A Comprehensive Strategy to Detect Glaucoma Worsening Earlier and With Fewer Tests

摘要 这是一份K23指导患者导向研究职业发展奖的申请书。这个 这项提议的目标是为应聘者提供建立独立的 青光眼诊断测试领域的研究方案，具有测试纠错和 未来青光眼预后的预测模型。为了促进这一长期目标，在目前的建议中， 候选人的主要研究目标是减少检测青光眼所需的时间和次数 通过(1)纠正先前获得的视野(VF)和乳突周围光学中的误差而恶化 使用贝叶斯分析(MLB)和生成性的多水平模型进行相干层析成像(OCT)测试 对抗性网络(GAN)(2)对青光眼快速恶化的高风险和低风险眼进行分层 使用深度卷积神经网络(DCNN)进行基线临床访问。这些目标是以高质量为基础的 初步数据表明：(1)VF可靠性指标和OCT信号强度对测试误差的影响 机器学习方法可以预测未来VF的风险 仅通过基准访问VF数据，并因此添加结构(OCT)和 基线访问的临床信息可能会提高模型的准确性。该模型的主要假设 提出的研究目标是(1)用MLB和GaN校正测试错误将减少所需的时间 分别检测10%和20%的恶化(2)结合基线访问结构(OCT)、功能(VF)和 临床数据作为DCNN的输入将使我们能够在AT的接收器操作曲线下获得区域 预测未来快速青光眼恶化的风险至少为0.8。候选人提出了一项全面的 培训计划，结合正式课程、会议、研讨会和研讨会，由他的不同团队监督 良师益友。具体培训目标包括：(1)接受多水平回归建模和 贝叶斯分析技术。(二)精通数据科学，特别注重学习 用于数据提取、操作和分析的Python。(3)加深对机器学习的认识 特别强调深度学习的技术，包括DCNN和GANS。(4)持续进修 研究的道德和负责任的行为。培训计划将与SET协调执行 上述研究活动的一部分。这项研究提案的结果将用于开发一种 随后的R01研究建议，将促进候选人向独立研究人员的过渡。

ABSTRACT This is an application for a K23 Mentored Patient-Oriented Research Career Development Award. The goal of this proposal is to provide the candidate with the advanced skills needed to establish an independent research program in the area of glaucoma diagnostic testing with special expertise in test error correction and predictive modeling of future glaucoma outcomes. To facilitate this long-term goal, in the current proposal, the candidate’s main research goal is to reduce the time and number of tests necessary to detect glaucoma worsening by (1) correcting for errors in previously obtained visual field (VF) and peripapillary optical coherence tomography (OCT) tests by using multilevel models with Bayesian analysis (MLB) and generative adversarial networks (GAN) (2) stratifying eyes at high and low risk for rapid glaucoma worsening at the baseline clinical visit using deep convolutional neural networks (DCNN). These aims are based on high quality preliminary data which show that: (1) the effect of VF reliability metrics and OCT signal strength on test error can be quantified and thus corrected for and (2) machine learning methods can predict risk of future VF progression with fair accuracy with baseline visit VF data alone and therefore adding structural (OCT) and clinical information from the baseline visit is likely to improve model accuracy. The main hypotheses of the proposed research aims are (1) correcting for test errors with MLB and GAN will reduce the time needed to detect worsening by 10 and 20% respectively (2) combining baseline visit structural (OCT), functional (VF) and clinical data as inputs into DCNNs will allow us to achieve an area under the receiver operating curve of at least 0.8 at predicting the risk of future rapid glaucoma worsening. The candidate proposes a comprehensive training plan, combining formal coursework, meetings, seminars and workshops overseen by his diverse group of mentors. Specific training goals include: (1) Receiving training in multi-level regression modeling and Bayesian analysis techniques. (2) Becoming adept at data science with a special emphasis on learning Python for data extraction, manipulation and analysis. (3) Furthering knowledge of machine learning techniques with a specific emphasis on deep learning including DCNNs and GANs. (4) Continuing training in the ethical and responsible conduct of research. The training plan will be executed in coordination with the set of research activities mentioned above. Results from this research proposal will be used to develop a subsequent R01 research proposal that will facilitate the candidate’s transition to an independent researcher.