Research On Medical Data Processing Based On Machine Learning Methodologies

The development of medical science heavily depends on the support fromrelated disciplines. With the rapid advance of information science, peoplenowadays are focusing on how to utilize the information technology to servethe medial research. Application of the information technology to the medicaldata analysis will play an important role in disease prevention, diseasediagnosis, disease therapy and other health care areas. At present, machinelearning has become one of the scientific frontiers in computer science, andhas been successfully applied to a wide variety of research fields. My researchis to combine the machine learning theories, methodologies and technologies,and apply it to the medical data analysis. The thesis can be divided into fourparts.1. The mixture random effect model (MR model) was proposed to improve theperformance of Meta analysis. MR model has the ability to capture the arbitrarycomplex distribution of the underlying effect sizes, and can be viewed as ageneralized model of the random effect model. An efficient algorithm for parameterestimation of MR model is the MCMC approach. MR model was used to evaluate theefficacy of nicotine replacement therapies. The complex distribution of theunderlying effect sizes was revealed, and the factors which have great influence onthe efficacy were found.2.The problem of integrating the Mean Values and Standard Deviation (MS) ofmultiple variables from scattered medical studies was proposed as MS informationlearning problem. Under the framework of EM, MSEM algorithm for estimating theprobability density function using MS information was designed. The result obtainedfrom MSEM can be used to analyze the distributional characteristics of variablesunder different conditions, to describe the statistical dependency between multiplevariables, and to predict MS information of the un-observed variables. In theexperiment, MSEM was used to integrate literature MS information into quantitativedescriptions of the functions of HPA hormones and cytokines under health or RAconditions. Some meaningful patterns of the defect of HPA-cytokine interactions inthe RA patients were obtained.3.Isomap, a newly proposed nonlinear dimensionality reduction method, wasintroduced to facilitate the analysis of high dimensional medical data. Based onIsomap, a novel supervised nonlinear dimensionality reduction method wasdeveloped, named as SIsomap. SIsomap1 and SIsomap2, two versions of SIsomap,can be used to reduce the dimensionality of data with continuous and two-class labelsupervised attributes respectively. A scheme of classifier design based on SIsomap2and RBF neural network was proposed. The successful application to several highdimensional medical datasets, including the lung tumor gene expression dataset,diabetes pathology dataset, etc, shows the power of the method.4.From the aspect of machine learning, the relationship between the patternrecognition problem and the integrative evaluation of repeated measurements withtwo controls was investigated. Based on such relationship, a new algorithm wasdeveloped. Our results on two pharmacological datasets showed that the proposedalgorithm was efficient for the analysis of repeated measurements with multiplegroups, multiple biomarkers and small sample size.