PROJECT SUMMARY Scientific approaches to detect and respond to gaps in the public health response to HIV face a broad challenge with profound implications: after more than a decade of intensive service scale-up, the magnitude and reasons for remaining gaps in the treatment and prevention cascades differ markedly across people, communities, organizations, and geographies. As a result, strategies to control the epidemic must efficiently identify people and places most in need of intensified HIV testing, prevention, and treatment support, and deliver to each a timely and effective intervention. This poses a distinct scientific task: methods are required to discover and leverage continually evolving information to optimize “local” estimates of how best to allocate limited resources for both measurement and intervention. To meet this need, we propose a set of theoretical and applied aims that advance the use of adaptive designs– i.e. repeated “learn-and- apply” cycles– as a strategic approach for implementation science in heterogenous environments. Our approach uses four steps. First, we will extend existing statistical theory to allow machine learning to direct ongoing adaptation in who to measure and how to intervene, thereby creating approaches that maximally leverage the increasingly rich program data available in the HIV epidemic response. Second, we will use existing data from several unique field studies to create “virtual laboratories” for method validation and evaluation: 1) a sampling study to assess mortality on treatment in 64 facilities and 165,000 patients in Zambia; 2) an 1800-person individually randomized trial to optimize retention in HIV treatment in Kenya; and, 3) a cluster randomized trial of HIV treatment as prevention in 32 communities and 150,000 adults in Kenya and Uganda. Third, we will use these data to compare the methods developed against benchmarks provided by standard approaches (e.g. fixed sampling or balanced randomization). Fourth, we will develop accessible software for several general classes of problems in the HIV epidemic response and implementation science more broadly. Aim 1 seeks to develop and evaluate adaptive sampling designs that target data gathering based on past information to better detect locations and subpopulations with elevated risk. Aim 2 seeks to develop and evaluate adaptive designs that continuously target intervention assignment to learn how best to distribute interventions with heterogeneous effects. Aim 3 seeks to develop and evaluate adaptive strategies that use past data to target combined data gathering and public health interventions to simultaneously discover regions of elevated risk and guide intervention decisions. Overall these methods offer the promise of detecting gaps in the HIV response more quickly, with fewer resources, and with fewer “missed signals”, and of intervening more effectively within a context of constrained resources by assigning interventions better tailored to individual and context-specific needs.

PROJECT NARRATIVE In the current HIV epidemic response, high variability in implementing contexts and epidemic settings demands epidemiological designs and analytic methods that are able to detect and respond to heterogeneity effectively and efficiently. We propose a series of aims that leverage the rich data increasingly generated in the course of the HIV epidemic response, and apply targeted machine learning to advance adaptive design and analytic approaches. For each aim, we: 1) propose theoretical advancements to current adaptive design methods; 2) use large-scale contemporary data from the HIV epidemic in East Africa to build virtual analytic laboratories; 3) test the performance of the proposed innovations in these settings; and, 4) develop statistical software that can be implemented by applied epidemiologists to address real world problems.