Where there is no death certificate: Using artificial intelligence to detect high-casualty epidemics from satellite imagery of burial sites - Resubmission - 1

检测高伤亡的流行病对于卫生当局前瞻性地减少疾病传播至关重要 并追溯到疫情发生后的健康状况--这对于传染病来说可以 包括癌症、糖尿病、神经发育障碍、“长冠状病毒病”等后遗症。不幸的是， 许多低收入和中等收入国家(LMIC)缺乏疫情检测的数据系统，因此缺乏足够的数据支持。 有能力调动资源以减少流行病的传播并解决其对健康的影响。这个 新冠肺炎大流行见证了使用卫星图像检测疫情期间死亡人数的一些首创努力 埋葬地点。尽管取得了成功，但这些都是小的“一次性”努力，因为对卫星的人工分析 图像是极其劳动密集型的。我们建议开发一种用于全自动测量的算法 使用卫星图像的墓地占有率。我们的探索性研究将集中在坦桑尼亚，它是一个典型的 这种算法的高优先级用例，因为它受到了过去两次最致命的流行病的沉重打击 世纪-艾滋病毒/艾滋病和新冠肺炎-自2020年5月以来就没有正式报告过新冠肺炎的统计数据。 我们的算法将作用于已经收集的卫星图像，这些图像可以在任何给定的区域获得 坦桑尼亚--实际上也是世界--可以追溯到不到两周前。在目标1中，我们将发展一个区域- 基于卷积神经网络(R-CNN)的自动计算墓地占有率的方法 最新的可用图像。我们将在图像中手动标记墓地，以进行算法训练和测试， 并将通过实地考察来验证标签，以计算墓地的真实占有率。在目标2中，我们将 开发一种新颖的“点对点”CNN(SD-CNN)来比较早期和后期图像中的入住率 同样的地点。我们假设SD-CNN将比R-CNN更准确，因为算法 将有关于站点在较早时间点的外观的信息，并且可以被训练来注意新的 埋葬地块，忽略照明和植被等“背景”变化。我们再次训练和测试 使用已标记图像的算法，并将通过观察数据的实地访问来验证我们的标记 墓地上的记号。最后，在目标3中，我们将测试该算法识别死亡率变化的能力 由于流行病。在坦桑尼亚，我们预计埋葬地点将显示由于艾滋病毒/艾滋病而导致的死亡率上升，这是由于 艾滋病毒治疗规模扩大，并因新冠肺炎而陡然上升。我们对卫星数据的初步观测 确认与2019年相比，坦桑尼亚的墓地占有率在2020年显著增加。 如果成功，我们的算法将使世界上第一个低成本、可扩展和全球公平的流行病 检测平台。回顾过去，我们的研究可以帮助确定新冠肺炎受灾最严重的地区，有助于 LMIC将调集急需的资金，以应对疫情的后果。展望未来，我们的研究 可以帮助人类更安全地抵御未来的大流行，特别是那些出现在LMICs和可能 否则，不被发现或不被报告，直到为时已晚。

Detecting high-casualty epidemics is essential for health authorities to prospectively reduce disease spread and retrospectively address health conditions in an epidemic’s aftermath – which for infectious diseases can include cancer, diabetes, neurodevelopmental disorders, “long COVID,” and other sequelae. Unfortunately, many low- and middle-income countries (LMICs) lack data systems for epidemic detection, and thus are ill- equipped to mobilize resources to reduce the spread of epidemics and address their health effects. The COVID-19 pandemic saw some of the first efforts to detect mortality during an epidemic using satellite imagery of burial sites. Though successful, these were small “one-off” efforts because manual analysis of satellite imagery is extremely labor-intensive. We propose to develop an algorithm for fully-automated measurement of burial site occupancy using satellite imagery. Our exploratory research will focus on Tanzania, which typifies a high-priority use case for such an algorithm because it was hard-hit by the two deadliest pandemics of the past century – HIV/AIDS and COVID-19 – and has not officially reported COVID-19 statistics since May 2020. Our algorithm will act upon already-collected satellite imagery, which can be obtained for any given area of Tanzania – and, indeed, the world – dating no more than two weeks back. In Aim 1, we will develop a region- based convolutional neural network (R-CNN) to automatically count the occupancy of burial sites using the most current available imagery. We will manually label burial plots in images for algorithm training and testing, and will validate the labeling with field visits to count the true occupancy of burial sites. In Aim 2, we will develop a novel “spot-the-difference” CNN (SD-CNN) to compare occupancy in earlier vs. later imagery of the same site. We hypothesize that the SD-CNN will be more accurate than the R-CNN because the algorithm would have information about what a site looked like at an earlier time-point and can be trained to notice new burial plots while ignoring “background” changes such as lighting and vegetation. We again train and test the algorithm using labeled imagery and will validate our labeling with field visits in which we will observe date markers on burial plots. Finally, in Aim 3, we will test the ability of the algorithm to identify changes in mortality due to epidemics. In Tanzania we expect burial sites to show a rise in mortality due to HIV/AIDS, a fall due to scale-up of HIV treatment, and an abrupt rise due to COVID-19. Our preliminary observations of satellite data confirm marked increases in burial site occupancy in Tanzania over the year 2020 relative to 2019. If successful, our algorithm will enable the world’s first low-cost, scalable, and globally equitable epidemic detection platform. Retrospectively, our research could help to identify areas hardest-hit by COVID-19, helping LMICs to marshal much-needed funding to address the pandemic’s aftermath. Prospectively, our research could help to keep humanity safer from future pandemics, especially those that arise in LMICs and may otherwise go undetected or unreported until it is too late.