Transfer learning to improve the re-usability of computable biomedical knowledge

应聘者：我的人工智能(PHD)、公共卫生信息学等多学科背景 (硕士)，流行病学和卫生统计学(MS)，预防医学(医学学士)，我的职业 目标是成为一名独立的调查员，在人工智能和 生物医学，最初特别强调机器学习和公共卫生。 培训计划：我的K99/R00培训计划强调机器学习、深度学习和 科学沟通技能(演讲、撰写文章和拨款申请)，这将是对 我目前在人工智能、统计学、医学和公共卫生方面的优势。我有一个非常强烈的 指导团队。我的导师迈克尔·贝奇博士(小学)、格雷戈里·库珀、黄恒和迈克尔 瓦格纳，他们都在研究和职业生涯发展方面经验丰富。 研究计划：我建议的K99/R00拨款的研究目标是增加对 可计算的生物医学知识，即以计算机可解释的形式表示的知识 例如贝叶斯网络和神经网络。我把这样的表述称为模型。尽管模特们 可以在其他设置中全盘重复使用，可能会造成性能损失，甚至更多 有问题的是，模型所需的数据与中提供的数据之间存在根本不匹配 新的环境使得它们不可能被重复使用。迁移学习领域为以下方面开发算法 将知识从一种环境转移到另一种环境。迁移学习，机器学习的一个子领域， 显式区分源设置(具有我们希望重用的模型)和 目标设置，其数据不足以从数据导出模型，因此需要重复使用模型 来自源设置。我建议开发和评估几种贝叶斯网络迁移学习(BN- TL)算法和卷积神经网络传递学习算法。我的具体研究目标是 目的是：(1)进一步开发和评估BN-TL，以便跨医疗保健共享可计算知识 设置；(2)开发和评估BN-TL以随时间更新可计算知识；以及(3)开发和 评估一种深度迁移学习算法，该算法结合了不同场景中的知识。我会做的 这项研究是关于用于自动检测流感等传染病病例的模型。 影响：拟议的研究利用了我之前开发的大型数据集；因此，我 预计很快就会产生结果，对如何从 最初正在经历疫情的地区到另一个希望有最佳病例检测的地点 能力越早越好。更广泛地说，它将带来对增强的机器学习的洞察 生物医学知识共享和更新。这笔培训补助金将为我独立工作和 领导开发计算解决方案，以满足未来R01项目中的生物医学需求。

Candidate: With my multidisciplinary background in Artificial Intelligence (PhD), Public Health Informatics (MS), Epidemiology and Health Statistics (MS), and Preventive Medicine (Bachelor of Medicine), my career goal is to become an independent investigator working at the intersection of Artificial Intelligence and Biomedicine, with a particular emphasis initially in machine learning and public health. Training plan: My K99/R00 training plan emphasizes machine learning, deep learning and scientific communication skills (presentation, writing articles, and grant applications), which will complement my current strengths in artificial intelligence, statistics, medicine and public health. I have a very strong mentoring team. My mentors, Drs. Michael Becich (primary), Gregory Cooper, Heng Huang, and Michael Wagner, all of whom are experienced with research and professional career development. Research plan: The research goal of my proposed K99/R00 grant is to increase the re-use of computable biomedical knowledge, which is knowledge represented in computer-interpretable formalisms such as Bayesian networks and neural networks. I refer to such representations as models. Although models can be re-used in toto in another setting, there may be loss of performance or, even more problematically, fundamental mismatches between the data required by the model and the data available in the new setting making their re-use impossible. The field of transfer learning develops algorithms for transferring knowledge from one setting to another. Transfer learning, a sub-area of machine learning, explicitly distinguishes between a source setting, which has the model that we would like to re-use, and a target setting, which has data insufficient for deriving a model from data and therefore needs to re-use a model from a source setting. I propose to develop and evaluate several Bayesian Network Transfer Learning (BN- TL) algorithms and a Convolutional Neural Network Transfer Learning algorithm. My specific research aims are to: (1) further develop and evaluate BN-TL for sharing computable knowledge across healthcare settings; (2) develop and evaluate BN-TL for updating computable knowledge over time; and (3) develop and evaluate a deep transfer learning algorithm that combines knowledge in heterogeneous scenarios. I will do this research on models that are used to automatically detect cases of infectious disease such as influenza. Impact: The proposed research takes advantage of large datasets that I previously developed; therefore I expect to quickly have results with immediate implications for how case detection models are shared from a region that is initially experiencing an epidemic to another location that wishes to have optimal case-detection capability as early as possible. More generally, it will bring insight into machine learning enhanced biomedical knowledge sharing and updating. This training grant will prepare me to work independently and lead efforts to develop computational solutions to meet biomedical needs in future R01 projects.