Thesis (M.A.)--West Virginia University, 2002. / Title from document title page. Document formatted into pages; contains ix, 61 p. : col. ill., col. maps. Includes abstract. Includes bibliographical references (p. 40-42).

Thesis (Ph. D.)--University of Texas at Austin, 2001. / Vita. Includes bibliographical references. Available also from UMI/Dissertation Abstracts International.

Thesis (M.S.)--Mississippi State University. Department of Electrical and Computer Engineering. / Title from title screen. Includes bibliographical references.

Thesis (M.S.)--University of Akron, Dept. of Biomedical Engineering, 2008. / "December, 2008." Title from electronic thesis title page (viewed 12/12/2009) Advisor, Dale H. Mugler; Committee members, Daniel B. Sheffer, Bruce C. Taylor; Department Chair, Daniel B. Sheffer; Dean of the College, George K. Haritos; Dean of the Graduate School, George R. Newkome. Includes bibliographical references.

This dissertation develops statistical and data mining based methods for insurance fraud detection. Insurance fraud is very costly and has become a world concern in recent years. Great efforts have been made to develop models to identify potentially fraudulent claims for special investigations. In a broader context, insurance fraud detection is a classification task. Both supervised learning methods (where a dependent variable is available for training the model) and unsupervised learning methods (where no prior information of dependent variable is available for use) can be potentially employed to solve this problem. First, an unsupervised method is developed to improve detection effectiveness. Unsupervised methods are especially pertinent to insurance fraud detection since the nature of insurance claims (i.e., fraud or not) is very costly to obtain, if it can be identified at all. In addition, available unsupervised methods are limited and some of them are computationally intensive and the comprehension of the results may be ambiguous. An empirical demonstration of the proposed method is conducted on a widely used large dataset where labels are known for the dependent variable. The proposed unsupervised method is also empirically evaluated against prevalent supervised methods as a form of external validation. This method can be used in other applications as well. Second, another set of learning methods is then developed based on the proposed unsupervised method to further improve performance. These methods are developed in the context of a special class of data mining methods, active learning. The performance of these methods is also empirically evaluated using insurance fraud datasets. Finally, a method is proposed to estimate the fraud rate (i.e., the percentage of fraudulent claims in the entire claims set). Since the true nature of insurance claims (and any level of fraud) is unknown in most cases, there has not been any consensus on the estimated fraud rate. The proposed estimation method is designed based on the proposed unsupervised method. Implemented using insurance fraud datasets with the known nature of claims (i.e., fraud or not), this estimation method yields accurate estimates which are superior to those generated by a benchmark naïve estimation method. / text

Many important scientific questions can not be studied properly using a single measurement as a response. For example, many phenotypes of interest in recent clinical research may be difficult to characterize due to their inherent complexity. It may be difficult to determine the presence or absence of disease based on a single measurement, or even a few measurements, or the phenotype may only be defined based on a series of symptoms. Similarly, a set of related phenotypes or measurements may be studied together in order to detect a shared etiology. In this work, we propose methods for studying complex phenotypes of these types, where the phenotype may be characterized either longitudinally or by a diverse set of continuous, discrete, or not fully observed components. In chapter 1, we seek to identify predictors that are related to multiple components of diverse outcomes. We take up specifically the question of identifying a multiple regulator, where we seek a genetic marker that is associated with multiple biomarkers for autoimmune disease. To do this, we propose sparse multiple regulation testing (SMRT) both to estimate the relationship between a set of predictors and diverse outcomes and to provide a testing framework in which to identify which predictors are associated with multiple elements of the outcomes, while controlling error rates. In chapter 2, we seek to identify risk profiles or risk scores for diverse outcomes, where a risk profile is a linear combination of predictors. The risk profiles will be chosen to be highly correlated to latent traits underlying the outcomes. To do this, we propose semiparametric canonical correlation analysis (sCCA), an updated version of the classical canonical correlation analysis. In chapter 3, the scientific question of interest pertains directly to the progression of disease over time. We provide a testing framework in which to detect the association between a set of genetic markers and the progression of disease in the context of a GWAS. To test for this association while allowing for highly nonlinear longitudinal progression of disease, we propose functional principal variance component (FPVC) testing.

Not available / text

This thesis presents an evaluation of algorithms for classification of functional MRI data. We evaluated the performance of probabilistic classifiers that use a Gaussian model against a popular non-probabilistic classifier (support vector machine, SVM). A pool of classifiers consisting of linear and quadratic discriminants, linear and non-linear Gaussian Naive Bayes (GNB) classifiers, and linear SVM, was evaluated on several sets of real and simulated fMRI data. Performance was measured using two complimentary metrics: accuracy of classification of fMRI volumes within a subject, and reproducibility of within-subject spatial maps; both metrics were computed using split-half resampling. Regularization parameters of multivariate methods were tuned to optimize the out-of-sample classification and/or within-subject map reproducibility. SVM showed no advantage in classification accuracy over Gaussian classifiers. Performance of SVM was matched by linear discriminant, and at times outperformed by quadratic discriminant or nonlinear GNB. Among all tested methods, linear and quadratic discriminants regularized with principal components analysis (PCA) produced spatial maps with highest within-subject reproducibility. We also demonstrated that the number of principal components that optimizes the performance of linear / quadratic discriminants is sensitive to the mean magnitude, variability and connectivity of simulated active signal. In real fMRI data, this number is correlated with behavioural measures of post-stroke recovery , and, in a separate study, with behavioural measures of self-control. Using the data from a study of cognitive aspects of aging, we accurately predicted the age group of the subject from within-subject spatial maps created by our pool of classifiers. We examined the cortical areas that showed difference in recruitment in young versus older subjects; this difference was demonstrated to be primarily driven by more prominent recruitment of task-positive network in older subjects. We conclude that linear and quadratic discriminants with PCA regularization are well-suited for fMRI data classification, particularly for within-subject analysis.

This thesis presents an evaluation of algorithms for classification of functional MRI data. We evaluated the performance of probabilistic classifiers that use a Gaussian model against a popular non-probabilistic classifier (support vector machine, SVM). A pool of classifiers consisting of linear and quadratic discriminants, linear and non-linear Gaussian Naive Bayes (GNB) classifiers, and linear SVM, was evaluated on several sets of real and simulated fMRI data. Performance was measured using two complimentary metrics: accuracy of classification of fMRI volumes within a subject, and reproducibility of within-subject spatial maps; both metrics were computed using split-half resampling. Regularization parameters of multivariate methods were tuned to optimize the out-of-sample classification and/or within-subject map reproducibility. SVM showed no advantage in classification accuracy over Gaussian classifiers. Performance of SVM was matched by linear discriminant, and at times outperformed by quadratic discriminant or nonlinear GNB. Among all tested methods, linear and quadratic discriminants regularized with principal components analysis (PCA) produced spatial maps with highest within-subject reproducibility. We also demonstrated that the number of principal components that optimizes the performance of linear / quadratic discriminants is sensitive to the mean magnitude, variability and connectivity of simulated active signal. In real fMRI data, this number is correlated with behavioural measures of post-stroke recovery , and, in a separate study, with behavioural measures of self-control. Using the data from a study of cognitive aspects of aging, we accurately predicted the age group of the subject from within-subject spatial maps created by our pool of classifiers. We examined the cortical areas that showed difference in recruitment in young versus older subjects; this difference was demonstrated to be primarily driven by more prominent recruitment of task-positive network in older subjects. We conclude that linear and quadratic discriminants with PCA regularization are well-suited for fMRI data classification, particularly for within-subject analysis.

In this study, we compare the two statistical techniques logistic regression and discriminant analysis to see how well they classify companies based on clusters – made from the solvency ratio ­– using principal components as independent variables. The principal components are made with different financial ratios. We use cluster analysis to find groups with low, medium and high solvency ratio of 1200 different companies found on the NASDAQ stock market and use this as an apriori definition of risk. The results shows that the logistic regression outperforms the discriminant analysis in classifying all of the groups except for the middle one. We conclude that this is in line with previous studies.