CAREER: Adaptable, Intelligible, and Actionable Models: Increasing the Utility of Machine Learning in Clinical Care

近年来，临床相关数据集的可获得性有了巨大的增长。很好地理解如何组织、处理这些数据并将其转化为可操作的知识是至关重要的。本研究旨在通过探索机器学习中新的基础研究方向和方法来释放这些数据的潜力。通过计算数据驱动的模型将被确定为高危的患者作为目标，可以以经济高效的方式减轻疾病负担。虽然医学领域的机器学习机会继续增加，但在翻译实践方面取得的成功相对较少。临床医生仍然将他们日常决策的大部分建立在相对少量的患者特定数据上。这里所做的技术贡献将使复杂的医疗数据能够得到有意义的使用。除了长期的社会影响外，这项工作还将通过与拟议目标相关的研究项目提供有价值的学生培训。侧重于计算研究的社会影响的有针对性的外联活动将吸引不同类型的研究生进入该领域。此外，这项工作将有助于为一门新的以项目为基础的课程奠定基础，该课程侧重于机器学习在临床护理中的应用。随着该领域的持续发展，这些课程将成为为下一代学生配备所需工具和洞察力的关键。最后，这项工作将促进计算机科学、工程和医学之间的关键部门间合作，从而丰富这两个领域。这项建议的主要研究目标是通过探索ML的新的基础研究方向和方法，增加机器学习在临床护理中的效用。为了让数据驱动的预测模型在临床护理中得到广泛和安全的采用，ML社区仍然必须解决几个关键的研究挑战：对患者群体和临床方案中复杂的意外变化的适应性较差，准确但无法解释的模型的可理解性不够，以及缺乏可操作性，其准确性克服了可操作性。PI建议开发新的迁移学习技术，以便在广泛的场景中学习健壮和自适应的模型。对大规模临床数据集的实验和评估将提供对这些数据如何随时间变化的洞察，并更好地理解模型应该何时以及如何适应。临床决策模型和软件很少被纳入实践，因为它们要么是黑匣子，要么是输出(虽然准确)没有提供任何关于如何行动的洞察力。提高模型可理解性的一种方法是专注于构建具有临床意义的特征。另一种提高清晰度的方法是通过稀疏。PI将研究特征工程/选择方法，以学习自动利用专家知识的有用抽象，并基于可操作的特征学习模型。PI将探索结构化正则化技术来选择可修改的特征。为了更好地了解不同的行为如何影响患者的风险，PI将在高维观察数据集的背景下解决因果推断的局限性。这项研究将产生产生临床有意义的输入的方法，以及联合优化稀疏性和可操作性的方法。这项拟议的工作将产生从患者数据中提取和建立适应性强、可理解和可操作的模型的新技术。强调适应性解决方案将确保此类技术能够长期安全地被采用。对处理数据内在异质性(例如，来自多个地点的不同患者群体)的技术的研究不仅将增加数据的实用性，而且将导致转移学习领域的更广泛的进步。对可理解性的关注--这是机器学习社区经常忽视的一个品质--有望增加此类模型的实用性，因为临床医生更有可能采用他们可以检查和理解的模型。确定可操作模型的优先顺序将产生在高维观测环境中进行因果分析的新策略。这反过来将使临床医学中关于因果关系的新假说得以产生。

In recent years, the availability of clinically relevant datasets has grown enormously. A good understanding of how to organize, process, and transform these data into actionable knowledge is crucial. This research aims to unlock the potential of these data through the exploration of new fundamental research directions and approaches in machine learning. Targeting patients identified as high-risk by through computational data-driven models could reduce the burden of disease in a cost-effective manner. While machine learning opportunities in medicine continue to grow, there have been relatively few successes regarding translation to practice. Clinicians still base the bulk of their daily decisions on relatively small amounts of patient-specific data. The technical contributions made here will enable the meaningful use of complex medical data. Beyond the long-term societal impact, this work will provide valuable student training through research projects related to the proposed objectives. Targeted outreach activities that focus on the societal impacts of computational research will attract a diverse set of graduate students to the field. In addition, this work will help lay the foundation for a new project-based course focusing on applications of machine learning in clinical care. As the field continues to grow, such courses will become critical for equipping the next generation of students with the required tools and insights. Finally, critical inter-departmental collaborations between computer science and engineering and medicine will grow as a result of this work, leading to the enrichment of both fields. The primary research objective of this proposal is to increase the utility of machine learning in clinical care, through the exploration of new fundamental research directions and approaches in ML. For data-driven predictive models to become widely and safely adopted in clinical care, there remain several key research challenges that the ML community must address: poor adaptability to complex unexpected changes in patient populations and clinical protocols, insufficient intelligibility of accurate but uninterpretable models, and absence of actionability, with accuracy overcoming actionability. The PI proposes the development of new transfer learning techniques for learning robust and adaptable models in a wide range of scenarios. Experiments and evaluations with large-scale clinical datasets will offer insight into how these data change over time, and a better understanding of when and how models should adapt. Clinical decision models and software are seldom incorporated into practice because they are either black-box or the output (while accurate) does not offer any insight into how to act. One way to increase the intelligibility of models is to focus on building clinically meaningful features. Another way to increase intelligibility is through sparsity. The PI will investigate feature engineering/selection methods for learning useful abstractions that automatically leverage expert knowledge and for learning models based on actionable features. The PI will explore structured regularization techniques to select modifiable features. To gain a better understanding of how different actions affect patient risk, the PI will address the limitations of causal inference in the context of high-dimensional observational datasets. This research will yield methods for producing clinically meaningful inputs, and methods for jointly optimizing sparsity and actionability. The proposed work will yield novel techniques for extracting and building adaptable, intelligible, and actionable models from patient data. An emphasis on adaptable solutions will ensure that such techniques can be safely adopted long-term. The study of techniques for dealing with the inherent heterogeneity of the data (e.g., different patient populations from across multiple sites) will not only increase the utility of the data but will lead to more general advances in the field of transfer learning. A focus on intelligibility - a quality that is often overlooked by the machine learning community - promises to increase the utility of such models, since clinicians are more likely to adopt a model they can check and understand. Prioritizing actionable models will yield new strategies for causal analysis in high-dimensional observational settings. This, in turn, will enable the generation of new hypotheses regarding causal relationships in clinical medicine.