The second research aspect is about signaling adverse drug reactions with different frequency.Since different frequent adverse drug reactions cause different risk and induce different latent problems,especially the rare ADRs usually hard to be detected in the clinical trial stage,which need to signal the identity ADRs in terms of patients,situations.We propose a multi-kernel multitask learning method to signal various frequency ADRs based on the patient's distinguishing clinical conditions.In this study,we categorize the feature space with regard to the feature properties and employing the kernel pool for each type of feature.Our method can assign the best kernel for the associated feature space via multiple kernel learning methods as well as construct the corresponding constrained convex optimization conditions and the regularization functions.In addition,the model is with the power of capturing different frequency adverse drug reactions.The model also can signal different frequency ADRs induced by patients' new drug combinations by transferring the trained weight from history patients.To evaluate the performance,we propose a new evaluation measurement,overall-HitRate@n,which is based on HitRate@n.The experiments show that our method can achieve an advanced performance against competing methods.The third research point is signaling the association between one-way drug and ADR with high true positive or causality.The problems of small sampling and confounding factors usually cause high false positive error rate when signaling associations between drugs and ADRs.To solve this problem,we propose the multivariate linear regression method with Gamma-Poisson-Shrinkage Bayesian estimator based on 4M in size patient records.This method assumes the data following the Gamma-Poisson conjugate prior distribution,and propose employing the multivariate linear regression model to adjust confounders.Our contributions can be concluded in the flowing three aspects.(1)We propose a multitask learning method to solve the personalized ranking ADRs problems.(2)We propose a multi-kernel multitask learning model to capturing ADRs in different frequency.(3)We propose a multivariate linear regression method using Gamma-Poisson-Shrinkage Bayesian estimator to adjust confounders to obtain true positive association between drug and ADRs.In conclusion,we propose a new perspective way in studying the problem of signaling adverse drug reactions.