Mapping and predicting HIV-transmission hotspots with phylogenetics and geospatial machine learning

通过这个项目，我们正在为艾滋病毒跟踪和预防带来一些非常新的东西，使用两套不同的大数据技术，我们已经领先于同行。 第一组技术是地理空间：我们正在验证新的方法来插入物理/社会障碍数据之间的直接观察到的城市blockface地区没有直接观察到的。 我们使用完整的土地覆盖数据（例如，卫星图像、税收数据），以考虑空间不连续性，如公园和主要道路。 我们正在研究如何在观测区域的外部空间边界之外进行推断。 我们正在编写一份方法手稿，我们的成功已经引起了当地政府机构的极大热情，它们同意提供环境数据并协助招募参与者。 第二组技术是时间性的：我们正在验证新的方法，以生成门诊病人的即时预测，即未来几个小时内药物渴望或压力的风险。 我们在机器学习模型中做到了这一点，该模型使用了几个小时的患者GPS跟踪与个人信息相结合。 我们拥有该工艺的临时专利（“移动的健康平台的方法和系统”，PCT/US 2016/029553），我们正在完成手稿以供出版。 这些技术可以共同应用于建立预防艾滋病毒的积极流行病学方法。 我们针对个人使用的时间尺度为小时的未来预测算法将适用于成为艾滋病毒传播热点的地区和社会场所的时间尺度为天、周或月。 使用病毒系统发育数据、社会接触数据和活动空间数据，我们打算在巴尔的摩市开发HIV储存库和传播风险的全面表面地图，开发用于预测地图变化的算法，并使用病毒系统发育数据为特定的关键人群（如有色人种的MSM）定制我们的算法 我们从研究参与者那里收集了数据，为这个项目提供了一些经验基础。 我们的一位博士后正在与一位校外艾滋病流行病学家合作，他可以访问城市和州的数据库，这些数据库将为我们的机器学习模型提供输入。 该项目的最终目标将是帮助将PrEP和其他预防资源集中到最需要它们的地理区域，使用预测性而不是反应性的战略。

With this project, we are bringing something very new to HIV tracking and prevention, using two different set of big-data techniques in which we are already ahead of our peers. The first set of techniques is geospatial: we are validating new ways to interpolate physical/social-disorder data between directly observed city blockface to regions not directly observed. We use full-land-coverage data (e.g., satellite imagery, tax data) to account for spatial discontinuities such as parks and major roads. We are working on ways to extrapolate beyond the outer spatial boundaries of the regions where observations have been made. We have a methods manuscript in preparation, and our success has already elicited great enthusiasm from local governmental agencies that have agreed to provide environmental data and assist in participant recruitment. The second set of techniques is temporal: we are validating new ways to generate live predictions of outpatients imminent risk of drug craving or stress several hours into the future. We do this in machine-learning models that use several hours of the patients GPS tracks in combination with person-level information. We have a provisional patent for this process ("Method and System for a Mobile Health Platform," PCT/US2016/029553) and we are finishing a manuscript for publication. Together, these techniques can be applied to create a proactive epidemiological approach to HIV prevention. The future-prediction algorithms that we use on a time scale of hours for individuals will be adapted to work on a time scale of days, weeks, or month for areas and social venues that become hotspots for HIV transmission. Using viral phylogenetic data, social-contact data, and activity-space data, we intend to develop wall-to-wall surface maps of HIV reservoir and transmission risk in the city of Baltimore, develop algorithms for prediction of changes in the maps, and use viral phylogenetic data to tailor our algorithms for specific key populations such as MSM of color We have collected data from research participants to provide some of the empirical grounding for this project. One of our postdocs is working with an extramural HIV epidemiologist, who has access to city and state databases that will provide input to our machine-learning models. The ultimate goal of the project will be to help focus PrEP and other prevention resources into the geographical areas where they are about to become most needed, using strategies that are predictive rather than reactive.