CAREER: Trustworthy and Robust Federated Learning for Computational Healthcare

美国国家工程院（National Academy of Engineering）已将“高级健康信息学”确定为21世纪改善患者护理和迅速应对突发公共卫生事件的重大挑战之一。计算医疗保健——一种数据驱动的医疗保健机器学习方法，通过支持循证医学实践、个性化患者治疗和降低成本，具有推动和革新医疗保健的巨大潜力。释放这一潜力需要无缝整合来自多个利益攸关方（如患者、医院、提供者以及地方、州和国家机构）的健康数据，并防止泄露或滥用信息的危险。然而，由于成本、患者隐私风险和数据保护法规的原因，整合来自多个利益相关者的健康数据是令人望而却步的。为了应对这些重要的挑战，利益相关者可以合作，共同构建和评估机器学习模型，而不需要使用分散的机器学习方法（称为联邦学习）共享数据。然而，在医疗保健领域使用联邦学习是脆弱的，因为存在许多挑战，包括异构的患者数据和利益相关者站点上不同的计算资源，疾病和患者的数据可用性有限，以及数据和模型容易受到对抗性攻击——所有这些都限制了联邦机器学习模型的开发。该项目通过为计算医疗保健应用程序的异构数据的生成、同化和分析开创下一代健壮和可信的联邦学习算法和方法，解决了医疗保健联邦学习的这些基本挑战，为国家对精准医疗计划的努力做出了贡献。该项目将通过以下几种方式加速医学发现和帮助临床决策，从而对现实世界产生影响：(a)安全地整合和学习分散的异构卫生数据孤岛；(b)生成疾病和患者的稳健表示；(c)通过提供灵活的开源评估工具包，建立临床医生和患者对数据驱动方法的信任。该项目还将通过课程作业和与医学专家一起进行的现实医疗保健项目，通过跨学科研究培训和本科生和研究生的技能发展，对教育和推广产生重大影响。将采取有针对性的努力，扩大妇女、代表性不足的群体和K-12群体对STEM教育的参与。该项目将开发新的算法、方法和软件，通过推进多视图学习、对抗学习和机器学习方面的最新技术，改进数据驱动的联邦学习，用于计算医疗保健。该项目的总体主题是健壮性和可信赖性，并分为三个相互关联的重点。该项目的第一个重点是“健壮的联邦学习”，它涉及开发多视图联邦学习算法和编码联邦学习方法，以解决医疗保健环境中固有的统计和系统异质性挑战。该项目的第二个重点是“可信赖的联邦学习”，在此基础上，将开发新的联邦对抗训练和可验证的联邦学习算法，以实现对抗性攻击和数据删除请求的弹性，并将采用公平和可解释的算法，使所提出的联邦学习解决方案对所有患者和利益相关者都是公正和公平的。最后，在该项目的第三个重点中，研究人员将研究和开发用于生成和评估用于联邦学习的现实综合健康数据的算法，以实现可重复性并加速联邦学习方法在医疗保健领域的发展。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The National Academy of Engineering has identified ‘Advanced Health Informatics’ as one of the grand challenges of the 21st century for improving patient care and swiftly responding to public health emergencies. Computational healthcare – a data-driven machine learning approach for healthcare, has a tremendous potential to advance and revolutionize healthcare by supporting the evidence-based practice of medicine, personalizing patient treatments, and reducing costs. Unlocking this potential requires the seamless integration of health data from multiple stakeholders such as patients, hospitals, providers, and local, state, and national agencies, and protection against the dangers of compromise or misuse of the information. However, integrating health data from multiple stakeholders is prohibitive due to cost, patient privacy risk, and data protection regulations. To tackle these important challenges, stakeholders can collaborate to jointly build and evaluate machine learning models without sharing data using a decentralized machine learning approach called Federated Learning. However, using federated learning in the healthcare domain is fragile due to a multitude of challenges, including heterogeneous patient data and varying computational resources at the stakeholders’ sites, limited data availability for diseases and patients, and vulnerability to adversarial attacks on the data and models – all of these limit federated machine learning model development. This project addresses these fundamental challenges of federated learning for healthcare by pioneering next-generation robust and trustworthy federated learning algorithms and methods for the generation, assimilation, and analysis of heterogeneous data for computational healthcare applications, contributing to the national effort toward precision medicine initiatives. The project will have a real-world impact by accelerating medical discovery and aiding in clinical decision-making in several ways: (a) securely integrating and learning from distributed heterogeneous siloes of health data; (b) yielding robust representations of diseases and patients; (c) building clinicians’ and patients’ trust in the data-driven methods by providing a flexible open-sourced evaluation toolkit. This project will also have a significant educational and outreach impact via interdisciplinary research training and skills development of undergraduate and graduate students through coursework and real-world healthcare projects with medical experts. Directed efforts will be undertaken to broaden the participation of women, underrepresented and K-12 groups in STEM education.This project will develop new algorithms, methodologies, and software to improve data-driven federated learning for computational healthcare - by advancing the state-of-the-art in multi-view learning, adversarial learning, and machine unlearning. This project’s overarching theme is robustness and trustworthiness, and is organized into three interrelated thrusts. The first thrust of the project focuses on ‘robust federated learning’, and it involves developing multi-view federated learning algorithms and coded federated learning methods to address the statistical and system heterogeneity challenges inherent in healthcare settings. The second thrust of the project is focused on ‘trustworthy federated learning’, under which novel federated adversarial training and verifiable federated unlearning algorithms will be developed to achieve resiliency to adversarial attacks and data-deletion requests, and fair and interpretable algorithms will be employed to make the proposed federated learning solutions unbiased and equitable to all patients and stakeholders. Finally, in the third thrust of the project, the researchers will study and develop algorithms for generating and evaluating realistic synthetic health data for federated learning to enable reproducibility and accelerate the development of federated learning methods in healthcare.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.