SCH: New Advanced Machine Learning Framework for Mining Heterogeneous Ocular Data to Accelerate

视力丧失是美国18岁以上成年人残疾的十大原因之一，也是最严重的原因之一 儿童常见的残疾状况。主要的眼部疾病是由视网膜慢性病变引起的 进行性神经变性，不幸的是是不可逆的和不可治愈的，因此早期诊断 眼科疾病是临床医生提供视网膜保护的关键。眼科新进展 成像和高通量基因分型和测序技术提供了令人兴奋的新机会， 最终提高我们对眼部疾病的理解，他们的遗传结构，以及他们对眼睛的影响。 内表型和功能。然而，现有的遗传学和视网膜图像的研究， 这就浪费了探索遗传学和视网膜图像之间相互作用的机会。因此，我们认为， 迫切需要新的机器学习和科学进步来揭示视网膜病变的遗传基础。 成像内表型和协同遗传学和成像理解疾病进展。我们 建议进行新的视网膜成像遗传学研究，以综合研究视网膜图像和 用于自动眼科疾病诊断和预后的遗传数据， 内表型和疾病进展预测。我们的团队对视网膜神经元进行了开创性的研究 遗传学，预测和图像分析，因此我们处于实现这些目标的独特位置。 具体来说，我们将研究以下目标：1）建立有效的数据集成模型， 来自多个来源的视网膜成像遗传学数据; 2）开发知识引导的学习模型， 识别高维视网膜成像遗传学数据之间的非线性关联; 3）检测 利用时间深度学习模型的视网膜数据中的纵向相互关系; 4）新的鲁棒公平度量 学习模型，以统一疾病预测和公平的度量选择; 5）应用并验证所提出的 将机器学习方法应用于来自多个独立队列的大规模视网膜成像遗传学数据。的 成功完成这一提案将产生尖端的机器学习工具，以促进 自动化疾病诊断和准确的疾病发展和进展的长期预测 这将加强疾病的早期预防和目前的临床管理， 为新的精密治疗开发提供见解。

Vision loss is among the top 10 causes of disability in the U.S in adults over the age of 18 and one of the most common disabling conditions in children. The major ocular diseases are caused by the retinal chronic progressive neurodegeneration and unfortunately are irreversible and incurable, thus the early diagnosis of ocular diseases is crucial for clinician to provide retinoprotection. Recent advances in ophthalmological imaging and high throughput genotyping and sequencing techniques provide exciting new opportunities to ultimately improve our understanding of ocular diseases, their genetic architecture, and their influences on endophenotype and function. However, existing studies of genetics and retinal images are only conducted separately, wasting the opportunity to explore the interplay between genetics and retinal images. Therefore, there is a critical need for new machine learning and scientific advances to reveal genetic basis of retinal imaging endophenotypes and to synergize genetics and imaging for understanding disease progression. We propose to conduct the novel retinal imaging genetics research to integratively study both retinal images and genetic data for automated ocular disease diagnosis and prognosis, genetic association study of endophenotype, and disease progression prediction. Our group has performed pioneering research on retinal genetics, prediction, and image analysis, therefore we are in a unique position to achieve these goals. Specifically, we will investigate the following aims: 1) build efficient data integration models to integrate retinal imaging genetics data from multiple sources; 2) develop knowledge guided learning models for identifying nonlinear associations among high-dimensional retinal imaging genetics data; 3) detect the longitudinal interrelations in retinal data utilizing temporal deep learning model; 4) new robust fair metric learning model to unify the disease prediction and fair metric selection; 5) apply and validate the proposed machine learning methods to large-scale retinal imaging genetics data from multiple independent cohorts. The successful completion of this proposal will produce cutting-edge machine learning tools to facilitate automated disease diagnosis and accurate long-term prediction of disease development and progression trajectory, which will enhance the early prevention and current clinical management of the disease and will provide insights for novel precision treatment development.