Multi-classifier Fusion Model And Its Application In Clinical Diagnosis And Treatment Decision-making

In recent years,the multi-classifier fusion model(Multi-Classifier Fusion Model,MCFM)has been widely used in clinical diagnosis and treatment decision-making tasks,such as diagnosis and prognosis prediction.Effective MCFM can be used for mining the information in medical image data and the internal relationship between clinical treatment parameters and diseases,providing better diagnosis information for clinical diagnosis and treatment decision-making and better helping clinicians make accurate diagnosis and develop individualized treatment plans for patients.MCFM obtains better classification performance than a single classifier by fusing different base classifiers.In MCFM,the prediction of the base classifier is required to be better than the random prediction,and the base classifier members are required to be complementary to each other.However,in reality,since most of MCFM is trained based on the same distribution of data,there is a high correlation between the adjoint base classifiers,and it is difficult to ensure the accuracy and diversity requirements at the same time.Here,MCFM and its application in clinical diagnosis and treatment decision-making were studied,and two approaches for creating MCFM is proposed.The main contents are as follows:First,a hierarchical fusion framework to integrate homogeneous and heterogeneous classifiers.Homogeneous ensemble is composed of a series of base classifiers based on the same mathematical theory,the emphasis is to improve the diversity of base classifiers.While heterogeneous ensemble is combined by different classifier algorithms,and how to make the final decision is the main research problem of this type of ensmeble method.We propose a hierarchical fusion model based on random projection.Random projection is performed on the original data set to increase the diversity of the data and the projected data is input into a two-layer multi-classifier fusion framework to exploit the characteristics of the two ensemble methods and improve the diversity and generalization ability of ensemble.The framework was comprehensively verified with 15 UCI data sets and 3 clinical data sets and the experimental results show that the framework is superior to the benchmark comparison method,verifying that it has practical clinical significance in assisting clinical diagnosis and treatment decision-making.Second,multi-classifier fusion model based on optimal random projection.Creating a multi-classifier system with good generalization performance can be regarded as an optimization problem,such as the selection problem of features,classifier parameters,classifier selection and so on.As a classic optimization method,genetic algorithm has been widely used in the research of MCFM.Based on the study of MCFM based on random projection,the system introduces genetic algorithm in the process of generating random projection matrix,and realizes the application of genetic algorithm to evolve the random projection matrix,so that the built ensmeble system can better predict a given classification task and overcome the limitations of machine learning algorithms depending on learning tasks,finally increase the diversity and robustness of the ensemble system.For example,for specific clinical decision-making tasks,such an auxiliary tool that is not affected by specific decision-making tasks has clinical practical significance.In this article,we conducted a preliminary verification of the effectiveness of the algorithm on 10 sets of UCI data.In this article,we innovatively proposed two multi-classifier fusion model based on random projection.In Chapter 3,we discussed from the perspective of ensemble framework and proposed a hierarchical fusion framework based on random projection.We verified its feasibility and effectiveness on 15 UCI datasets and three datasets of clinical application.In Chapter 4,we propose a multi-classifier fusion ensemble based on optimal random projection from the perspective of optimization,and conduct preliminary experimental verification on 10 sets of UCI data.The experimental results show that the algorithm is effective.