Improving Lyme disease diagnosis through bioinformatics

点击翻译按钮获取中文摘要

Due to limited physician knowledge of diagnostic criteria for Lyme disease (LD)and excessive utilization of laboratory tests, over-diagnosis and over-treatment of LD has become a significant public health problem. This grant is the first phase of a two-part process that will develop a bioinformatic framework to maximize the predictive power of new and existing laboratorytests. The current two-step serologic approach detects antibody to Borrelia burgdorferi by whole-cell EIA, followed by Western blot confirmation of positive or equivocal EIA results; while this approach is highly specific, it lacks sensitivity for early LD. We will evaluate a new, patented algorithm based on Bayes' theorem that combines the pretest risk of LD with multi-antibody serology to generate a posterior probability for each patient. This new algorithm has demonstrated sensitivity superior to the two-step method in a pilot study. To provide comparators to the algorithm, multivariate models will be developed concurrently using partial ROC regression and logistic-linear regression, assisted by penalized likelihood functions. Data from 280 LD patients and 559 controls, already tested using the two- step method and VIsE, C6, and pepdO ElAs, will be used to develop these models. The new algorithm will utilize those variables selected by partial ROC regression that contribute significantly to the predictive model. The pretest risk of LD will be estimated for each patient and control based on available clinical data. A partial ROC curve (at least 80% specific) will be generated for each predictive method and the two-step approach by varying their respective posterior probability cutoffs. The diagnostic power of each method will be determined by the area under its partial ROC curve (AUC), and compared to that of the two-step approach using a bootstrap technique. The primary study end-point is to identify at least one new predictive method with performance equivalent or superior to the two-step approach, thereby justifying a Phase II study. Phase II will consist of a prospective multi-centered study to collect both serum and clinical data from a diverse set of LD patients and controls. Using the biostatistical techniques developed in Phase I, new clinical and serologic predictive models will be developed in Phase II. This bioinformatic approach has the potential for integrating clinical and serologic diagnostic approaches using a standardized serum bank.