The Joint Diagnosis Of Multiple Biomarkers

In medical diagnosis,biomarkers are often used as a diagnostic tool to identify patients with certain diseases.Since the diagnosis based on a single biomarker may not be highly accurate,in order to improve the diagnostic accuracy,it is usually necessary to combine multiple biomarkers for comprehensive diagnosis.Under the evaluation criterion of the area under the ROC curve(AUC)or Youden index,there are some combined diagnostic methods of biomarkers in the literature,such as the parametric method based on normal hypothesis,min-max,stepwise and pairwise without distribution hypothesis.Literature shows that parametric method performs poorly in non-normal situations,and min-max method has advantages of fast calculation and robustness compared with stepwise and pairwise method.But in some cases,the diagnostic accuracy is often much lower than the other three methods,such as symmetric distribution.It is noted that min-max method only uses the maximum and minimum values of multiple biomarkers,and lacks the average level information of individual indicators,and the data normalization method adopted is greatly influenced by the sample size.In this paper,an improved method of min-max method is proposed: a new method of data normalization(control standardization)is defined,which combines the maximum,minimum and truncated mean of standardized data of multiple biomarkers and is recorded as minmax-trim method.The diagnostic performance of the new method is simulated and compared under AUC and Youden index criteria respectively.The structure of the paper is as follows: Chapter 2 mainly introduces the specific algorithms of minmax-trim combination method and other four methods under AUC criterion,and simulates these five combination methods.The simulation results show that minmax-trim method improves the AUC of min-max method,especially in the normal case.Compared with stepwise and pairwise method,minmax-trim method has shorter running time,and in most nonnormal cases,the AUC is larger.In the normal case,the AUC of minmax-trim method can be close to the parameter method based on normal assumption and stepwise method.Chapter 3mainly considers the algorithm and performance of the five combinatorial methods in Chapter 2 under Youden index criterion,and makes comparative analysis through simulation.The simulation results show that the stepwise method works best in the case of multivariate normal distribution,but the Youden index obtained by minmax-trim method is close to it and takes less time.In other cases,minmax-trim method performs well.Chapter 4 applies the new method to APS data,growth-related hormone data and fetal growth data.