Logic Forest: A Statistical Method for Biomarker Discovery

逻辑森林：生物标志物发现的统计方法

• 批准号: 7686707
• 负责人: Elizabeth H. Slate
• 金额: $ 7.29万
• 依托单位: MEDICAL UNIVERSITY OF SOUTH CAROLINA
• 依托单位国家: 美国
• 财政年份: 2008
• 资助国家: 美国
• 起止时间: 2008-09-15 至 2010-08-31
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/7686707
• 关键词: Biological; Biological Markers; Cancer Detection; Categories; Cause of Death; Classification; Clinical; Complex; Development; Diagnosis; Diagnostic; Diagnostic tests; Disease; Division of Cancer Prevention; Early Diagnosis; Evaluation; Genes; Goals; Individual; Logic; Malignant Neoplasms; Methodology; Methods; Modeling; Outcome; Patients; Performance; Physicians; Primary Prevention; Public Health; Recurrent Malignant Neoplasm; Research; Risk; Risk Assessment; Screening procedure; Sensitivity and Specificity; Severity of illness; Statistical Methods; System; Techniques; Technology; Testing; Translations; Trees; Voting; anticancer research; base; cancer diagnosis; cancer initiation; cancer risk; cancer therapy; forest; improved; mortality; outcome forecast; pre-clinical; public health relevance; tool; tool development

DESCRIPTION (provided by applicant): Despite the many advances in cancer diagnosis and therapy, cancer remains the second leading cause of death in the US. This is due in part to lack of adequate methods for early detection [1-3]. Multiple studies cite the need for a noninvasive test that would be both sensitive and specific for a particular cancer [5-8]. A diagnostic test based on multiple biomarkers may be necessary to achieve the desired sensitivity and specificity [12]. Statistical methodologies that can model complex biologic interactions and that are easily interpretable allow for the translation of biomarker research into diagnostic tools. Logic regression, a relatively new multivariable regression method that predicts binary outcomes using logical combinations of binary predictors, has the capability to model the complex interactions in biologic systems in easily interpretable models [9]. The three specific aims for this proposal will develop and assess new statistical methods that extend the capability of current logic regression methodology to improve identification and evaluation of combinations of biomarkers for cancer. Aim 1 extends logic regression from a single logic tree model to an ensemble of logic trees for analysis of binary predictors and a binary outcome. The ensemble model will classify binary outcomes by popular vote in the ensemble. 1a: The predictive accuracy of the ensemble of logic trees model will be compared to the predictive accuracy of a single logic tree model and competing ensemble methods. 1b: The ability of the ensemble model to correctly identify important individual predictors will be evaluated. 1c: The ability of the ensemble to correctly identify combinations of predictors will be evaluated. Aim 2 will extend logic regression methodology to handle classification of ordinal outcomes rather than binary outcomes only. Aim 3 will extend the method developed in aim 2 from a model including only a single ordinal logic tree to an ensemble of ordinal logic trees model for analysis of binary predictors and ordinal outcomes. 3a: Assess the ability of an ensemble of ordinal logic trees to correctly classify observations compared to a single ordinal logic tree model and competing ensemble methods. 3b: The ability of an ensemble of ordinal logic trees to correctly identify important predictors will also be evaluated. 3c: The ability of an ensemble of ordinal logic trees to correctly identify important combinations of predictors will be evaluated. This research fits the NCI Division of Cancer Prevention goal of "Identification, development, and evaluation of biological analytic techniques, methodologies, and clinical technologies relevant to pre-clinical cancer detection and prevention of primary and recurrent cancers." This research is relevant to public health because the objective is to provide analytic tools that will aid in the development of tools for use in cancer risk assessment, screening, prognosis, and treatment which have the potential to decrease the rate of cancer related mortality.

描述（由申请人提供）： 尽管癌症诊断和治疗取得了许多进展，但癌症仍然是美国第二大死亡原因。部分原因是缺乏足够的早期检测方法[1-3]。多项研究指出需要一种对特定癌症既敏感又特异的非侵入性检测[5-8]。为了达到所需的敏感性和特异性，可能需要基于多种生物标志物的诊断测试[12]。可以模拟复杂的生物相互作用并且易于解释的统计方法可以将生物标志物研究转化为诊断工具。逻辑回归是一种相对较新的多变量回归方法，它使用二元预测变量的逻辑组合来预测二元结果，能够以易于解释的模型对生物系统中的复杂相互作用进行建模[9]。该提案的三个具体目标将开发和评估新的统计方法，扩展当前逻辑回归方法的能力，以改进癌症生物标志物组合的识别和评估。目标 1 将逻辑回归从单个逻辑树模型扩展到逻辑树集合，以分析二元预测变量和二元结果。集成模型将通过集成中的大众投票对二元结果进行分类。 1a：逻辑树集成模型的预测精度将与单个逻辑树模型和竞争集成方法的预测精度进行比较。 1b：将评估集成模型正确识别重要个体预测变量的能力。 1c：将评估集成正确识别预测变量组合的能力。目标 2 将扩展逻辑回归方法来处理序数结果的分类，而不仅仅是二元结果。目标 3 将目标 2 中开发的方法从仅包含单个序数逻辑树的模型扩展到用于分析二元预测变量和序数结果的序数逻辑树集合模型。 3a：与单个序数逻辑树模型和竞争集成方法相比，评估序数逻辑树集成正确分类观测值的能力。 3b：还将评估序数逻辑树集合正确识别重要预测变量的能力。 3c：将评估序数逻辑树集合正确识别预测变量的重要组合的能力。这项研究符合 NCI 癌症预防部门的目标“识别、开发和评估与临床前癌症检测和预防原发性和复发性癌症相关的生物分析技术、方法和临床技术”。这项研究与公共卫生相关，因为目的是提供分析工具，帮助开发用于癌症风险评估、筛查、预后和治疗的工具，从而有可能降低癌症相关死亡率。