JMIR Publications

ABSTRACT

Background:

Urban arboviruses remain one of the main public health challenges in Brazil, and in 2024 the Federal District recorded the highest national dengue incidence rate. Conventional entomological surveillance, such as the Rapid Index Survey for Aedes aegypti (LIRAa), depends on household inspection by field agents and is limited by sampling periodicity, spatial granularity, and difficulty of access to breeding sites. Digital tools for participatory surveillance address predictive surveillance, civic engagement, and health communication in a fragmented manner and are directed primarily at health professionals rather than citizens.

Objective:

To describe the development of a mobile application prototype in which an artificial intelligence (AI) agent actively guides the resident through a household inspection for Aedes aegypti larval breeding sites, with image classification performed on the device itself.

Methods:

The Design Science Research Methodology (DSRM) was adopted, reporting stages 1 to 4. Requirements were elicited through a literature review, a prior-art search in the Brazilian National Institute of Industrial Property (INPI) Computer Program database, technical meetings with Federal District health surveillance agencies, and five workshops with 17 public secondary school students. The prototype integrates an embedded computer vision module (YOLOv10n with 8-bit post-training quantization), a hybrid AI agent that uses a language model on in-house infrastructure when connectivity is available and a local approximate-matching engine otherwise, and a georeferenced risk map. The detection model was trained on three public image datasets, remapped into three public health–relevant superclasses and split by source photograph to prevent leakage. Performance was estimated on 154 independent photographs with 95% confidence intervals, at the threshold used by the application.

Results:

The INPI search identified 28 computer program registrations addressing dengue control in Brazil between 2013 and 2026, seven with related functionalities. None combines a hybrid conversational interface, embedded computer vision, a dedicated spatial database, and active institutional partnership. On the test set, the model flagged 121 of 137 photographs with breeding sites (sensitivity 88.3%, 95% CI 81.7–93.2) and raised no alarm in any of the 17 photographs without them (specificity 100%, 95% CI 80.5–100), with an mAP50 of 61.3% (95% CI 54.9–72.5). The quantized model was 2.7 MB. Performance depended on image source: data from other settings improved results almost exclusively within their own domain. Six of seven planned functionalities reached the functional prototype stage.

Conclusions:

An AI agent that actively guides the search for larval foci was technically feasible for on-device inference with a compact quantized model and without continuous connectivity, contributing to participatory entomological surveillance. The dependence of performance on image source indicates that a local photograph dataset is a prerequisite for operational use. Validation and usability studies in pilot areas are needed to assess engagement, acceptance, and changes in Ae. aegypti density indicators.