Jack's Place

Inspiration

We started with a question: could we use dengue cases, weather and neighbourhood data to identify where risk might rise next?

The information exists across different datasets, but connecting it is difficult. We wanted DengueRadar to bring it together and help teams plan inspections and community outreach earlier.

What it does

DengueRadar aims to forecast dengue risk across Singapore’s planning areas over the next 14 days.

The dashboard displays low, medium and high risk on a map. Users can select an area to see its predicted probabilities, contributing factors and uncertainty.

How we built it

We started with a dashboard prototype and a modelling plan. We identified inputs such as recent dengue cases, rainfall, temperature, age distribution and nearby parks, hawker centres and construction sites.

We chose multinomial logistic regression to estimate the probability of each risk category. This approach also allows us to examine how different features relate to the predictions.

We designed an evaluation plan to compare the model with a simple forecast based on recent cases and check whether its predicted probabilities match actual outcomes.

The figures in our prototype are illustrative. Connecting the datasets, training the model and validating its performance are still ahead of us.

Challenges we faced

The datasets do not fit together neatly. Some describe planning areas, while others cover larger regions. They also have different update schedules and coverage. For example, the construction dataset includes only HDB projects.

Another challenge was deciding how to present risk clearly. A colour on a map is easy to understand, but it can make a prediction look more certain than it is. We included probabilities and uncertainty intervals to address this.

What we learned

We learned that preparing and aligning the data is just as important as choosing the model. Every input needs to match the location and time period being studied.

We also learned to be careful when explaining predictions. A relationship between a neighbourhood feature and dengue cases does not prove that the feature causes dengue.

What’s next

Our next step is to connect the data and test whether the model provides useful information beyond recent case trends. We then want to refine the dashboard with potential users and work out when an alert would be helpful enough to act on.