Font Size: a A A

Study On A Unified Statistical Software For Heterogeneity In Treatment Effect

Posted on:2019-07-09Degree:MasterType:Thesis
Country:ChinaCandidate:M Y HaoFull Text:PDF
GTID:2394330566496859Subject:Computer technology
Abstract/Summary:PDF Full Text Request
Randomized clinical trials are systematic studies of medications and other treatments.The purpose of clinical trials is to provide the information of the beneficiaries,and even the sufferer on the efficacy of the medication.A subject that exhibits the same type of therapeutic effect for a medication is called a subgroup.The process of finding subgroups that exhibit heterogeneity with respect to medication efficacy from clinical trial data is called subgroup analysis/identification.With the help of pattern recognition and machine learning techniques,many researchers have achieved results with different advantages in the research of subgroup identification methods.However,the current subgroup identification methods deeply depend on correlations between characte ristics and the outcome of subject,and cannot adapt to the needs of scalable multi-class subgroup identification.In view of this situation,this paper proposes a Markov Chain Monte Carlo(MCMC)based subgroup identification method.In our study,first we review the source of subgroup analysis problems,the development of subgroup identification methods,and introduce some related concepts such as clinical trials and subgroup heterogeneity besides computer science;second,according to the characteristics o f clinical trial data,a systematic,reproducible,and statistically controllable clinical trial data simulation methods have been designed and followed by a detailed description of MCMC based subgroup identification methods;third,a set of data with different statistical characteristics are obtained from the data simulation experiment and then the statistical characteristics of the data are evaluated,and the efficacy of the proposed method is evaluated on the data set.In the end,a browser-based subgroup identification platform developed with the R-language extended package Shiny was successfully online.There are two main innovations in our study:(1)Clinical trial data simulation.If the number of subjects of the input data for subgroup identification method is too small,the process of subgroups generation may be stopped prematurely due to the subjects of a subgroup is too few or the differences between the subjects are too small,for this kind of situation may lead to a higher type ? error rate.Moreover,since clinical trial data does not have true subgroup markers,it is difficult to accurately assess the performance of different subgroup identification methods.Based on the multivariate Gaussian distribution,we designed a method for generating simulated clinical trial data.Using this method,a group of simulated clinical trial data with controllable number of subjects,controllable subgroup number,controllable subgroup properties,controllable outcome differences between intervention groups and control groups and true marker markers(mark_ori)was gathered.The validity of the subgroup identification method can be evaluated by measuring the consistency between mark_ori and the prediction results of the proposed subgroup identification method.(2)The design of MCMC based subgroup identification method.The MCMC-based subgroup identification method proposed in our study assigns a group of subgroups as a systematic state,which is transferred between different systematic states by the equal probability state transfer matrix,uses the subgroup state assessment function to control the direction of the transfer,and convergence to the stationary distribution,the global optimal subgroup state.While maintaining the average identification efficacy consistent with the classical method: 1.Compared with the subgroup identification based on tree structure which is dependent on the nature of greedy selection,our method has global optimization ability;2.Due to the final steady state of the Markov chain Existence,this method will automatically avoid overfitting;3.If the method proposed in our paper is used for dealing with multi-subgroup types,only the subgroup state evaluation function needs to be modified,and the application scenario can be easily extended.
Keywords/Search Tags:subgroup identification, multivariate Gaussian distribution, Markov Chain Monte Carlo
PDF Full Text Request
Related items