Integrating Machine Learning and Physician Expertise for Breast Cancer Diagnosis

描述（由申请人提供）：本研究的目标是开发新的机器学习技术，将医生的专业知识和机器学习的逻辑规则整合到图形模型中，以准确估计乳房活检后的乳腺癌风险。我们的多学科团队有一个记录（包括NIH资助和医学和计算机科学文献中的出版物），说明了一个创新的研究计划，该计划融合了尖端的机器学习算法，包括归纳逻辑编程和统计关系学习，以训练图形模型来预测乳腺癌风险。然而，与之前的工作相比，我们正在测试一种全新的方法，我们称之为基于建议的学习（ABLe）。通过开发ABLe，我们的团队旨在在机器学习和医生专业知识之间建立一个创新的协作循环。我们建议测试这样一个假设，即这种循环将提高精确度，超出机器或人类单独完成的能力。具体来说，我们首先假设使用传统机器学习算法训练的传统训练图形模型可以准确预测核心活检后乳腺癌的风险，并且比目前的临床实践表现更好；我们的新初步数据有利地预示了一个关键的目标，但这是劳动密集型的，因为我们必须完善我们独特的临床数据，以准确地代表临床经验。其次，使用ABLe训练的图形模型可以将多关系数据与医生专业知识结合起来，并且比传统训练的图形模型和当前临床实践显著提高预测准确性。第三，我们用ABLe训练的最佳图形模型可以准确地估计新的临床病例中乳腺活检后恶性肿瘤的概率，比单独的医生更好，这是一种有潜力改善护理的工具。

DESCRIPTION (provided by applicant): The goal of this research is to develop novel machine learning techniques to integrate physician expertise and machine learned, logical rules in a graphical model that will accurately estimate breast cancer risk after breast biopsy. Our multidisciplinary team has a track record (including NIH funding and publications in the medical and computer science literature) illustrating an innovative research program that merges cutting edge machine learning algorithms including inductive logic programming and statistical relational learning to train graphical models to predict breast cancer risk. However, in contrast to prior work, we are testing a completely new methodology which we call Advice-Based-Learning (ABLe). By developing ABLe, our team aims to establish an innovative, collaborative cycle between machine-learning and physician expertise. We propose to test the hypothesis that this cycle will increase accuracy beyond what either the machine or human can accomplish alone. Specifically, we hypothesize first that a conventionally-trained graphical model trained with conventional machine learning first algorithms can accurately predict the risk of breast cancer after core biopsy and perform better than current clinical practice; a critical aim that is favorably foreshadowed by our new preliminary data but is labor intensive because we must perfect our unique clinical data that accurately represents clinical experience. Second, a graphical model trained using ABLe can incorporate multi-relational data with physician expertise and significantly improve the predictive accuracy over conventionally trained graphical models and current clinical practice. Third, our best graphical model trained with ABLe will accurately estimate the probability of malignancy after breast biopsy on new clinical cases better than physicians alone resulting in a tool that has the potential to improve care. Our clinical application is as compelling as our algorithmic work. Image-guided core needle biopsy of the breast is a common procedure that is imperfect, has high-stakes, and is particularly amenable to improvement with automated decision support. Breast core biopsy, the standard of care for breast cancer diagnosis, can be "non-definitive" in 5-15% of women undergoing this procedure. This means that between 35,000-105,000 women will require additional biopsies or radiologic follow-up to cement a diagnosis and risk the possibility of missed breast cancers, delays in diagnosis, and unnecessary surgeries. This important problem is emblematic of a plethora of clinical situations where rigorous and accurate risk estimation of rare events provides the opportunity for automated decisions support tools to personalize and strategically target health care interventions to improve decision-making for health-care providers and patients. This award will enable us not only to produce graphical models that provide improved decision support in the breast cancer clinic, but also, and more significantly, to develop a methodology that integrates heterogeneous predictive data and physician knowledge within a graphical model, thereby developing and validating a new algorithmic paradigm for creating accurate, comprehensible, adaptable decision support tools well-suited for clinical translation.