Informatics Techniques to Optimize Mammography for Aging Populations

描述（由申请人提供）：本研究的目标是开发、整合和评估新的信息学技术，以优化老年女性乳腺癌的乳房x线摄影诊断。65岁以上患乳腺癌的妇女死亡率更高，预后也更差。这些妇女还承受着“过度诊断”的最重负担，在这种现象中，筛查发现的癌症可能不会继续引起症状或死亡。在2000年至2050年期间，65岁以上的女性人数预计将增加一倍以上（从2000万至4000万），这一认识使这些发人深省的统计数据变得更加紧迫。在这个研究项目中，我们建议开发工具，以提高浸润性乳腺癌的早期诊断，减少不必要的侵入性手术（减少假阳性），并随之而来的解决过度诊断。具体来说，我们的目标是1)开发一个概率计算机模型，该模型由一种新的机器学习算法训练，即使用逻辑提升模型（PLUM）进行预测，该模型可以根据年龄增长的女性量身定制乳腺癌风险估计；2)使用决策分析模型，利用重要的病理预后指标，即年龄背景下的细胞学分级，确定最佳乳腺活检阈值；3)使用比较有效性分析，确定这些个性化的风险预测策略和最佳行动阈值将如何改善目前65岁以上妇女的乳腺癌筛查政策。我们的多学科团队拥有创新研究的记录（包括NIH资助和在医学、工程和计算机科学文献方面的出版物），这些研究整合了最先进的信息学算法，以改善乳腺癌诊断。利用之前的经验和基础设施，我们正在构建一个全新的机器学习算法，PLUM，它使用1)归纳逻辑编程（ILP）从多关系数据中准确学习；2)使用年龄作为分区的隆起模型，以及3)将规则纳入我们的概率模型以进行准确的风险预测。我们必须使用独特丰富的临床数据，对疾病过程的深刻理解，以及这些计算工具的创造性集成。这次重新提交的新初步数据预示着科学上的成功和临床转化。如果得到支持，该项目将证明风险预测和决策分析工具可以准确评估乳腺癌风险，确定最佳的个性化活检阈值，并提供比目前在美国使用的65岁以上妇女更好的乳腺癌筛查政策。

DESCRIPTION (provided by applicant): The goal of this research is to develop, integrate, and evaluate novel informatics techniques to optimize the mammographic diagnosis of breast cancer in aging women. Women over age 65 who develop breast cancers have greater death rates and poorer outcomes. These women also bear the heaviest burden of "overdiagnosis" a phenomenon in which screening identifies cancer which may not go on to cause symptoms or death. These sobering statistics are made more urgent by the realization that the number of women > 65 is projected to more than double (from 20-40 million) between 2000 and 2050. In this research program, we propose to develop tools that will improve the early diagnosis of invasive breast cancer, minimize unnecessary invasive procedures (decrease false positives) and concomitantly address overdiagnosis. Specifically we aim to 1) develop a probabilistic computer model trained by a novel machine learning algorithm, Prediction using Logical Uplift Modeling (PLUM), that tailors breast cancer risk estimations to aging women; 2) use a decision analytic model to determine the optimal breast biopsy threshold using an important pathologic prognostic indicator, cytologic grade in the context of age, and 3) use comparative effectiveness analysis to determine how these personalized risk prediction strategies and optimal thresholds for action will improve on current breast cancer screening policies in women > 65. Our multidisciplinary team has a track record (including NIH funding and publications in the medical, engineering, and computer science literature) of innovative research that integrates state-of-the-art informatics algorithms to improve breast cancer diagnosis. Using prior experience and infrastructure, we are building a completely new machine learning algorithm, PLUM, which uses 1) inductive logic programming (ILP) to accurately learn from multi-relational data; 2) uplift modeling that uses age as a partition, and 3) rule incorporation into our probabilistic model for accurate risk prediction. We necessarily use a uniquely rich clinical data, a deep understanding of disease processes, and creative integration of these computational tools. New preliminary data presented in this resubmission foreshadow scientific success and clinical translation. If supported, this project will prove that risk prediction and decision analytic tools can accurately assess breast cancer risk, determine an optimal, personalized biopsy threshold, and provide a superior breast cancer screening policy than is currently employed in the US for women over age 65.