Bayesian data mining techniques in public health and biomedical applications.

摘要

This thesis is focused on Bayesian data mining techniques for non-traditional data in the health care and biomedical fields, and comprises three main topics.;Meta-analysis of clinical trials with incomplete data set.;This methodology is motivated by meta-analysis that compares the effectiveness of glass-ionomer (GI) to resin-based (RB) sealants in preventing caries in permanent molars of children. While GI and RB sealants are two of the most commonly used dental sealant materials, there is strong debate regarding their comparative effectiveness. Although GI to RB sealant meta-analyses have been performed in the past, this research is the first to find difference by including studies that were excluded previously solely because they failed to report data in a paired manner. The conclusion is based on 12 clinical studies performed on 2,012 children, that had incomplete information and inconsistent conclusions.;To retrieve full paired table information from studies that only reported marginal totals, we propose a Bayesian approach based on Metropolis random walk and a hierarchical Bayesian model. The procedure relies on Markov chain Monte Carlo simulation, recovers a full table and reports its accuracy in terms of credible sets for the cell counts. The performance of the model is remarkable; from 17 matched-pair study examples, four of the estimated tables coincided exactly with the original tables and nine tables deviated only (+1,-1) in terms of cell counts.;Analysis of spatial repeated measures with excess zero and no covariate information.;Water fluoridation, the addition of fluoride to drinking water to prevent tooth decay, is listed as one of ten greatest public health achievements of the 20th century by the Centers for Disease Control and Prevention. However, fluoride is present naturally in water in concentrations well above recommended levels, which may have several long-term adverse effects. Therefore, describing the availability of natural fluoride by county can inform public health efforts in prevention of tooth decay and dental fluorosis.;This work is to estimate United States county level natural fluoride occurrence in the ground water. To accommodate the limitations of the data, we propose a two-step hierarchical Bayesian approach. It models spatial repeated measures with excess zeroes and it is robust even when covariate information is not available. The approach starts with: (i) an autologistic model that estimates the probability of zero observations, and proceeds to (ii) an excess-zero model that describes the underlying physical phenomena of the zero-inflated data. The proposed hierarchical Bayesian structure produces improved discretized estimates and reflects the continuous behavior of the natural fluoride occurrence better. By employing the hyper prior structure, state level estimates can be obtained as well. Furthermore, the model can answer pragmatic questions such as the probability that a specific county has naturally occurring fluoride above EPA recommended level".;Image classification based on the overall image regularity in wavelet domain.;Breast cancer is one of the most common forms of cancer among women in the United States; an estimated 1 in 8 women born today will be diagnosed during her lifetime. Since the causes of breast cancer are not yet fully understood, early detection is still the best strategy for improving prognoses. Mammography is currently the most effective method for detecting breast cancer early; however, radiological interpretation of mammogram images is a challenging task. The appearance of even normal tissue is highly variable and complex, and signs of early disease are often indistinct.;We propose a diagnostic based on the properties of overall image backgrounds; this procedure currently is an unused diagnostic modality in mammograms. Generally speaking, normal/healthy images tend to be more irregular than cancer images. The overall regularity of the image is assessed through wavelet analysis, which is then summarized by a few measures. These estimators are evaluated based on their ability to classify digitized mammogram images from a clinical database, for which the true disease status is known by biopsy. This research presents a two-fold approach: (i) generalization of the covariance wavelet spectra to the complex domain and (ii) the estimation of Hurst parameter and phase information as discriminatory descriptors. The most accurate classification rates from this work achieve 86%; these rates vary slightly with the choice of wavelet basis, levels used and size of training set.