Adaptive Interventions for Optimizing Malaria Control: a Cluster-Randomized SMART Trial

In the past decade, massive scale-up of long-lasting insecticide-treated nets (LLINs) and indoor residual spraying (IRS) in Africa have led to significant reductions in malaria mortality and mobility. However, current first-line interventions are not sufficient to eliminate malaria in most countries. The widespread use of pyrethroid insecticides has resulted in resistant vector populations, and high coverage of LLINs and IRS has led to increased outdoor human feeding behavior and resting behavior. These changes in vector ecology and behaviors have significantly limited the effectiveness of current first-line interventions that target indoor biting and resting mosquitoes. Furthermore, as a result of ecological changes and intervention measures, malaria risk in a locality is dynamic, and the utility of malaria intervention tools may vary as new tools are being approved and introduced and the cost of each tool differs among locations and over time. Such variations in malaria risk, vector ecology, and utility of intervention tools exemplify the need to develop optimal adaptive interventions tailored to local malaria risks, vector ecology and supply chains. The central objective of this application is to design optimal adaptive combinations of vector control interventions to maximize reductions in malaria burden based on local malaria transmission risks, changing vector ecology, and available mix of interventions approved by the Ministry of Health in each target country. The central hypothesis is that an adaptive approach based on local malaria risk and changing vector ecology will lead to significant reductions in malaria incidence and transmission risk. To accomplish this objective, the investigators propose the following three specific aims: 1. Measure malaria incidence and predict risk using environmental, biological, social, and climatic features with machine learning approaches. Hypothesis: Malaria risk prediction can be improved through the use of machine learning techniques that include environmental, biological, socioeconomic, and climatic features. Approach: Each site will measure malaria incidence, prevalence and social economic factors through community surveys. Classification-based and regression-based approaches will be used to develop malaria risk predictive models, and model performance will be validated. Outcome: This Aim will establish improved malaria risk prediction models and lay an important foundation for developing intervention strategies adaptive to local vector ecology and future malaria risks using reinforced machine learning approaches. 2. Use a cluster-randomized sequential, multiple assignment randomized trial (SMART) design to develop an optimal adaptive intervention strategy. Hypothesis: Malaria control interventions that are adapted to local malaria risk and vector ecology and are cost effective can be identified using a cluster-randomized SMART design. Approach: Cluster-randomized SMART design will be used in a high transmission areas in Kenya to evaluate the impact of adaptive interventions that involve sequential and combination use of next-generation nets, indoor spraying of non-pyrethroid insecticides, and larval source management for malaria control. 3. Evaluate the cost-effectiveness and impact of an adaptive intervention approach on secondary endpoints related to malaria risk and transmission. Hypothesis: Intervention strategies adapted to local malaria risk and vector ecology will be more cost-effective in reducing malaria incidence and transmission risk than the currently-used LLIN intervention. Approach: The economic costs of individual interventions or combinations thereof will be assessed from both a provider and societal perspective using standard economic evaluation methodologies. Cost-effectiveness will be measured in terms of cost per person protected. The study will examine changes in drug and insecticide resistance and infection prevalence attributable to the adaptive interventions. Malaria interventions adapted to rapidly changing malaria risk and vector ecologies are critically needed to improve the effectiveness of malaria control measures. This study will use new techniques, including machine learning and a novel cluster-randomized SMART design, to develop optimal adaptive malaria intervention strategies. The investigators will use 84 clusters in Kisumu County in Western Kenya to conduct the trial. Since it is a sequential multiple assignment randomized trail, the trial will include several intervention stages. At each stage there will be different interventions. If an intervention is effective (i.e., yields an above threshold reduction in malaria incidence) at Stage 1, the intervention will be continued, otherwise, the intervention will be replaced by another one at Stage 2. The replacement intervention may be decided by different ways, e.g., an known effective intervention or an intervention determined by a machine learning algorithm. Since interventions in some clusters may be continued (i.e., effective) by next stage, other interventions may be replaced by different interventions, the number of interventions arms can vary from stage to stage. This is very different from ordinary cluster randomized trials. In this trial, the investigators planned to start with piperonyl butoxide (PBO) treated long-lasting insecticidal nets (PBO LLIN), indoor residual spraying with Actellic(R) insecticide, and using the routine LLIN intervention as control. Both Actellic IRS and PBO LLIN have been tested to be effective against pyrethroid resistant Anopheles malaria vectors and reduce clinical malaria. Therefore, the initial stage will have three arms, i.e., regular LLIN, PBO LLIN, and regular LLIN plus Actellic IRS. Since the investigators don't know if the effectiveness of these interventions in different clusters, the stage 2 interventions may include up to 7 arms, i.e., some arms may be split into two arms, based on the evaluation at the end of Stage 1 intervention. The investigators will begin the trial with a two-year smaller scale trial using 36 cluster and randomly assign the three interventions, i.e., regular LLIN, PBO LLIN and regular LLIN plus Actellic IRS, into these 36 cluster, with 12 clusters for each intervention. This pre-trial trial is to determine the optimal way for conducting the full-scale 84 cluster trial, including operational and effectiveness evaluation procedures, as well as cost-effectiveness analysis. The full scale 84 cluster trial will be started by Year 3. The full trial will be started from fresh, i.e., the same three interventions will be randomly assigned to the 84 clusters with 28 clusters for each interventions. Clinical malaria will be monitored using a cohort active case surveillance, parasite prevalence and vector density will be monitored using cross-sectional samplings. The results of these surveillance at the end of Stage 1 trial will be used to evaluate the effectiveness of interventions at each cluster for the Stage 1 interventions. Stage 2 interventions will be determined for each cluster based on the above evaluations, e.g., continue the same intervention or replace the intervention with different ones.