报告题目： Integrative Classification by Structural Equation Modeling of Homeostasis
报告摘要：We consider a new classification approach in the high-dimensional multiple source data setting, in which the number of features is much larger than the number of observations. The approach is motivated from the typical disease classification where the features in one group may have certain correlation structures that do not present in the other group, and this setting has become commonplace in biological and medical applications. Linear discriminant analysis (with a link function) is a classical method for this problem and more recently various machine learning algorithms are also being used, but the correlation structures have not been accounted in most methods. In this talk, we develop an integrative classification approach utilizing structural equations to model the homeostasis inherent in the control group and as characterized by phenotypes and biomarkers. The correlation structures and the parameters are estimated using only the data from the control group. The structural equation models no longer hold in the diseased group where the correlation often become less informative. The specificity and sensitivity are determined by choosing the confidence intervals of estimated parameters. The method is then used to analyze a real multi-platform genomics dataset and compared with the logistic regression analysis. The proposed method performs well.
报告人简介： 方宏彬, 乔治城大学生物、生物信息与生物数学系教授。主要研究兴趣为癌症相关的问题研究，包括临床试验数据的分析、诊断检验的评估、生物标志物和药物组合分析等。