Developing Artificial Intelligence Approaches to Predicting Progressive Myopia and Risk of Myopic Complications Based on Optometry Data

在世界范围内，未矫正的屈光不正是造成视力损害的主要原因，影响1.163亿人。近视是最常见的疾病，随着我们见证“近视流行病”，其患病率正在上升。2010年，19亿人近视，占世界人口的27%，其中7000万人（2.8%）高度近视。据估计，到2050年，这些百分比将分别增加到52%和10% （Fricke et al., 2018）。近视造成的视力损害具有重大的经济影响和对生活质量的不利影响，病理性近视尤其有害，因为它会导致眼睛后部的退行性变化，导致失明。近视是白内障、青光眼、视网膜脱离和近视性黄斑变性的危险因素。据估计，高达11%的病理性近视患者会出现脉络膜新生血管，轴向伸长会导致乳头周围区域畸变，导致青光眼和视野丧失（Wong et al., 2014）。随机对照试验表明，干预措施（如双焦点隐形眼镜、低剂量阿托品）可以减缓儿童近视的进展（Chamberlain等人，2019），早期干预可能更有效。然而，目前还没有令人满意的方法来识别有进行性近视风险的个体或识别最有可能从治疗干预中受益的个体。更好地确定风险的能力将使治疗能够针对那些最有可能受益的人，提高成功率，并使验光师有信心增加他们对创新治疗的使用。如果有可能更好地识别有近视变化风险的个体，就有更大的机会进行成功的预防性治疗。该项目将利用苏格兰临床验光和眼科网络电子研究（SCONe）合作，这是一个由格拉斯哥卡利多尼亚大学和爱丁堡大学的学术合作伙伴开发的研究储存库。在苏格兰，验光师每年捕获超过100万张视网膜图像，为研究提供了丰富的基于人群的资源。SCONe正在利用验光实践中的图像建立一个精心策划的数据集，将视网膜照片和光学相干断层扫描（OCT）图像与临床信息联系起来。这是一个不断增长的纵向资源，是开发人工智能（AI）工具以改善临床决策的理想选择。该项目的主要目的是利用SCONe开发人工智能算法，该算法可用于验光实践，根据视网膜照片、OCT、屈光和其他临床特征的信息识别近视进展的高风险个体。在繁忙的实践中，验光师可能会错过治疗那些可能从预防性治疗中受益的人的机会，而人工智能工具可以帮助识别合适的个体。开发这样一种工具的目的是将眼睛分为高、中、低风险，或者根据进行性变化的可能性给出个性化的评分，即近视进展指数（MPI）。SCONe数据集的纵向性质，以及它对从苏格兰各地的初级保健验光（即高街验光师的做法）获得的数据的关注，为该项目提供了一个独特的机会。第二个目标是开发和验证一个额外的人工智能算法，仅从视网膜照片的特征来确定轴长和折射，并测试该算法区分近视、高度近视和病理性近视患者与健康个体的能力。第二种算法，再次利用SCONe数据集，旨在量化已知与近视相关的视网膜特征；例如，乳头周围萎缩的迹象；并有助于识别近视的新的视网膜生物标志物，例如，视网膜模式的变化

Worldwide, uncorrected refractive error is the leading cause of visual impairment, affecting 116.3 million people. Myopia (short-sightedness) is the most common disorder, and its prevalence is increasing as we witness a 'myopia epidemic'. In 2010, 1.9 billion people, 27% of the world's population, were myopic, with 70 million (2.8%) highly myopic. It is estimated that these percentages will increase to 52% and 10% respectively by 2050 (Fricke et al., 2018). Visual impairment from myopia has a significant economic impact and adverse effect on quality of life, with pathologic myopia particularly harmful as it leads to degenerative changes at the back of the eye causing blindness.Myopia is a risk factor for cataract, glaucoma, retinal detachment and myopic macular degeneration. It is estimated that up to 11% of people with pathological myopia develop choroidal neovascularization and axial elongation can cause distortion of the peripapillary region leading to glaucoma and loss of visual field (Wong et al., 2014). Randomized control trials have shown interventions (e.g. bifocal contact lenses, low-dose atropine) can slow progression of childhood myopia (Chamberlain et al., 2019) and earlier interventions are likely to be more effective. However, at present there is no satisfactory way to identify individuals at risk of progressive myopia or to identify those most likely to benefit from treatment intervention. The ability to better determine risk would enable treatments to be targeted for those most likely to benefit, increasing success rates, and giving optometrists confidence to increase their use of innovative treatments. If it is possible to better identify individuals at risk of myopic changes, there is greater opportunity for successful preventative treatment.The project will utilise the Scottish Clinical Optometry and Ophthalmology Network e-research (SCONe) collaboration, a research repository developed by academic partners at Glasgow Caledonian University and University of Edinburgh. In Scotland, over 1 million retinal images are captured by optometrists each year, providing a rich population-based resource for research. SCONe is utilizing images from optometry practices to build a curated dataset, incorporating retinal photographs and optical coherence tomography (OCT) images linked to clinical information. This is a growing, longitudinal resource, ideal for developing Artificial Intelligence (AI) tools to improve clinical decision making.The main aim of this project is to use SCONe to develop AI algorithms that can be used in optometric practices to identify individuals at high risk of myopic progression based on information from retinal photographs, OCT, refraction, and other clinical features. Within a busy practice, optometrists may be missing opportunities to treat those likely to benefit from preventative treatments, and an AI tool could facilitate identification of suitable individuals. The objective of developing such a tool would be to categorize eyes as high, medium, or low risk, or attribute an individualized score indicating the likelihood of progressive changes, i.e., a Myopia Progression Index (MPI). The longitudinal nature of the SCONe dataset, and its focus on data acquired from primary care optometry (i.e. high street opticians' practices) from across Scotland, provides a unique opportunity for this project. A secondary aim is to develop and validate an additional AI algorithm to determine axial length and refraction from features of retinal photographs alone and to test the algorithm in its ability to differentiate patients with myopia, high myopia and pathological myopia from healthy individuals. This second algorithm, again utilising the SCONe dataset, will aim to quantify retinal features already known to be associated with myopia; for example, signs of peripapillary atrophy; and facilitate identification of new retinal biomarkers of myopia, for example, changes in patterns of retin