A Machine Learning-Based Mobile Application and Cloud Platform to Enable Accurate and Streamlined Surveillance of Soil-Transmitted Helminth Infection and Schistosomiasis

项目摘要/摘要 土壤传播蠕虫(STH)感染和血吸虫病影响着20亿人，并具有显著的 对健康有害的影响。实施性传播感染和血吸虫病干预的战略目前依赖于 通过对粪便样本进行显微分析来检测这些寄生虫，以检测寄生虫卵并鉴定虫卵 物种。准确的监督检测和及时准确的结果报告是有效的 在规划层面上进行决策，以实施感染控制战略。增加的方法 基于显微镜的测试和简化报告的速度和准确性标准化可能会有所帮助 消灭STH感染和血吸虫病。 我们建议开发一种基于手机的STH-血吸虫卵识别和计数工具，该工具 采用机器学习(深度学习)，在没有互联网连接的情况下工作。有了这款应用， 用户将收集监控数据，以便集成到云平台中。然后可以将监控数据可视化 在仪表板上为控制疾病的干预措施提供信息。 我们的方法与其他已发表的开发机器学习算法的工作有根本的不同 对于STH和血吸虫病，因为它将在监测活动中非常准确地识别虫卵类型， 它将在应用程序中提供给用户，并与云存储和报告集成。我们的跨学科 团队结合了全球健康研究人员、产品可用性测试专家、显微镜专家和 数据科学家。 在R21阶段，我们将收集有史以来最大的STH和血吸虫卵的显微镜图像集(&gt；15 000)。我们将训练一种基于卷积神经网络的算法，这种算法可以制造高精度的寄生虫卵 分类(物种识别)，并将此算法嵌入到无需互联网的移动应用程序中 连通性。为了促进APP的实用性，我们将在监控环境中评估其准确性和可用性。我们 通过构建一个Web应用程序来服务于 深度学习模型，识别STH和血吸虫卵，准确率为98%。 只有在实现了明确的里程碑时，才会进行R33阶段。我们会进一步发展 移动应用程序作为数据捕获系统，将与云存储和动态数据可视化相集成 系统能够提高STH和血吸虫病随时间和跨地区监测的准确性 地理位置。​验证研究将评估系统在节省时间和成本方面的​优势，并 在监测活动期间收集的数据的质量。这项工作的总体目标是提高精度 并简化性传播感染和血吸虫病监测，以便在疾病控制中做出有效决策。

PROJECT SUMMARY/ABSTRACT Soil-transmitted helminth (STH) infections and schistosomiasis affect 2 billion people and have significant detrimental effects on health. Strategies to implement STH and schistosomiasis interventions currently rely on testing for these parasites by microscopic analysis of stool samples to detect parasite eggs and identify egg species. Accurate surveillance testing and timely and accurate reporting of results are required for effective decision-making at the programmatic level to implement infection control strategies. Approaches that increase the speed and standardize the accuracy of microscopy-based testing and streamline reporting could help eliminate STH infections and schistosomiasis. We propose to develop a mobile phone-based STH-schistosome egg identification and counting tool that employs machine learning (deep learning) and works in the absence of an internet connection. With this app, users will collect surveillance data for integration into a cloud platform. Surveillance data can then be visualized in dashboards to inform interventions to control disease. Our approach is fundamentally different from other published work that develop machine learning algorithms for STH and schistosomiasis because it will very accurately identify egg types during surveillance activities, and it will be available to users in an app and integrate with cloud storage and reporting. Our interdisciplinary team combines the expertise of global health researchers, product usability testing experts, microscopists, and data scientists. In the R21 phase, we will collect the largest ever microscopy image set of STH and schistosome eggs (> 15 000). We will train an algorithm based on convolutional neural networks that make highly accurate parasite egg classification (species identification) and embed this algorithm into a mobile app that works without internet connectivity. To promote app utility, we will evaluate its accuracy and usability in a surveillance setting. We established the feasibility of our approach in preliminary data by building a web app that serves the results of a deep learning model that identifies STH and schistosome eggs with > 98% accuracy. The R33 phase will be only undertaken if well-defined milestones are achieved. We will further develop the mobile app as a data capture system that will integrate with cloud storage and a dynamic data visualization system to enable increased accuracy in STH and schistosomiasis surveillance over time and across geographic location. ​Validation studies will assess the​ benefits of the system to time and cost savings and quality of data collected during surveillance activities. The overall goal of this work is to increase the accuracy and streamline STH and schistosomiasis surveillance to enable effective decision-making in disease control.