NoMoSquito

Inspiration

Dengue remains a recurring public-health challenge in Singapore, yet existing surveillance primarily tells us where dengue is happening now. We were inspired by a simple question: what if we could identify where dengue risk is likely to rise before an outbreak develops? This led to NoMoSquito and our proposed solution, DengueRadar—a predictive early-warning system designed to shift dengue management from reactive response towards earlier, targeted prevention.

What it does

DengueRadar aims to forecast Low, Medium or High dengue risk two weeks ahead at the finest spatial resolution supported by available data. Instead of simply visualising existing clusters, it combines:

• recent dengue cases and active clusters;
• rainfall and temperature;
• population exposure; and
• land-cover characteristics such as vegetation and built-up areas. As new dengue and weather data become available, DengueRadar can update its predictors and refresh its forecasts. The final interface will show where risk may rise, how risk is changing, and the key factors driving each prediction, helping residents take earlier precautions and enabling dengue-control teams to prioritise preventive interventions.

How we built it

We designed DengueRadar as an end-to-end pipeline on Databricks. Lakeflow and Delta Lake will ingest and manage dengue, weather, population and land-cover data. The datasets will then be aligned across space and time and transformed into predictive features such as recent case trends, nearby clusters, cumulative rainfall over the previous 7–28 days and recent temperature conditions. Using MLflow, we plan to train, evaluate and compare forecasting models against a simple recent-case baseline. The selected model will generate a two-week dengue-risk forecast, while an interactive Databricks App / AI-BI dashboard will translate the results into an explainable spatial risk map.

Challenges we ran into

One of our biggest challenges is spatial resolution. Dengue, weather, population and land-cover datasets may be collected at different geographical scales. A highly detailed prediction is only meaningful if historical dengue outcomes are available at a comparable resolution. We therefore plan to forecast at the finest spatial scale consistently supported by the data, rather than creating false precision. Another challenge is determining which predictors genuinely improve forecasting. Rainfall, temperature, land cover and population may all be associated with dengue risk, but adding more variables does not necessarily produce a better model. We therefore plan to evaluate their predictive contribution rather than assuming that every available variable should be included. Finally, we need to balance our ambition with what can realistically be delivered within the two-week sprint.

Accomplishments that we're proud of

We are proud of developing a solution that goes beyond creating another dengue dashboard. Our proposal connects real-time surveillance with predictive analytics, while keeping the output understandable and actionable for its users. We have also designed the project around explainability. Rather than simply labelling an area “High Risk,” DengueRadar aims to show the factors contributing to that forecast—helping users understand why risk may be increasing. Most importantly, we have developed a technically feasible MVP that can be expanded as more granular data and modelling capabilities become available.

What we learned

Developing NoMoSquito taught us that dengue risk is a spatiotemporal problem. Knowing where risk is high is not enough; effective early warning requires understanding where and when risk is changing. We also learned the importance of distinguishing between prediction and causation. A variable such as rainfall or vegetation may help predict future dengue activity without necessarily being the sole cause of an outbreak. This reinforced the importance of evaluating models against historical outcomes and maintaining an interpretable forecasting process. Finally, we learned that good data science is not necessarily about building the most complicated model—it is about producing a prediction that is reliable, explainable and useful for a real decision.

What's next for NoMoSquito

Our immediate goal is to build the DengueRadar MVP: integrate the datasets, engineer temporal and spatial predictors, train and validate the two-week forecasting model, and develop an interactive risk map. Beyond the sprint, we hope to develop DengueRadar into a continuously updated early-warning system. As new dengue and weather observations become available, forecasts could automatically refresh, while periodic model retraining could allow the system to learn from newly observed outbreaks. With more granular surveillance and environmental data, future versions could provide increasingly localised forecasts and support more precise allocation of dengue-control resources.

Built With

• api
• databricks
• deltalake
• geospatialanalysis
• machine-learning
• mlflow
• nea
• onemap
• pyspark
• python