An Adaptive Strategy for Preventing and Treating Lapses of Retention in Adult HIV Care II (ADAPT-R II)

摘要 这个R34应用程序最好在NIH资助的顺序多重 分配随机试验（SMART）“自适应策略，以预防和治疗失 保留（ADAPT-1）”接近完成，未来的试验（ADAPT-3）由观察结果驱动 从ADAPT-1长期坚持艾滋病毒治疗是一种典型的 复杂的公共卫生问题，需要创新的解决方案。强度的多样性和 参与的各种障碍意味着没有任何一种干预措施是所有人都需要的，也不会对所有人都有效。 都需要帮助例如，咨询可以帮助患者经历耻辱，但不会帮助 一个想来但负担不起交通费用的人。为了解决这个难题， 我们进行了一个SMART（ADAPT-1）测试家庭的自适应保留策略。通过 在那些做得好的人中保持较低强度的干预， 效率，而在那些做得不好的升级提高了有效性。在ADAPT-1中，我们 最初将患者随机分配到三种较低强度干预之一（标准护理（SOC）， 短信和有条件的现金转账）。只有那些未能持续保留的人 被重新随机分配到三种更密集的干预措施之一（SOC外展，短信 有条件现金转移或导航器）。适应技术方案1的新成果（即将 出版物）证实了我们最初的假设， 行为改善结果，该研究还揭示了额外的机会，以延长 “精准公共卫生”范式。具体来说，我们观察到不同的患者（基于 社会人口统计学、临床和实验室特征）对不同的 适应性保留策略。这一观察结果引出了一个进一步的假设：使用预测 分析（使用尖端的机器学习技术进行优化）， 干预（例如，SOC、现金转账、短信），以最有可能对该方案做出反应的患者 干预措施可以在效力和效率上比任何单一的顺序干预措施都有更大的提高。 保持战略，即使战略本身已经适应。我们计划未来的R 01应用程序， 测试基于机器学习的保留干预措施的分布，与最佳单一 序贯适应性干预（来自ADAPT-1）。为了准备新的审判，我们建议 R34，以（1）开发和测试提供按需预测的信息技术基础 该领域的医疗保健工作者，（2）完善机器学习能力的统计基础 通过模拟进行预测，以及（3）评估基于机器学习的建议的适合性 在肯尼亚卫生系统的组织、政策和道德背景下。

Abstract This R34 application is best understood in the context of both a NIH-funded sequential multiple assignment randomized trial (SMART) “Adaptive Strategies to Prevent and Treat Lapses of Retention (ADAPT-1)” nearing completion and a future trial (ADAPT-3) motivated by observations from the ADAPT-1. Retention in HIV treatment over long periods of time represents an archetypal complex public health problem and requires innovative solutions. The diversity of intensities and types of barriers to engagement mean that no single intervention is needed by all nor will work for all in need. For example, counseling could help a patient experiencing stigma, but will not help an individual who wants to come but cannot afford transportation. To respond to this conundrum, we carried out a SMART (ADAPT-1) to test a family of adaptive retention strategies. By maintaining lower intensity interventions in those doing well, adaptive strategies optimize efficiency, while escalating in those not doing well enhances effectiveness. In ADAPT-1, we initially randomized patients to one of three lower intensity interventions (standard of care (SOC), SMS messages and a conditional cash transfer). Only those who fail to be consistently retained are re-randomized to one of three more intensive interventions (SOC outreach, SMS message with a conditional cash transfer, or a navigator). Emerging ADAPT-1 results (in forthcoming publications) confirm our original hypothesis that pegging the retention intervention to patient behavior improves outcomes, the study also revealed additional opportunities to extend a “precision public health” paradigm. Specifically, we observed that different patients (based on sociodemographic, clinical and laboratory characteristics) respond differentially to different adaptive retention strategies. This observation begs a further hypothesis: use of predictive analytics (optimized with cutting-edge machine learning techniques) to distribute each intervention (e.g., SOC, cash transfer, SMS) to those patients most likely to respond to that intervention can achieve further gains in effectiveness and efficiency over any single sequenced retention strategy, even if strategy is itself already adaptive. We plan a future R01 application to test a machine learning based distribution of retention interventions as compared to best single sequential adaptive interventions (from ADAPT-1). To prepare for the novel trial, we propose this R34, to (1) develop and test the information technology basis for delivering on-demand predictions to health care workers in the field, (2) refine the statistical foundations of machine learning ability to predict through simulations and (3) assess the fit of machine learning based recommendations in the organizational, policy and ethical context of health systems in Kenya.