An immune system for the city: a new paradigm for control of urban disease vectors

Conventional surveillance system (control): Community members could report the presence of insects through the conventional system (in which insect specimens have to be brought for identification to a health post) or through Alerta Chirimacha, a notification system in which they can report through photographs in WhatsApp or verbal description if they are unable to take pictures. For the active search, we will assign to a vector control personnel a list of houses to visit. This list will be compiled through optimization algorithms to maximize the dual objectives of visiting high-risk houses and achieving high spatial coverage of the search area. If a house is found to be infested with T. infestans, either by active search or report, the house and its immediate neighbors will be treated.


Immune: Reports about the presence of the insect will be received in either of two forms, conventional or through AlertaChirimcha. For this approach, we allow control personnel autonomy to guide their own search based on their own experience and on the quantitative risk estimates presented in a color ramp provided via an app. In the case of T. infestans presence, a vector control specialist will be sent to conduct a thorough inspection of the household and to cover a "risk area” of 200 m around the infested household. Within this intervention zone, they will: 1) visit homes are offered entomologic inspections to those who opt for them 2) distribute flyers to all households, 2) post flyers in all local shops willing to let us do so, and 3) create targeted Facebook posts for 10 days for residents living within 1 km of the infested house to let them know an infestation was found in the area, encourage them to be aware of the insect presence and report in case they detect it. All houses found to be positive in the area and their neighbors will be treated with insecticide.

2021-10-19

2025-12-31

1a. Analytical approach. Ascertainment bias—a systematic difference in the ability or effort to detect an outcome—plagues studies of vector surveillance. A naïve evaluation of an intervention leads to perverse interpretations because a better surveillance system might uncover and eliminate more cases, while a poorer one would leave cases undetected for a longer period, thereby allowing the creation, and eventual discovery, of more cases. To avoid this paradox, our approach is to estimate the rate of detection and elimination of infestations. Using an approach from Queuing theory we estimate the rate of detecting and removing infested households, through iterative filtering, that allows for inference on parameters for stochastic processes. These methods have been integrated into an R package (pomp) allowing for ease of use and replication. Using pomp we will derive point estimates and credible intervals for the removal rate for both arms of the trial. By comparing the credible intervals around these estimates, we will ascertain if the immune arm significantly increases the rate of detection and elimination of vector foci.

Alternative analytical approach. Our methods are new, and, as with any new approach, we may run into computational issues that prevent us from implementing them in the manner that we envision. We therefore have a more traditional approach to testing our hypothesis, which, although less powerful, is easily implemented. Under the alternative approach the primary study hypothesis is: the number of infested houses uncovered will be greater in the immune arm early in the study, and greater in the conventional arm later in the study. We will model the number of confirmed infestations using a negative binomial regression, and considering as covariates the study arm, the year of the study (1-4) and the interaction of the two (arm * year). Our expectation, under our alternative primary hypothesis, is a significantly positive coefficient for the study arm, and a significantly negative coefficient for the interaction of study arm and year.

Originally planned Year 3 Interim analysis (2023-Mar-01) We compare the rate of detection of infested households (defined as the number of households detected / the number of person days of effort in the arm) in an intervention versus control arm.
*We hypothesize that we will find more infested households in the intervention than in the control arm.

FOLLOWING THE INTERIM ANALYSIS: The interim analysis showed significantly more infested houses detected in the intervention arm than control, and prompted the decision to provide the intervention to all arms. Control arms began receiving the intervention on May 15th, 2023. Our hypothesis for the following year (May 15 2023 - May 15 2024) is: The number of infested households detected will be greater in the catchments originally assigned to the control arm (because the control arm, presumably, was less efficient detecting and eliminating the infestations in the original trial) than those originally assigned to the intervention.

Secondary outcomes
As we will apply the framework from Implementation Science—the Reach, Effectiveness, Adoption, Implementation, and Maintenance (RE-AIM) we will also evaluate the following outcomes, comparing the intervention and control arms:

The percent of households who know the steps for reporting a Triatomine.
The percent of households who can recognize a Triatomine.
The time (days) between vector sighting and re-treatment of the infested household to assess system-wide delay.
The spatial coverage of search routes of MOH vector control specialists
The size of infestation (number of insects captured)
The cost of each surveillance-and-response strategy over 2 year-time periods.

Cluster randomized trial

Not available

We balanced the arms based on the number of years that had passed since application of insecticide to the catchment by the Ministry of Health (The 'attack' phase), the prevalence of infestation at the time of insecticide application, and the number of households in the catchment that were included in the attack phase treatment. We assigned twice the weight to the number of households as compared to the other two variables. We used the R package nbpmatching, which minimizes weighted Mahalanobis distances to create balanced pairs of catchments, and randomly assigned one set to the intervention and the control through a random number generator.

Catchments of ~2,225 households

Yes

Original: 32
Post-covid: 60 (increased to maximize power for a shortened intervention)

NA

Original Power calculations:
We estimated power via simulation.We simulated expected counts of infestations in each catchment (of 3,000 households), assuming our hypothesis were true (higher counts in years 1 & 2 in the immune arm, and lower thereafter, and the converse in the conventional arm). We set the mean expectation to 1 instance per catchment per year (low) and 3 instances per catchment per year (high). We set the dispersion parameter of the negative binomial such that 58% of catchments would have no instances of infestation in the lower count, drawing from our historical data. We then analyzed these count data using the negative binomial regression including three covariates (trial arm, year of the trial, and the interaction of the two). To estimate power, we tally the number of 1000 simulations in which the p-value for each covariate is <0.05. Given these assumptions, a 4-year trial with a sample of 32 catchments (16/arm) would provide us with 90.6% power to detect a significant effect of the arm of the trial, and, importantly, 88.8% power to detect a significant effect of both trial arm and the interaction between trial arm and year.

Post covid: following delay due to covid lockdowns in Peru we increased our sample size to 30 catchments in immune and 30 catchments in control to maximize our power during a shortened intervention